BRICS nations will work on early warning data integration by using a shared reference framework to share risk information, improve weather forecasting, and take early action against disasters. Watch for detailed analysis on cross-border emergency response models and technological integrations.


BRICS set of guidelines for disaster management

Since 2 August, Anthropic has been marking content produced by Claude with two separate signals: an invisible watermark embedded in generated text, and C2PA Content Credentials attached to generated image files. Claude Fable 5.1 and Mythos 5.1 carry the mark from launch, along with every model released since that date.&nbsp;The company has also published what defeats it, which is the more useful part of the announcement.&nbsp;How The Text Mark Works&nbsp;The approach is more sophisticated than the tricks that preceded it, and the distinction matters.&nbsp;Earlier attempts at marking AI text relied on things like invisible Unicode characters — trivially removed by anyone who knew to look, and often destroyed accidentally by ordinary copy-paste. Anthropic's watermark lives in the statistical pattern of the tokens the model chose, seeded on a secret key.&nbsp;There is no character to strip, and no letter substitution touches it. The signal is the sequence of words itself, which is why it survives being copied between applications and, the company says, can survive some editing.&nbsp;For detecting the straightforward case — text pasted unchanged from Claude into a document — it works.&nbsp;What Defeats It&nbsp;Anthropic states the limitation directly: if another model rewrites the text, translates it, or paraphrases it thoroughly, the signal vanishes.&nbsp;This follows necessarily from the design. The watermark is computed over the specific chain of words the model produced. Replace those words with different words carrying the same meaning and there is nothing left to measure.&nbsp;The practical consequence is worth stating without euphemism. Defeating this watermark requires pasting the text into a second model and asking it to rewrite the passage. That is not a technical attack. It is one additional step, available to anyone, costing seconds.&nbsp;The image credentials have a parallel weakness. C2PA metadata is a cryptographically signed note in the file, and it is robust against tampering — but re-saving the file through software that does not preserve the metadata removes it, and stripping metadata is a routine operation.&nbsp;Why It Is Still Worth Doing&nbsp;It would be easy to conclude the whole exercise is theatre. That conclusion is too quick.&nbsp;Most AI text is not laundered by an adversary trying to hide it. It is pasted by a student, an employee or a contractor who has not thought about detection at all, and for that population the watermark works as intended. A measure that catches the careless majority while failing against the motivated minority is still a measure — that describes most security controls in ordinary use.&nbsp;The provenance infrastructure also has to exist before it can be required. C2PA is the same standard used by camera manufacturers and photo-editing software, and a future regulation mandating disclosure needs something to mandate.&nbsp;Where The Honesty Runs Out&nbsp;The risk is not the watermark. It is what people conclude from its absence.&nbsp;A detector that returns a positive tells you something reliable. A detector that returns nothing tells you almost nothing — the text may be human, or it may be Claude's output passed through one paraphrase. If institutions start treating a negative result as evidence of human authorship, the tool will have made things worse than no tool at all, and it will be the confident users rather than the engineering that caused it.&nbsp;Anthropic has been clear about this in its own documentation. Whether anyone deploying a checker reads that far is a different question.&nbsp;What To Watch&nbsp;Whether other providers adopt compatible marking, since a standard covering one company's output has limited value. Whether paraphrase-resistant approaches emerge, which is an unsolved research problem rather than an engineering gap. And whether institutions build policies around what these tools actually establish — because a watermark that is trusted more than it deserves is the failure mode here.

Claude's text watermark lives in the statistical pattern of tokens, so it survives copy-paste. Anthropic also states what removes it: one paraphrase through another model.

Since 2 August, Anthropic has been marking content produced by Claude with two separate signals: an invisible watermark embedded in generated text, and C2PA Content Credentials attached to generated image files. Claude Fable 5.1 and Mythos 5.1 carry the mark from launch, along with every model released since that date. The company has also published what defeats it, which is the more useful part of the announcement. How The Text Mark Works The approach is more sophisticated than the tricks that preceded it, and the distinction matters. Earlier attempts at marking AI text relied on things like invisible Unicode characters — trivially removed by anyone who knew to look, and often destroyed accidentally by ordinary copy-paste. Anthropic's watermark lives in the statistical pattern of the tokens the model chose, seeded on a secret key. There is no character to strip, and no letter substitution touches it. The signal is the sequence of words itself, which is why it survives being copied between applications and, the company says, can survive some editing. For detecting the straightforward case — text pasted unchanged from Claude into a document — it works. What Defeats It Anthropic states the limitation directly: if another model rewrites the text, translates it, or paraphrases it thoroughly, the signal vanishes. This follows necessarily from the design. The watermark is computed over the specific chain of words the model produced. Replace those words with different words carrying the same meaning and there is nothing left to measure. The practical consequence is worth stating without euphemism. Defeating this watermark requires pasting the text into a second model and asking it to rewrite the passage. That is not a technical attack. It is one additional step, available to anyone, costing seconds. The image credentials have a parallel weakness. C2PA metadata is a cryptographically signed note in the file, and it is robust against tampering — but re-saving the file through software that does not preserve the metadata removes it, and stripping metadata is a routine operation. Why It Is Still Worth Doing It would be easy to conclude the whole exercise is theatre. That conclusion is too quick. Most AI text is not laundered by an adversary trying to hide it. It is pasted by a student, an employee or a contractor who has not thought about detection at all, and for that population the watermark works as intended. A measure that catches the careless majority while failing against the motivated minority is still a measure — that describes most security controls in ordinary use. The provenance infrastructure also has to exist before it can be required. C2PA is the same standard used by camera manufacturers and photo-editing software, and a future regulation mandating disclosure needs something to mandate. Where The Honesty Runs Out The risk is not the watermark. It is what people conclude from its absence. A detector that returns a positive tells you something reliable. A detector that returns nothing tells you almost nothing — the text may be human, or it may be Claude's output passed through one paraphrase. If institutions start treating a negative result as evidence of human authorship, the tool will have made things worse than no tool at all, and it will be the confident users rather than the engineering that caused it. Anthropic has been clear about this in its own documentation. Whether anyone deploying a checker reads that far is a different question. What To Watch Whether other providers adopt compatible marking, since a standard covering one company's output has limited value. Whether paraphrase-resistant approaches emerge, which is an unsolved research problem rather than an engineering gap. And whether institutions build policies around what these tools actually establish — because a watermark that is trusted more than it deserves is the failure mode here.

Anthropic has been marking Claude-generated text with an invisible statistical watermark since August, and attaching C2PA Content Credentials to generated image files. The engineering is more robust than earlier approaches. The company has also stated the limitation plainly: run the text through another model and the signal disappears.

Anthropic built a watermark for AI, then published limitation that defeats it

Anthropic has launched Claude for Teachers as a free Enterprise offering for schools and districts, with centrally managed access and single sign-on.&nbsp;The pricing will lead the coverage. The administrative detail is the more consequential part.&nbsp;Why Central Management Is The Point&nbsp;Teachers have been using AI tools in classrooms for two years, mostly without institutional involvement — personal accounts, personal subscriptions, no visibility for the school and no record of what was used or how.&nbsp;Central administration with single sign-on changes what a school can do about that. It gives an institution a list of who has access, the ability to revoke it, a consistent configuration, and an audit trail. For a district lawyer or a data protection officer, that is the difference between a tool they can sanction and one they have to pretend is not happening.&nbsp;Free removes the remaining obstacle, which is not usually the licence cost. It is the procurement process that a licence cost triggers — the evaluation, the budget line, the committee. A free tier with enterprise controls can be adopted by a district without any of that.&nbsp;The Strategy, Named Plainly&nbsp;Free distribution into education is one of the oldest plays in the technology industry, and it works.&nbsp;Apple did it with school computers. Google did it with Chromebooks and Workspace, and in doing so captured a generation's default assumptions about what a document is and where it lives. The mechanism is not brand loyalty in any sentimental sense — it is that the tool a person learns on becomes the tool they are fluent in, and fluency is what people carry into workplaces that then buy licences.&nbsp;For a company preparing to go public on projected revenue growth, a channel that builds future demand at the cost of serving it is a rational investment. Saying so is not cynicism; it is the reason the offer exists and can be taken at face value.&nbsp;The Questions Schools Should Ask&nbsp;Three things are worth establishing before a district deploys this, and they are not answered by the announcement.&nbsp;What happens to student and teacher inputs — whether they are used for training, how long they are retained, and what the deletion path is. Whether 'free' is permanent or introductory, because a tool embedded in lesson planning across a district is expensive to remove later. And what the exit looks like: whether material created in the tool can be exported in a usable form if the school leaves.&nbsp;These are the standard questions for any institutional software adoption. Free tends to suppress them, because the procurement process that would normally ask them is the process being bypassed.&nbsp;The Part That Is Genuinely Useful&nbsp;None of that makes this a bad offer for schools, and it would be wrong to imply otherwise.&nbsp;Teachers are already using these tools. The realistic alternative to a managed deployment is not abstention — it is the current situation, in which staff use personal accounts with no oversight, no data agreement and no institutional record. A sanctioned tier with administrative controls is straightforwardly better than that, on privacy grounds as much as on pedagogical ones.&nbsp;The correct posture is to take the offer and read the terms, rather than to treat free as a reason to skip the reading.&nbsp;What To Watch&nbsp;Whether the free tier carries usage limits that make it workable for a whole district or only for pilots. Whether OpenAI and Google respond with comparable education tiers, which would turn this into a distribution contest. And whether education authorities set procurement standards for AI tools — because at present each district is negotiating alone against companies that have done this many times.

Claude for Teachers is a free Enterprise tier with central management and single sign-on. The admin controls are the substance — and free distribution into education has a long, well-understood history.

Anthropic has launched Claude for Teachers as a free Enterprise offering for schools and districts, with centrally managed access and single sign-on. The pricing will lead the coverage. The administrative detail is the more consequential part. Why Central Management Is The Point Teachers have been using AI tools in classrooms for two years, mostly without institutional involvement — personal accounts, personal subscriptions, no visibility for the school and no record of what was used or how. Central administration with single sign-on changes what a school can do about that. It gives an institution a list of who has access, the ability to revoke it, a consistent configuration, and an audit trail. For a district lawyer or a data protection officer, that is the difference between a tool they can sanction and one they have to pretend is not happening. Free removes the remaining obstacle, which is not usually the licence cost. It is the procurement process that a licence cost triggers — the evaluation, the budget line, the committee. A free tier with enterprise controls can be adopted by a district without any of that. The Strategy, Named Plainly Free distribution into education is one of the oldest plays in the technology industry, and it works. Apple did it with school computers. Google did it with Chromebooks and Workspace, and in doing so captured a generation's default assumptions about what a document is and where it lives. The mechanism is not brand loyalty in any sentimental sense — it is that the tool a person learns on becomes the tool they are fluent in, and fluency is what people carry into workplaces that then buy licences. For a company preparing to go public on projected revenue growth, a channel that builds future demand at the cost of serving it is a rational investment. Saying so is not cynicism; it is the reason the offer exists and can be taken at face value. The Questions Schools Should Ask Three things are worth establishing before a district deploys this, and they are not answered by the announcement. What happens to student and teacher inputs — whether they are used for training, how long they are retained, and what the deletion path is. Whether 'free' is permanent or introductory, because a tool embedded in lesson planning across a district is expensive to remove later. And what the exit looks like: whether material created in the tool can be exported in a usable form if the school leaves. These are the standard questions for any institutional software adoption. Free tends to suppress them, because the procurement process that would normally ask them is the process being bypassed. The Part That Is Genuinely Useful None of that makes this a bad offer for schools, and it would be wrong to imply otherwise. Teachers are already using these tools. The realistic alternative to a managed deployment is not abstention — it is the current situation, in which staff use personal accounts with no oversight, no data agreement and no institutional record. A sanctioned tier with administrative controls is straightforwardly better than that, on privacy grounds as much as on pedagogical ones. The correct posture is to take the offer and read the terms, rather than to treat free as a reason to skip the reading. What To Watch Whether the free tier carries usage limits that make it workable for a whole district or only for pilots. Whether OpenAI and Google respond with comparable education tiers, which would turn this into a distribution contest. And whether education authorities set procurement standards for AI tools — because at present each district is negotiating alone against companies that have done this many times.

Anthropic has launched Claude for Teachers, a free Enterprise offering for schools and districts with centrally managed access and single sign-on. The administrative features are the substance of the announcement — and free distribution into education is a strategy with a long and well-understood history in this industry.

Anthropic is giving Claude to schools free, and the reason is not philanthropy

In February, Anthropic closed a $30 billion funding round at a $380 billion valuation.&nbsp;Investors have since offered close to $800 billion. Reporting now indicates backers are targeting a valuation of at least $2 trillion in an initial public offering, with the timeline pushed toward mid-October and the raise expected to exceed $60 billion.&nbsp;That is a fivefold move in roughly seven months, in a company whose financial disclosures remain confidential.&nbsp;The Revenue Case&nbsp;There is a number behind the optimism, and it is a real one.&nbsp;Anthropic's backers expect annualised revenue to reach $100 billion to $120 billion by the end of the year — more than ten times its level at the start of 2026. If that materialises, a valuation in the trillions is not obviously unmoored; it implies a multiple that, while high, is within range of what fast-growing software businesses have commanded.&nbsp;The load-bearing word is 'expect'. This is a projection held by people with capital committed to the outcome, not a reported figure, and the year is not over.&nbsp;Why The Range Itself Is The Story&nbsp;Follow the figures attached to this company across 2026 and the pattern is more informative than any single number.&nbsp;A $350 billion valuation was reported around Google's investment. $380 billion in February. $965 billion has been reported for a round in May. Offers near $800 billion now. A $2 trillion target for October. Earlier reporting mentioned $61.5 billion as recently as early 2025.&nbsp;Those figures are not all measuring the same thing — some are round prices, some are secondary offers, some are targets — but the spread is extraordinary, and it does not move in one direction. An $800 billion offer sits below a $965 billion figure reported months earlier.&nbsp;What that indicates is not a company being repriced by new information. It is a company being priced by appetite, in a market where the quantity of capital seeking AI exposure exceeds the number of places to put it.&nbsp;What The Filing Will Settle&nbsp;The IPO is the mechanism that ends this, which is the genuinely useful thing about it.&nbsp;Anthropic has submitted a confidential draft S-1. When the public version lands it must disclose what private rounds never required: cost of revenue, gross margin, the economics of serving frontier models at scale, and customer concentration. Revenue growth of ten times in a year is impressive; whether it is profitable at the margin is the question no one outside the company can currently answer.&nbsp;Compute costs are the crux. A business whose largest input cost scales with usage has a different margin structure from conventional software, and that difference is precisely what the filing will make visible.&nbsp;The Case For Caution&nbsp;Two things are worth holding in view.&nbsp;Google and Amazon hold large positions in Anthropic, and their stakes have contributed materially to reported profits at both companies. Those positions are currently marked against negotiated private valuations. A public listing replaces that with a market price, which can move in either direction and will be visible quarterly.&nbsp;And a company listing at the top of its own valuation range transfers the risk of that range being wrong from a small number of institutions to a broad base of public shareholders. That is what an IPO does. It is not a criticism, but it is worth naming while the numbers are still climbing.&nbsp;What To Watch&nbsp;Whether the public S-1 arrives before the mid-October window and what gross margin it discloses. Whether the $100-120 billion revenue projection is stated in the filing or remains an investor expectation. And where the listing actually prices — because the gap between $380 billion in February and $2 trillion in October is not a repricing, it is a different asset class, and the market is about to choose.

Anthropic raised at $380bn in February. Investors have offered near $800bn and are targeting $2tn in an October IPO — a fivefold move with no public financials to explain it.

In February, Anthropic closed a $30 billion funding round at a $380 billion valuation. Investors have since offered close to $800 billion. Reporting now indicates backers are targeting a valuation of at least $2 trillion in an initial public offering, with the timeline pushed toward mid-October and the raise expected to exceed $60 billion. That is a fivefold move in roughly seven months, in a company whose financial disclosures remain confidential. The Revenue Case There is a number behind the optimism, and it is a real one. Anthropic's backers expect annualised revenue to reach $100 billion to $120 billion by the end of the year — more than ten times its level at the start of 2026. If that materialises, a valuation in the trillions is not obviously unmoored; it implies a multiple that, while high, is within range of what fast-growing software businesses have commanded. The load-bearing word is 'expect'. This is a projection held by people with capital committed to the outcome, not a reported figure, and the year is not over. Why The Range Itself Is The Story Follow the figures attached to this company across 2026 and the pattern is more informative than any single number. A $350 billion valuation was reported around Google's investment. $380 billion in February. $965 billion has been reported for a round in May. Offers near $800 billion now. A $2 trillion target for October. Earlier reporting mentioned $61.5 billion as recently as early 2025. Those figures are not all measuring the same thing — some are round prices, some are secondary offers, some are targets — but the spread is extraordinary, and it does not move in one direction. An $800 billion offer sits below a $965 billion figure reported months earlier. What that indicates is not a company being repriced by new information. It is a company being priced by appetite, in a market where the quantity of capital seeking AI exposure exceeds the number of places to put it. What The Filing Will Settle The IPO is the mechanism that ends this, which is the genuinely useful thing about it. Anthropic has submitted a confidential draft S-1. When the public version lands it must disclose what private rounds never required: cost of revenue, gross margin, the economics of serving frontier models at scale, and customer concentration. Revenue growth of ten times in a year is impressive; whether it is profitable at the margin is the question no one outside the company can currently answer. Compute costs are the crux. A business whose largest input cost scales with usage has a different margin structure from conventional software, and that difference is precisely what the filing will make visible. The Case For Caution Two things are worth holding in view. Google and Amazon hold large positions in Anthropic, and their stakes have contributed materially to reported profits at both companies. Those positions are currently marked against negotiated private valuations. A public listing replaces that with a market price, which can move in either direction and will be visible quarterly. And a company listing at the top of its own valuation range transfers the risk of that range being wrong from a small number of institutions to a broad base of public shareholders. That is what an IPO does. It is not a criticism, but it is worth naming while the numbers are still climbing. What To Watch Whether the public S-1 arrives before the mid-October window and what gross margin it discloses. Whether the $100-120 billion revenue projection is stated in the filing or remains an investor expectation. And where the listing actually prices — because the gap between $380 billion in February and $2 trillion in October is not a repricing, it is a different asset class, and the market is about to choose.

Anthropic closed a $30 billion round at a $380 billion valuation in February. Investors have since offered close to $800 billion, and reporting indicates backers are targeting a valuation of at least $2 trillion in an IPO now pushed toward mid-October. The figures are moving faster than any underlying disclosure explains.

Anthropic's valuation has moved from $380 billion to a $2 trillion target in seven months

Twelve disclosed AI chip funding rounds between January and August raised $5.37 billion. Inference-focused companies accounted for eight of the twelve deals and about 67 per cent of the capital.&nbsp;For most of the current AI cycle, the money and the attention went to training hardware — the chips used to build models. That is no longer where the majority of new chip capital is going.&nbsp;The Distinction, Briefly&nbsp;Training is the process of building a model: enormously compute-intensive, performed a limited number of times, dominated by a small number of organisations with the capital to attempt it.&nbsp;Inference is running the finished model to answer a query. Each individual operation is far cheaper, but it happens continuously, at whatever scale the product has users.&nbsp;Over a successful model's life, inference is where most of the compute is ultimately consumed. It is the recurring cost, and it scales with adoption rather than with ambition.&nbsp;Why Investors Have Turned&nbsp;Two reasons, and they point in the same direction.&nbsp;The first is market structure. Training hardware is a market Nvidia dominates comprehensively, and a startup entering it competes against both the silicon and the software ecosystem built around it over fifteen years. Inference is more fragmented, more sensitive to cost per query and latency, and more open to specialised designs that do one thing efficiently rather than everything adequately.&nbsp;The second is that the shape of AI spending is changing. If models are being deployed into production — and agentic systems running continuously are the clearest example — then the volume of inference grows with usage while training spend stays lumpy and episodic. Investors backing inference are backing the recurring half of the bill.&nbsp;What It Signals&nbsp;This is a bet that the AI buildout is entering its deployment phase rather than its construction phase.&nbsp;Capital allocated to inference assumes models good enough to deploy already exist, and that the constraint ahead is serving them economically rather than building better ones. That is a meaningfully different view of the next few years from the one implied by ever-larger training clusters.&nbsp;Both can be true simultaneously — frontier labs will keep building larger models while everyone else optimises the cost of running what exists. But the capital is now split in a way it was not eighteen months ago.&nbsp;The Caution&nbsp;Two qualifiers belong on these numbers.&nbsp;Twelve disclosed rounds is a small sample, and disclosure is selective — private rounds that never surface are absent by definition, which may bias the picture in either direction. And $5.37 billion, while substantial for chip startups, is modest against the tens of billions flowing into AI infrastructure overall. This is a signal about where new venture capital is going, not a measurement of total spending.&nbsp;The direction is more reliable than the magnitude.&nbsp;What To Watch&nbsp;Whether any of these companies reaches volume deployment against Nvidia's inference offerings, which is where previous challengers have failed rather than at the design stage. Whether the ratio holds through the rest of the year or reverts as new training clusters are announced. And whether the large cloud providers keep building inference silicon in-house, which would compress the market these startups are being funded to enter.

Twelve disclosed AI chip rounds raised $5.37bn this year, and inference companies took eight of them and about 67% of the capital — a bet that the constraint ahead is serving models, not building them.

Twelve disclosed AI chip funding rounds between January and August raised $5.37 billion. Inference-focused companies accounted for eight of the twelve deals and about 67 per cent of the capital. For most of the current AI cycle, the money and the attention went to training hardware — the chips used to build models. That is no longer where the majority of new chip capital is going. The Distinction, Briefly Training is the process of building a model: enormously compute-intensive, performed a limited number of times, dominated by a small number of organisations with the capital to attempt it. Inference is running the finished model to answer a query. Each individual operation is far cheaper, but it happens continuously, at whatever scale the product has users. Over a successful model's life, inference is where most of the compute is ultimately consumed. It is the recurring cost, and it scales with adoption rather than with ambition. Why Investors Have Turned Two reasons, and they point in the same direction. The first is market structure. Training hardware is a market Nvidia dominates comprehensively, and a startup entering it competes against both the silicon and the software ecosystem built around it over fifteen years. Inference is more fragmented, more sensitive to cost per query and latency, and more open to specialised designs that do one thing efficiently rather than everything adequately. The second is that the shape of AI spending is changing. If models are being deployed into production — and agentic systems running continuously are the clearest example — then the volume of inference grows with usage while training spend stays lumpy and episodic. Investors backing inference are backing the recurring half of the bill. What It Signals This is a bet that the AI buildout is entering its deployment phase rather than its construction phase. Capital allocated to inference assumes models good enough to deploy already exist, and that the constraint ahead is serving them economically rather than building better ones. That is a meaningfully different view of the next few years from the one implied by ever-larger training clusters. Both can be true simultaneously — frontier labs will keep building larger models while everyone else optimises the cost of running what exists. But the capital is now split in a way it was not eighteen months ago. The Caution Two qualifiers belong on these numbers. Twelve disclosed rounds is a small sample, and disclosure is selective — private rounds that never surface are absent by definition, which may bias the picture in either direction. And $5.37 billion, while substantial for chip startups, is modest against the tens of billions flowing into AI infrastructure overall. This is a signal about where new venture capital is going, not a measurement of total spending. The direction is more reliable than the magnitude. What To Watch Whether any of these companies reaches volume deployment against Nvidia's inference offerings, which is where previous challengers have failed rather than at the design stage. Whether the ratio holds through the rest of the year or reverts as new training clusters are announced. And whether the large cloud providers keep building inference silicon in-house, which would compress the market these startups are being funded to enter.

Between January and August, twelve disclosed AI chip funding rounds raised $5.37 billion. Inference-focused companies took eight of those deals and roughly 67 per cent of the capital — a reversal from the years when training hardware attracted almost everything, and a signal about where investors think the durable spending lies.

The AI chip money has quietly moved from training to inference

The US Cybersecurity and Infrastructure Security Agency has added seven vulnerabilities to its Known Exploited Vulnerabilities catalogue. One of them is in the AI stack.&nbsp;CVE-2026-59822, carrying a CVSS score of 8.8, is an authentication bypass in the MCP Streamable HTTP endpoint of LiteLLM — an open-source proxy widely used to sit between applications and the various language model providers they call.&nbsp;What The Listing Means&nbsp;The KEV catalogue is a specific claim, and it is stronger than a vulnerability disclosure.&nbsp;Entries are added when there is reliable evidence of active exploitation in the wild. This is not a researcher's proof of concept or a theoretical attack path — it means someone has been using it. For US federal civilian agencies, KEV inclusion carries a binding remediation deadline, and most large enterprises treat the catalogue as a de facto patch priority list.&nbsp;Why This Component Matters&nbsp;LiteLLM is infrastructure most people using it never think about, which is precisely what makes it worth attention.&nbsp;A proxy of this kind sits at an unusually sensitive point. To do its job it typically holds provider API keys for every model an organisation calls, and it sees the traffic in both directions — the prompts going out and the completions coming back. That traffic frequently contains exactly the material an organisation would least like to lose: internal documents, customer records, source code, credentials pasted into a prompt by someone in a hurry.&nbsp;An authentication bypass on a component like that is not a peripheral issue. It is access to the keys and the conversation at once.&nbsp;The Broader Point&nbsp;This is the part worth generalising, because it will recur.&nbsp;Enormous attention has gone to whether models themselves can be manipulated — prompt injection, jailbreaks, data extraction from training sets. Far less has gone to the ordinary software assembled around them: proxies, gateways, orchestration layers, vector databases, the connective tissue that turns a model into a product.&nbsp;That layer is conventional software with conventional vulnerabilities. It has often been adopted very fast, by teams under pressure to ship AI features, with less security review than an equivalently sensitive component would normally receive. And it is frequently open source maintained by a small number of people relative to how widely it has been deployed.&nbsp;The attack surface of an AI system is not primarily the model. It is everything that was built around the model in a hurry.&nbsp;What Organisations Should Actually Do&nbsp;The immediate action is narrow: identify whether LiteLLM is deployed anywhere in the estate, patch it, and rotate every provider API key it held. A bypass that may have exposed stored credentials is not remediated by patching alone.&nbsp;The broader action is an inventory. Most organisations that adopted AI tooling quickly do not have a clear list of what is sitting in the path between their applications and their model providers, and cannot answer what any of those components hold.&nbsp;What To Watch&nbsp;Whether further AI infrastructure components appear in the KEV catalogue in the coming months, which would confirm this as a pattern rather than an incident. Whether the AI tooling ecosystem adopts the disclosure and patching discipline of more mature infrastructure software. And whether enterprises begin treating the model-adjacent stack as security-critical, which most currently do not.

CISA has added LiteLLM's CVE-2026-59822, an authentication bypass scoring 8.8, to its Known Exploited Vulnerabilities catalogue — meaning it is being used, on a component that holds API keys and sees every prompt.

The US Cybersecurity and Infrastructure Security Agency has added seven vulnerabilities to its Known Exploited Vulnerabilities catalogue. One of them is in the AI stack. CVE-2026-59822, carrying a CVSS score of 8.8, is an authentication bypass in the MCP Streamable HTTP endpoint of LiteLLM — an open-source proxy widely used to sit between applications and the various language model providers they call. What The Listing Means The KEV catalogue is a specific claim, and it is stronger than a vulnerability disclosure. Entries are added when there is reliable evidence of active exploitation in the wild. This is not a researcher's proof of concept or a theoretical attack path — it means someone has been using it. For US federal civilian agencies, KEV inclusion carries a binding remediation deadline, and most large enterprises treat the catalogue as a de facto patch priority list. Why This Component Matters LiteLLM is infrastructure most people using it never think about, which is precisely what makes it worth attention. A proxy of this kind sits at an unusually sensitive point. To do its job it typically holds provider API keys for every model an organisation calls, and it sees the traffic in both directions — the prompts going out and the completions coming back. That traffic frequently contains exactly the material an organisation would least like to lose: internal documents, customer records, source code, credentials pasted into a prompt by someone in a hurry. An authentication bypass on a component like that is not a peripheral issue. It is access to the keys and the conversation at once. The Broader Point This is the part worth generalising, because it will recur. Enormous attention has gone to whether models themselves can be manipulated — prompt injection, jailbreaks, data extraction from training sets. Far less has gone to the ordinary software assembled around them: proxies, gateways, orchestration layers, vector databases, the connective tissue that turns a model into a product. That layer is conventional software with conventional vulnerabilities. It has often been adopted very fast, by teams under pressure to ship AI features, with less security review than an equivalently sensitive component would normally receive. And it is frequently open source maintained by a small number of people relative to how widely it has been deployed. The attack surface of an AI system is not primarily the model. It is everything that was built around the model in a hurry. What Organisations Should Actually Do The immediate action is narrow: identify whether LiteLLM is deployed anywhere in the estate, patch it, and rotate every provider API key it held. A bypass that may have exposed stored credentials is not remediated by patching alone. The broader action is an inventory. Most organisations that adopted AI tooling quickly do not have a clear list of what is sitting in the path between their applications and their model providers, and cannot answer what any of those components hold. What To Watch Whether further AI infrastructure components appear in the KEV catalogue in the coming months, which would confirm this as a pattern rather than an incident. Whether the AI tooling ecosystem adopts the disclosure and patching discipline of more mature infrastructure software. And whether enterprises begin treating the model-adjacent stack as security-critical, which most currently do not.

The US Cybersecurity and Infrastructure Security Agency has added seven vulnerabilities to its Known Exploited Vulnerabilities catalogue, among them an authentication bypass in LiteLLM — a widely deployed proxy that sits between applications and the language models they call. Inclusion in the catalogue means it is not theoretical.

A flaw in a widely used AI proxy is now on the US list of actively exploited bugs

Nvidia's equity investments in AI companies stood at roughly $99 billion as of late July, according to a CNBC tally. A year earlier the figure was about $7 billion.&nbsp;That is a fourteen-fold expansion in twelve months, and it has made the world's most important chip supplier simultaneously one of the largest investors in the companies that buy its chips.&nbsp;Where The Money Went&nbsp;The single largest position is OpenAI, where Nvidia invested $30 billion as part of a $110 billion funding round.&nbsp;Beyond that, its 2026 equity commitments run past $40 billion across the stack, including seven multi-billion-dollar stakes in public companies — among them up to $3.2 billion for Corning and $2.1 billion for IREN. Separately, the company has signed memoranda of understanding targeting more than $500 billion of third-party capital over time.&nbsp;The Structural Detail&nbsp;The most consequential fact about this strategy is what it is not.&nbsp;Nvidia has vertically integrated the AI economy without acquiring any of the participants. Every position is a minority equity stake, not a purchase — which means none of it passes through merger review, in any jurisdiction, at any threshold.&nbsp;An acquisition of comparable scope would attract antitrust scrutiny in Washington, Brussels and Beijing simultaneously. A minority stake attracts almost none, while delivering a substantial share of the strategic benefit: influence over roadmaps, early visibility into demand, and a claim on the upside of customers whose growth Nvidia's own supply enables.&nbsp;This is not a loophole in any illegal sense. It is a straightforward reading of where the regulatory perimeter sits, and a decision to operate just inside it.&nbsp;The Circularity Problem&nbsp;There is a criticism of this arrangement that has been made repeatedly, and it is not baseless.&nbsp;When a supplier invests billions in its own customers, some portion of the revenue those customers generate for the supplier was funded by the supplier. Demand of that kind is real in an accounting sense but is not independent evidence that the underlying market is as large as the revenue suggests.&nbsp;Nvidia's counter is that these are minority positions in companies with substantial independent funding, that it has no control over purchasing decisions, and that the AI infrastructure buildout would proceed with or without its participation. Both readings can be partly true, and the honest position is that nobody outside the companies involved can currently separate the two.&nbsp;What it does mean is that anyone reading Nvidia's demand figures as a clean signal of independent market appetite should discount them somewhat, and the size of that discount is unknown.&nbsp;The Risk It Creates&nbsp;The strategy concentrates a second kind of exposure on top of the first.&nbsp;Nvidia already carries the risk that AI infrastructure demand slows. It now additionally carries the mark-to-market risk on $99 billion of equity in companies whose valuations depend on the same demand. Those risks are not independent — they would arrive together, and a correction in AI valuations would hit both sides of the balance sheet at once.&nbsp;For a company of Nvidia's size that is survivable. It is still a materially different risk profile from selling chips for cash.&nbsp;What To Watch&nbsp;Whether regulators show any interest in minority stakes at this concentration, which would be a novel application of merger doctrine. Whether the $500 billion of memoranda converts into deployed capital or stays aspirational. And how much of Nvidia's reported demand can be traced to entities it has itself funded — a disclosure no one currently requires it to make.

Nvidia's AI equity book went from about $7bn to roughly $99bn in twelve months, $30bn of it in OpenAI. Every position is a minority stake — which keeps the whole strategy outside merger review.

Nvidia's equity investments in AI companies stood at roughly $99 billion as of late July, according to a CNBC tally. A year earlier the figure was about $7 billion. That is a fourteen-fold expansion in twelve months, and it has made the world's most important chip supplier simultaneously one of the largest investors in the companies that buy its chips. Where The Money Went The single largest position is OpenAI, where Nvidia invested $30 billion as part of a $110 billion funding round. Beyond that, its 2026 equity commitments run past $40 billion across the stack, including seven multi-billion-dollar stakes in public companies — among them up to $3.2 billion for Corning and $2.1 billion for IREN. Separately, the company has signed memoranda of understanding targeting more than $500 billion of third-party capital over time. The Structural Detail The most consequential fact about this strategy is what it is not. Nvidia has vertically integrated the AI economy without acquiring any of the participants. Every position is a minority equity stake, not a purchase — which means none of it passes through merger review, in any jurisdiction, at any threshold. An acquisition of comparable scope would attract antitrust scrutiny in Washington, Brussels and Beijing simultaneously. A minority stake attracts almost none, while delivering a substantial share of the strategic benefit: influence over roadmaps, early visibility into demand, and a claim on the upside of customers whose growth Nvidia's own supply enables. This is not a loophole in any illegal sense. It is a straightforward reading of where the regulatory perimeter sits, and a decision to operate just inside it. The Circularity Problem There is a criticism of this arrangement that has been made repeatedly, and it is not baseless. When a supplier invests billions in its own customers, some portion of the revenue those customers generate for the supplier was funded by the supplier. Demand of that kind is real in an accounting sense but is not independent evidence that the underlying market is as large as the revenue suggests. Nvidia's counter is that these are minority positions in companies with substantial independent funding, that it has no control over purchasing decisions, and that the AI infrastructure buildout would proceed with or without its participation. Both readings can be partly true, and the honest position is that nobody outside the companies involved can currently separate the two. What it does mean is that anyone reading Nvidia's demand figures as a clean signal of independent market appetite should discount them somewhat, and the size of that discount is unknown. The Risk It Creates The strategy concentrates a second kind of exposure on top of the first. Nvidia already carries the risk that AI infrastructure demand slows. It now additionally carries the mark-to-market risk on $99 billion of equity in companies whose valuations depend on the same demand. Those risks are not independent — they would arrive together, and a correction in AI valuations would hit both sides of the balance sheet at once. For a company of Nvidia's size that is survivable. It is still a materially different risk profile from selling chips for cash. What To Watch Whether regulators show any interest in minority stakes at this concentration, which would be a novel application of merger doctrine. Whether the $500 billion of memoranda converts into deployed capital or stays aspirational. And how much of Nvidia's reported demand can be traced to entities it has itself funded — a disclosure no one currently requires it to make.

Nvidia's equity investments in AI companies reached roughly $99 billion as of late July, up from about $7 billion a year earlier, according to CNBC's tally. The structure is what stands out: the company has positioned itself across the entire AI stack through minority stakes rather than acquisitions, which keeps it out of the path of merger review.

Nvidia has taken $99 billion in equity stakes across AI in a year

The frontier AI labs have spent years making broadly the same argument about regulation: that they take safety seriously, that they welcome sensible oversight, and that rules should be workable. In Massachusetts, that consensus has visibly broken.&nbsp;A state Senate proposal, carried as part of a larger economic development bill, would require large AI developers to commission independent evaluators to assess catastrophic risks in their models every four months.&nbsp;Anthropic is backing it and has described its preferred version as the most robust AI safety legislation in the nation. OpenAI and Google are opposing it.&nbsp;What Is Actually In Dispute&nbsp;The disagreement is narrower than 'safety versus no safety', and it is worth stating precisely.&nbsp;Nobody in this argument is opposing third-party review. The dispute is about cadence and independence. Anthropic wants independent risk assessments every four to six months. OpenAI favours something closer to the approach Illinois took — annual third-party audits.&nbsp;The gap between four months and twelve months sounds procedural. It is not. Frontier models now ship substantially revised versions several times a year, and an annual audit can assess a model that has already been superseded twice. A four-month cadence is an attempt to make evaluation track the release cycle rather than the calendar.&nbsp;The Case Against, Taken Seriously&nbsp;OpenAI's objection is not obviously self-serving, and it should be engaged with rather than dismissed.&nbsp;Its argument is that aggressive state-by-state regulation produces a patchwork of inconsistent rules, and that the compliance burden of satisfying fifty different regimes delays the release of beneficial models — including, it notes, tools used for cybersecurity defence.&nbsp;That is a real problem. A company shipping into all fifty states under conflicting evaluation regimes faces genuine cost, and there is no federal framework to preempt it. The patchwork concern is the strongest argument available to the opposing side.&nbsp;The weakness is that it is an argument for federal legislation, and the same companies have not conspicuously pushed for a federal standard strict enough to preempt anything. An objection to state rules that does not come with a proposed alternative functions, in practice, as an objection to rules.&nbsp;Why Anthropic's Position Is Also Not Free&nbsp;It would be naive to read Anthropic's stance as pure principle.&nbsp;The company has built its market identity on safety, and a regulation that makes rigorous evaluation mandatory converts that identity into a competitive moat. Compliance costs that are burdensome for a smaller competitor are absorbable for a company already running the evaluations internally.&nbsp;That does not make the position wrong. A rule can be simultaneously good policy and commercially convenient for whoever proposed it, and the coincidence is a reason to examine the rule rather than the motive.&nbsp;What Makes This Consequential&nbsp;State law has become the actual venue for AI regulation in the United States, in the absence of anything federal. What Massachusetts settles on will be copied, adapted or reacted against by other legislatures, in the way California's privacy law set the template a decade ago.&nbsp;It is also the first time the frontier labs have publicly taken opposing sides on a specific binding rule rather than on principles. That breaks the industry's ability to speak with one voice to legislators, which is a more significant development than the bill itself.&nbsp;What To Watch&nbsp;Whether the four-month cadence survives into the final text or is negotiated toward the annual model. Whether other states move before Massachusetts concludes. And whether any of these companies now supports a federal standard with teeth — because that is the test of whether the patchwork objection was about the patchwork.

Massachusetts would require independent evaluations of catastrophic AI risk every four months. Anthropic backs it; OpenAI and Google want annual audits — the first public split between frontier labs on a binding rule.

The frontier AI labs have spent years making broadly the same argument about regulation: that they take safety seriously, that they welcome sensible oversight, and that rules should be workable. In Massachusetts, that consensus has visibly broken. A state Senate proposal, carried as part of a larger economic development bill, would require large AI developers to commission independent evaluators to assess catastrophic risks in their models every four months. Anthropic is backing it and has described its preferred version as the most robust AI safety legislation in the nation. OpenAI and Google are opposing it. What Is Actually In Dispute The disagreement is narrower than 'safety versus no safety', and it is worth stating precisely. Nobody in this argument is opposing third-party review. The dispute is about cadence and independence. Anthropic wants independent risk assessments every four to six months. OpenAI favours something closer to the approach Illinois took — annual third-party audits. The gap between four months and twelve months sounds procedural. It is not. Frontier models now ship substantially revised versions several times a year, and an annual audit can assess a model that has already been superseded twice. A four-month cadence is an attempt to make evaluation track the release cycle rather than the calendar. The Case Against, Taken Seriously OpenAI's objection is not obviously self-serving, and it should be engaged with rather than dismissed. Its argument is that aggressive state-by-state regulation produces a patchwork of inconsistent rules, and that the compliance burden of satisfying fifty different regimes delays the release of beneficial models — including, it notes, tools used for cybersecurity defence. That is a real problem. A company shipping into all fifty states under conflicting evaluation regimes faces genuine cost, and there is no federal framework to preempt it. The patchwork concern is the strongest argument available to the opposing side. The weakness is that it is an argument for federal legislation, and the same companies have not conspicuously pushed for a federal standard strict enough to preempt anything. An objection to state rules that does not come with a proposed alternative functions, in practice, as an objection to rules. Why Anthropic's Position Is Also Not Free It would be naive to read Anthropic's stance as pure principle. The company has built its market identity on safety, and a regulation that makes rigorous evaluation mandatory converts that identity into a competitive moat. Compliance costs that are burdensome for a smaller competitor are absorbable for a company already running the evaluations internally. That does not make the position wrong. A rule can be simultaneously good policy and commercially convenient for whoever proposed it, and the coincidence is a reason to examine the rule rather than the motive. What Makes This Consequential State law has become the actual venue for AI regulation in the United States, in the absence of anything federal. What Massachusetts settles on will be copied, adapted or reacted against by other legislatures, in the way California's privacy law set the template a decade ago. It is also the first time the frontier labs have publicly taken opposing sides on a specific binding rule rather than on principles. That breaks the industry's ability to speak with one voice to legislators, which is a more significant development than the bill itself. What To Watch Whether the four-month cadence survives into the final text or is negotiated toward the annual model. Whether other states move before Massachusetts concludes. And whether any of these companies now supports a federal standard with teeth — because that is the test of whether the patchwork objection was about the patchwork.

A Massachusetts proposal would require large AI developers to hire independent evaluators to assess catastrophic risk in their models every four months — a cadence stricter than anything else at state level in the United States. Anthropic supports it. OpenAI and Google are opposing it, arguing for annual third-party audits instead.

Anthropic has broken with OpenAI and Google over the toughest AI safety bill in America

Artificial intelligence is no longer being used only to analyse human language. Scientists are now using it to study how animals communicate, raising an extraordinary possibility: could AI eventually help humans understand or even communicate with other species?&nbsp;According to Bloomberg, researchers are using machine learning to process enormous collections of animal sounds that would be almost impossible to study manually. The work spans species including crows, whales and wolves, with scientists looking for patterns hidden inside thousands of recordings.<h2>AI helps decode crow calls</h2>Biologist Vittorio Baglione has spent around three decades studying carrion crows in Spain.In 2018, he and his research partner Daniela Canestrari fitted small microphones to 43 birds. The experiment produced thousands of hours of recordings involving adult crows, chicks and other birds.&nbsp;The collection has since grown to about 150,000 recordings. Researchers turned to an AI model developed by the Earth Species Project to separate sounds and identify which birds produced them. One call particularly interested Baglione. He observed that when a buzzard attacked a crow's nest, the bird produced a soft sound before other crows arrived to defend it.&nbsp;He interpreted the call as roughly meaning: “Hey! Come help!” That does not prove crows have a human-like language. But AI could help scientists test whether repeated sounds consistently correspond to particular situations.<h2>Could AI let humans talk to animals?</h2>The potential is enormous. Researchers are combining huge amounts of audio and visual data to search for patterns across animal behaviour. Investor and animal-communication supporter Jeremy Coller has predicted that two-way communication could become possible by 2030.&nbsp;But animal communication is far more complex than speech. Animals can use sounds, smells, colours, movements and even sonar. That means simply translating animal sounds may not be enough. AI may need to understand the animal's surroundings and behaviour at the same time.<h2>The technology could also harm animals</h2>There is another side to the breakthrough.Experts warn that understanding animal communication could make it easier to manipulate animals at scale. César Rodríguez-Garavito, director of NYU’s More-Than-Human Life Program, has warned that AI could create “greater harm at scale”.&nbsp;Researchers are particularly concerned about playback experiments, in which recorded calls are played back to animals to study their reactions. Baglione has warned that such experiments can cause stress and disrupt social groups. The race to understand animals, therefore, comes with a difficult question: if AI can crack the code, who should be allowed to use it and how?

AI is helping scientists analyse huge collections of animal sounds, including crow calls, in a search for patterns humans cannot detect. Researchers hope this could eventually enable communication with animals, but experts warn the technology could also be misused.

Artificial intelligence is no longer being used only to analyse human language. Scientists are now using it to study how animals communicate, raising an extraordinary possibility: could AI eventually help humans understand or even communicate with other species? According to Bloomberg, researchers are using machine learning to process enormous collections of animal sounds that would be almost impossible to study manually. The work spans species including crows, whales and wolves, with scientists looking for patterns hidden inside thousands of recordings. AI helps decode crow calls Biologist Vittorio Baglione has spent around three decades studying carrion crows in Spain. In 2018, he and his research partner Daniela Canestrari fitted small microphones to 43 birds. The experiment produced thousands of hours of recordings involving adult crows, chicks and other birds. The collection has since grown to about 150,000 recordings. Researchers turned to an AI model developed by the Earth Species Project to separate sounds and identify which birds produced them. One call particularly interested Baglione. He observed that when a buzzard attacked a crow's nest, the bird produced a soft sound before other crows arrived to defend it. He interpreted the call as roughly meaning: “Hey! Come help!” That does not prove crows have a human-like language. But AI could help scientists test whether repeated sounds consistently correspond to particular situations. Could AI let humans talk to animals? The potential is enormous. Researchers are combining huge amounts of audio and visual data to search for patterns across animal behaviour. Investor and animal-communication supporter Jeremy Coller has predicted that two-way communication could become possible by 2030. But animal communication is far more complex than speech. Animals can use sounds, smells, colours, movements and even sonar. That means simply translating animal sounds may not be enough. AI may need to understand the animal's surroundings and behaviour at the same time. The technology could also harm animals There is another side to the breakthrough. Experts warn that understanding animal communication could make it easier to manipulate animals at scale. César Rodríguez-Garavito, director of NYU’s More-Than-Human Life Program, has warned that AI could create “greater harm at scale”. Researchers are particularly concerned about playback experiments, in which recorded calls are played back to animals to study their reactions. Baglione has warned that such experiments can cause stress and disrupt social groups. The race to understand animals, therefore, comes with a difficult question: if AI can crack the code, who should be allowed to use it and how?

AI is helping scientists analyse huge collections of animal sounds, including crow calls, in a search for patterns humans cannot detect. Researchers hope this could eventually enable communication with animals, but experts warn the technology could also be misused.

AI is decoding animal sounds: Is it possible we could soon talk to animals?

An American AI company has decided the easiest place to build a very large data centre is not America.&nbsp;The Deal&nbsp;Together AI, which serves open-source models to customers, announced a partnership on August 31 with HUMAIN, the Saudi state-backed artificial intelligence company, for a 250 megawatt facility in the kingdom.&nbsp;The site would house roughly 120,000 semiconductors. Together AI will use HUMAIN's chips and give HUMAIN a share of the revenue in return, rather than buying the hardware outright.&nbsp;The company, valued at $8.3 billion in July, says it expects the facility to produce more than $5 billion in annualised revenue. It describes the deal as among the largest AI infrastructure agreements for open source anywhere.&nbsp;Why Not America&nbsp;The stated reason is the interesting part, and it has been reported plainly rather than hinted at: building in Saudi Arabia lets the company bypass American backlash over data centres.&nbsp;That backlash is real and growing. Data centres consume enormous quantities of electricity and, in many designs, water. American communities have increasingly organised against them, utilities have struggled to connect them, and several states have seen local fights over land, grid capacity and residential power prices rising alongside new construction.&nbsp;For a company that needs 250 megawatts, those are not abstractions. They are years of permitting and the possibility of losing anyway.&nbsp;Saudi Arabia offers the inverse. Power is abundant and cheap, land is not contested in the same way, and the state is actively courting exactly this construction as part of its economic diversification away from oil.&nbsp;What Saudi Arabia Is Assembling&nbsp;This is not a single opportunistic deal. HUMAIN has been signing partners at pace.&nbsp;Microsoft announced an enhanced partnership with HUMAIN, including new enterprise AI offerings and an AI PC. Qualcomm has announced AI infrastructure work in the kingdom with the same company. HUMAIN has separately arranged a $1.2 billion financing package with Infra to fund 250 megawatts of data centre space, and is working with center3 on plans described as reaching a gigawatt.&nbsp;The pattern is a state-backed entity positioning itself as the counterparty for anyone who wants large-scale compute without the domestic difficulty of building it — and taking equity, revenue shares or both in exchange.&nbsp;The Question That Follows&nbsp;Compute is the strategic asset of this decade, and the United States has spent two years constructing an elaborate apparatus to control where it goes.&nbsp;Export controls determine which chips may be sold to whom. The Remote Access Security Act, passed by the House, would extend those controls to renting compute remotely, closing the gap that let Chinese laboratories reach restricted hardware through third countries. Washington has treated the physical location of advanced chips as a matter of national security.&nbsp;American companies voluntarily relocating their compute to a foreign state-backed partner sits awkwardly inside that framework. Not illegally — Saudi Arabia is an ally, and there is no suggestion this arrangement breaches any rule. But the policy logic that says it matters where chips sit does not obviously stop applying when the company moving them is American and the destination is friendly.&nbsp;The Open Source Dimension&nbsp;There is a further wrinkle specific to what Together AI does.&nbsp;The company serves open-weight models — systems whose weights are published for anyone to download and run. The argument for open models has always been that they prevent concentration: no single company controls who may use the capability.&nbsp;A 250 megawatt facility dedicated to serving those models, financed and hardware-supplied by a single state-backed entity taking a revenue share, is a considerable concentration of the infrastructure that open models actually run on. Open weights do not help if the practical ability to serve them at scale sits in a small number of hands.&nbsp;That is not an argument against the deal. It is a reminder that openness at the model layer and concentration at the infrastructure layer are entirely compatible, and that the second is where the leverage tends to end up.&nbsp;What To Watch&nbsp;Whether other American AI companies follow. The domestic constraints Together AI is routing around apply to everyone, and the Gulf states have capital, power and political will in a combination almost nowhere else offers.&nbsp;If a second and third company announce comparable facilities in the next two quarters, the question stops being about one deal and becomes about whether the next phase of American AI infrastructure gets built somewhere else entirely — and what Washington intends to do about that, having spent two years insisting that where compute lives is a matter of national security.

Together AI has partnered with Saudi state-backed HUMAIN on a 250MW data centre holding 120,000 chips, expecting over $5 billion in annualised revenue — built abroad partly to sidestep American opposition to data centres.

An American AI company has decided the easiest place to build a very large data centre is not America. The Deal Together AI, which serves open-source models to customers, announced a partnership on August 31 with HUMAIN, the Saudi state-backed artificial intelligence company, for a 250 megawatt facility in the kingdom. The site would house roughly 120,000 semiconductors. Together AI will use HUMAIN's chips and give HUMAIN a share of the revenue in return, rather than buying the hardware outright. The company, valued at $8.3 billion in July, says it expects the facility to produce more than $5 billion in annualised revenue. It describes the deal as among the largest AI infrastructure agreements for open source anywhere. Why Not America The stated reason is the interesting part, and it has been reported plainly rather than hinted at: building in Saudi Arabia lets the company bypass American backlash over data centres. That backlash is real and growing. Data centres consume enormous quantities of electricity and, in many designs, water. American communities have increasingly organised against them, utilities have struggled to connect them, and several states have seen local fights over land, grid capacity and residential power prices rising alongside new construction. For a company that needs 250 megawatts, those are not abstractions. They are years of permitting and the possibility of losing anyway. Saudi Arabia offers the inverse. Power is abundant and cheap, land is not contested in the same way, and the state is actively courting exactly this construction as part of its economic diversification away from oil. What Saudi Arabia Is Assembling This is not a single opportunistic deal. HUMAIN has been signing partners at pace. Microsoft announced an enhanced partnership with HUMAIN, including new enterprise AI offerings and an AI PC. Qualcomm has announced AI infrastructure work in the kingdom with the same company. HUMAIN has separately arranged a $1.2 billion financing package with Infra to fund 250 megawatts of data centre space, and is working with center3 on plans described as reaching a gigawatt. The pattern is a state-backed entity positioning itself as the counterparty for anyone who wants large-scale compute without the domestic difficulty of building it — and taking equity, revenue shares or both in exchange. The Question That Follows Compute is the strategic asset of this decade, and the United States has spent two years constructing an elaborate apparatus to control where it goes. Export controls determine which chips may be sold to whom. The Remote Access Security Act, passed by the House, would extend those controls to renting compute remotely, closing the gap that let Chinese laboratories reach restricted hardware through third countries. Washington has treated the physical location of advanced chips as a matter of national security. American companies voluntarily relocating their compute to a foreign state-backed partner sits awkwardly inside that framework. Not illegally — Saudi Arabia is an ally, and there is no suggestion this arrangement breaches any rule. But the policy logic that says it matters where chips sit does not obviously stop applying when the company moving them is American and the destination is friendly. The Open Source Dimension There is a further wrinkle specific to what Together AI does. The company serves open-weight models — systems whose weights are published for anyone to download and run. The argument for open models has always been that they prevent concentration: no single company controls who may use the capability. A 250 megawatt facility dedicated to serving those models, financed and hardware-supplied by a single state-backed entity taking a revenue share, is a considerable concentration of the infrastructure that open models actually run on. Open weights do not help if the practical ability to serve them at scale sits in a small number of hands. That is not an argument against the deal. It is a reminder that openness at the model layer and concentration at the infrastructure layer are entirely compatible, and that the second is where the leverage tends to end up. What To Watch Whether other American AI companies follow. The domestic constraints Together AI is routing around apply to everyone, and the Gulf states have capital, power and political will in a combination almost nowhere else offers. If a second and third company announce comparable facilities in the next two quarters, the question stops being about one deal and becomes about whether the next phase of American AI infrastructure gets built somewhere else entirely — and what Washington intends to do about that, having spent two years insisting that where compute lives is a matter of national security.

Together AI has signed a partnership with HUMAIN, the Saudi state-backed AI company, for a 250 megawatt data centre housing around 120,000 semiconductors. Together AI, valued at $8.3 billion in July, expects the facility to generate more than $5 billion in annualised revenue. It is among the largest data centres built specifically to serve open-source models — and it is being built abroad partly to sidestep domestic opposition.

An American AI company is building its 250MW data centre in Saudi Arabia to avoid the fight at home

DeepSeek is preparing to order at least 160,000 of Huawei's next-generation Ascend 950DT accelerators for a new gigawatt-scale data centre at Ulanqab in Inner Mongolia, according to a Bloomberg report.&nbsp;If it completes, it would be one of the largest known clusters of domestically manufactured Chinese AI accelerators anywhere — a roughly $2.56 billion order placed by China's most closely watched AI lab, with Nvidia locked out of the market.&nbsp;It reads as a decoupling milestone. The details complicate that reading considerably.&nbsp;The Caveat&nbsp;DeepSeek does not currently plan to use the 950DT chips for training.&nbsp;That is the part worth sitting with. Huawei designed and marketed the 950DT for training — the computationally brutal step where a model is actually built. DeepSeek intends to use them to run its models, not to create them, and has so far relied on Nvidia accelerators for training even after attempting to train earlier models on Huawei silicon.&nbsp;So the largest domestic-silicon order yet placed by a Chinese lab is, on current plans, an inference order. The dependency that actually constrains DeepSeek — access to Western chips for training — is not what this purchase resolves.&nbsp;The Second Constraint&nbsp;There is also a supply problem, and it is Huawei's.&nbsp;Component shortages, particularly for high-end memory, are reported to cap total 950DT output at the low hundreds of thousands of units this year. Against that ceiling, a single 160,000-unit order is enormous, and fulfilling it could stretch beyond twelve months.&nbsp;The data centre is scheduled to come online partially by late 2027 or early 2028 — a timeline that tells you this is a bet on capacity that does not exist yet, from a supplier that cannot currently build it at the required rate.&nbsp;Why It Still Matters&nbsp;None of which makes the order insignificant. Three things follow from it regardless.&nbsp;It commits Huawei's constrained output to a domestic customer, at a scale that squeezes everyone else in China queuing for the same chips. It gives Huawei a reference deployment at gigawatt scale, which is the credential its accelerator business has lacked. And it demonstrates that a serious Chinese lab will now plan multi-year infrastructure around domestic silicon rather than waiting for export policy to loosen.&nbsp;That last point is the durable one. Export controls were designed to make Chinese labs dependent on American hardware. A 160,000-chip order placed against a supplier that cannot yet deliver is not the behaviour of a company expecting that dependency to be restored.&nbsp;The Energy Question&nbsp;A gigawatt is the other number in this story, and it gets less attention than it deserves.&nbsp;That is a full nuclear reactor's worth of generation dedicated to a single facility. Inner Mongolia is a plausible site for exactly that reason — cheap land, cold air, and proximity to coal and wind generation — but a build of this scale is a regional energy commitment as much as a computing one.&nbsp;What To Watch&nbsp;Whether the order is confirmed at the reported volume or quietly reduced as Huawei's memory constraints bite. Whether DeepSeek's stated inference-only plan holds, or whether it shifts training onto the 950DT once the cluster exists — which would be the genuinely significant development. And whether other Chinese labs follow with orders of their own, or wait to see whether this one is delivered.&nbsp;For now the correct reading is narrower than the headline number suggests. This is a very large bet on Chinese silicon for the easier half of the problem.

DeepSeek is set to order 160,000 Huawei Ascend 950DT accelerators for a gigawatt data centre in Inner Mongolia. But it does not plan to train on them — and Huawei may not be able to build them fast enough.

DeepSeek is preparing to order at least 160,000 of Huawei's next-generation Ascend 950DT accelerators for a new gigawatt-scale data centre at Ulanqab in Inner Mongolia, according to a Bloomberg report. If it completes, it would be one of the largest known clusters of domestically manufactured Chinese AI accelerators anywhere — a roughly $2.56 billion order placed by China's most closely watched AI lab, with Nvidia locked out of the market. It reads as a decoupling milestone. The details complicate that reading considerably. The Caveat DeepSeek does not currently plan to use the 950DT chips for training. That is the part worth sitting with. Huawei designed and marketed the 950DT for training — the computationally brutal step where a model is actually built. DeepSeek intends to use them to run its models, not to create them, and has so far relied on Nvidia accelerators for training even after attempting to train earlier models on Huawei silicon. So the largest domestic-silicon order yet placed by a Chinese lab is, on current plans, an inference order. The dependency that actually constrains DeepSeek — access to Western chips for training — is not what this purchase resolves. The Second Constraint There is also a supply problem, and it is Huawei's. Component shortages, particularly for high-end memory, are reported to cap total 950DT output at the low hundreds of thousands of units this year. Against that ceiling, a single 160,000-unit order is enormous, and fulfilling it could stretch beyond twelve months. The data centre is scheduled to come online partially by late 2027 or early 2028 — a timeline that tells you this is a bet on capacity that does not exist yet, from a supplier that cannot currently build it at the required rate. Why It Still Matters None of which makes the order insignificant. Three things follow from it regardless. It commits Huawei's constrained output to a domestic customer, at a scale that squeezes everyone else in China queuing for the same chips. It gives Huawei a reference deployment at gigawatt scale, which is the credential its accelerator business has lacked. And it demonstrates that a serious Chinese lab will now plan multi-year infrastructure around domestic silicon rather than waiting for export policy to loosen. That last point is the durable one. Export controls were designed to make Chinese labs dependent on American hardware. A 160,000-chip order placed against a supplier that cannot yet deliver is not the behaviour of a company expecting that dependency to be restored. The Energy Question A gigawatt is the other number in this story, and it gets less attention than it deserves. That is a full nuclear reactor's worth of generation dedicated to a single facility. Inner Mongolia is a plausible site for exactly that reason — cheap land, cold air, and proximity to coal and wind generation — but a build of this scale is a regional energy commitment as much as a computing one. What To Watch Whether the order is confirmed at the reported volume or quietly reduced as Huawei's memory constraints bite. Whether DeepSeek's stated inference-only plan holds, or whether it shifts training onto the 950DT once the cluster exists — which would be the genuinely significant development. And whether other Chinese labs follow with orders of their own, or wait to see whether this one is delivered. For now the correct reading is narrower than the headline number suggests. This is a very large bet on Chinese silicon for the easier half of the problem.

DeepSeek is preparing to order at least 160,000 of Huawei's next-generation Ascend 950DT accelerators for a gigawatt-scale data centre in Inner Mongolia, in what would be one of the largest known clusters of Chinese-made AI silicon. But the company does not currently plan to use them for training — the step Huawei built the chip to handle.

DeepSeek is buying 160,000 Huawei chips, and the detail that matters is what it will not use them for

Export licences for graphics processors are not a traditional instrument of peacemaking. According to a Wall Street Journal report, they were used as one in the South Caucasus.&nbsp;US negotiators are said to have used the promise of access to Nvidia Blackwell processors for an Armenian AI data centre as an incentive in the talks that produced a provisional peace agreement between Armenia and Azerbaijan, signed at the White House in August.&nbsp;The Armenian side has not treated this as an awkward revelation. Prime Minister Nikol Pashinyan has said publicly that he personally lobbied President Donald Trump to expedite the chip-transfer approvals, and framed the matter as part of the peace agreement itself.&nbsp;The Project&nbsp;The data centre at the centre of this is real and already approved.&nbsp;The AI startup Firebird received US government clearance to export Nvidia chips to Armenia for a supercomputer project — a 100-megawatt facility backed by an initial $500 million investment, using Dell servers and Nvidia's Blackwell processors. The first phase was slated for mid-2026.&nbsp;For Armenia, a landlocked country of under three million people, a facility of that size is not an incremental technology investment. It is an attempt to become infrastructure that other countries need.&nbsp;What Makes This Unusual&nbsp;Export controls on advanced chips have been used as a stick for years. Washington restricts what Nvidia can sell to China, and the policy is straightforwardly punitive in intent.&nbsp;Using the same controls as a carrot is a different proposition. It treats an export licence as a sovereign favour to be traded — which, functionally, is what it has become. The scarcity is manufactured by policy, which means the ability to relieve it is a diplomatic asset that costs the United States nothing to create.&nbsp;That is an efficient form of leverage. It is also one that only works while the scarcity holds.&nbsp;The Uncomfortable Part&nbsp;There is a reading of this that is genuinely positive: a peace agreement was signed, and if compute access helped get it there, that is a cheaper inducement than most of the alternatives available to American diplomacy.&nbsp;The harder question is what it establishes. If chip access is a bargaining chip in territorial negotiations, then the allocation of AI infrastructure becomes a function of geopolitical alignment rather than economic demand — and countries learn that the route to compute runs through Washington's foreign policy priorities.&nbsp;It also creates a dependency with a short half-life. An export approval can be withdrawn, and a 100-megawatt facility built around a specific vendor's processors is not easily repurposed if the political weather changes.&nbsp;The Physical Cost&nbsp;The facility's demands are worth stating plainly, because they are rarely part of the diplomatic conversation.&nbsp;A 100-megawatt data centre consumes roughly the annual electricity of a city of 120,000 people, and cooling it can require more than 500,000 gallons of water per day. In a region where water allocation is itself a source of tension, that is not a neutral footprint.&nbsp;What To Watch&nbsp;Whether the provisional agreement holds, and whether the chip approvals survive it if it does not. Whether Firebird's first phase actually comes online on schedule. And whether other governments now approach Washington with the same trade in mind — because the precedent, once set, is available to anyone with a border dispute and a plot of land.&nbsp;The lesson other capitals will take from this is not about Armenia. It is that compute is now something you can negotiate for.

US negotiators used the promise of Nvidia Blackwell access for an Armenian AI data centre as an incentive in the Armenia-Azerbaijan peace talks, the WSJ reports. Armenia's PM says he lobbied Trump directly.

Export licences for graphics processors are not a traditional instrument of peacemaking. According to a Wall Street Journal report, they were used as one in the South Caucasus. US negotiators are said to have used the promise of access to Nvidia Blackwell processors for an Armenian AI data centre as an incentive in the talks that produced a provisional peace agreement between Armenia and Azerbaijan, signed at the White House in August. The Armenian side has not treated this as an awkward revelation. Prime Minister Nikol Pashinyan has said publicly that he personally lobbied President Donald Trump to expedite the chip-transfer approvals, and framed the matter as part of the peace agreement itself. The Project The data centre at the centre of this is real and already approved. The AI startup Firebird received US government clearance to export Nvidia chips to Armenia for a supercomputer project — a 100-megawatt facility backed by an initial $500 million investment, using Dell servers and Nvidia's Blackwell processors. The first phase was slated for mid-2026. For Armenia, a landlocked country of under three million people, a facility of that size is not an incremental technology investment. It is an attempt to become infrastructure that other countries need. What Makes This Unusual Export controls on advanced chips have been used as a stick for years. Washington restricts what Nvidia can sell to China, and the policy is straightforwardly punitive in intent. Using the same controls as a carrot is a different proposition. It treats an export licence as a sovereign favour to be traded — which, functionally, is what it has become. The scarcity is manufactured by policy, which means the ability to relieve it is a diplomatic asset that costs the United States nothing to create. That is an efficient form of leverage. It is also one that only works while the scarcity holds. The Uncomfortable Part There is a reading of this that is genuinely positive: a peace agreement was signed, and if compute access helped get it there, that is a cheaper inducement than most of the alternatives available to American diplomacy. The harder question is what it establishes. If chip access is a bargaining chip in territorial negotiations, then the allocation of AI infrastructure becomes a function of geopolitical alignment rather than economic demand — and countries learn that the route to compute runs through Washington's foreign policy priorities. It also creates a dependency with a short half-life. An export approval can be withdrawn, and a 100-megawatt facility built around a specific vendor's processors is not easily repurposed if the political weather changes. The Physical Cost The facility's demands are worth stating plainly, because they are rarely part of the diplomatic conversation. A 100-megawatt data centre consumes roughly the annual electricity of a city of 120,000 people, and cooling it can require more than 500,000 gallons of water per day. In a region where water allocation is itself a source of tension, that is not a neutral footprint. What To Watch Whether the provisional agreement holds, and whether the chip approvals survive it if it does not. Whether Firebird's first phase actually comes online on schedule. And whether other governments now approach Washington with the same trade in mind — because the precedent, once set, is available to anyone with a border dispute and a plot of land. The lesson other capitals will take from this is not about Armenia. It is that compute is now something you can negotiate for.

US negotiators used the promise of Nvidia Blackwell access for an Armenian AI data centre as an incentive in the talks that produced a provisional Armenia-Azerbaijan peace agreement, according to a Wall Street Journal report. Armenia's prime minister has said he personally lobbied Donald Trump to speed the chip approvals, framing the request as part of the deal.

Nvidia chips were reportedly used as leverage to help broker a peace deal in the Caucasus

Anthropic has confirmed that it submitted a draft registration statement on Form S-1 to the Securities and Exchange Commission on a confidential basis. Reporting around that filing points to a Nasdaq listing as soon as October, with Goldman Sachs, JPMorgan and Morgan Stanley leading an offering expected to raise more than $60 billion.&nbsp;If it completes near those figures it would be among the largest public offerings in the history of the stock market.&nbsp;A Caution On The Numbers&nbsp;The valuations attached to this listing deserve scepticism, and the range is the reason.&nbsp;Google's investment earlier in 2026 was struck at a reported $350 billion valuation. A Series H-1 in May has been reported at $965 billion. The New York Times reported in late August that an eventual IPO could value the company at $2 trillion.&nbsp;Those figures span a factor of nearly six within a single year. Some of that is genuine repricing in a fast-moving market, but a spread that wide across sources and dates is not a valuation — it is a range of opinions about one, and no single number in it should be reported as settled.&nbsp;Because the S-1 was submitted confidentially, it is not available on EDGAR. Anthropic has not set a share price, a share count, or a listing date, and has said the timing and completion of any offering depend on market conditions.&nbsp;The Risk Factor&nbsp;The detail that makes this filing genuinely unusual was reported by CNBC in late August: the S-1 is expected to name public backlash against AI among its risk factors.&nbsp;Risk-factor sections are typically defensive boilerplate, drafted to inoculate against future litigation. But this one is close to the company's foundational premise. Anthropic has positioned itself for years as the lab that takes AI risk seriously — and it is now, reportedly, obliged to tell prospective shareholders that public hostility to its industry could damage the business.&nbsp;There is no contradiction in that, strictly speaking. A company can believe the technology is dangerous, believe it should be built carefully, and disclose that others disagree. But it does compress an unusual amount of tension into a regulatory document.&nbsp;What The Filing Will Actually Reveal&nbsp;The financial disclosure is the part worth waiting for.&nbsp;Anthropic's annualised revenue was reported at around $30 billion as of April 2026. What has never been public is the cost of producing it: compute spend, gross margin, the economics of serving frontier models at scale. Private rounds do not require that; a public listing does.&nbsp;The same filing will settle a related question. Google and Amazon have both taken large positions in Anthropic — Google committing up to $40 billion in total, Amazon roughly $33 billion — and their stakes have contributed materially to reported profits at both companies. An IPO puts a market price on those holdings rather than a negotiated one.&nbsp;Why Now&nbsp;Two readings, and both are defensible.&nbsp;The optimistic one: the company has the revenue trajectory to justify a listing and is choosing to raise while conditions are favourable, funding compute commitments that run to many tens of billions.&nbsp;The sceptical one: private valuations have run far ahead of what late-stage investors will keep funding, and a public listing transfers that risk to a broader base of buyers. Companies rarely list at the bottom of their own perceived value.&nbsp;What To Watch&nbsp;Whether the public S-1 arrives before the October window and what gross margin it discloses. Whether the AI-backlash risk factor survives into the final document or is softened. And whether the listing prices anywhere near the figures now circulating — because the gap between $350 billion and $2 trillion is not a rounding error, and the market is about to resolve it.

Anthropic has confirmed a confidential S-1 with the SEC, with reporting pointing to an October Nasdaq listing raising $60bn+. Quoted valuations range from $350bn to $2 trillion.

Anthropic has confirmed that it submitted a draft registration statement on Form S-1 to the Securities and Exchange Commission on a confidential basis. Reporting around that filing points to a Nasdaq listing as soon as October, with Goldman Sachs, JPMorgan and Morgan Stanley leading an offering expected to raise more than $60 billion. If it completes near those figures it would be among the largest public offerings in the history of the stock market. A Caution On The Numbers The valuations attached to this listing deserve scepticism, and the range is the reason. Google's investment earlier in 2026 was struck at a reported $350 billion valuation. A Series H-1 in May has been reported at $965 billion. The New York Times reported in late August that an eventual IPO could value the company at $2 trillion. Those figures span a factor of nearly six within a single year. Some of that is genuine repricing in a fast-moving market, but a spread that wide across sources and dates is not a valuation — it is a range of opinions about one, and no single number in it should be reported as settled. Because the S-1 was submitted confidentially, it is not available on EDGAR. Anthropic has not set a share price, a share count, or a listing date, and has said the timing and completion of any offering depend on market conditions. The Risk Factor The detail that makes this filing genuinely unusual was reported by CNBC in late August: the S-1 is expected to name public backlash against AI among its risk factors. Risk-factor sections are typically defensive boilerplate, drafted to inoculate against future litigation. But this one is close to the company's foundational premise. Anthropic has positioned itself for years as the lab that takes AI risk seriously — and it is now, reportedly, obliged to tell prospective shareholders that public hostility to its industry could damage the business. There is no contradiction in that, strictly speaking. A company can believe the technology is dangerous, believe it should be built carefully, and disclose that others disagree. But it does compress an unusual amount of tension into a regulatory document. What The Filing Will Actually Reveal The financial disclosure is the part worth waiting for. Anthropic's annualised revenue was reported at around $30 billion as of April 2026. What has never been public is the cost of producing it: compute spend, gross margin, the economics of serving frontier models at scale. Private rounds do not require that; a public listing does. The same filing will settle a related question. Google and Amazon have both taken large positions in Anthropic — Google committing up to $40 billion in total, Amazon roughly $33 billion — and their stakes have contributed materially to reported profits at both companies. An IPO puts a market price on those holdings rather than a negotiated one. Why Now Two readings, and both are defensible. The optimistic one: the company has the revenue trajectory to justify a listing and is choosing to raise while conditions are favourable, funding compute commitments that run to many tens of billions. The sceptical one: private valuations have run far ahead of what late-stage investors will keep funding, and a public listing transfers that risk to a broader base of buyers. Companies rarely list at the bottom of their own perceived value. What To Watch Whether the public S-1 arrives before the October window and what gross margin it discloses. Whether the AI-backlash risk factor survives into the final document or is softened. And whether the listing prices anywhere near the figures now circulating — because the gap between $350 billion and $2 trillion is not a rounding error, and the market is about to resolve it.

Anthropic has confirmed it submitted a confidential draft S-1 to the SEC, with reporting pointing to an October Nasdaq listing that could raise more than $60 billion. The valuations attached to it vary enormously depending on the source and the date — and the filing is reported to name public hostility to AI among its risk factors.

A company built on AI safety is about to ask public markets to price it

Microsoft AI has released MAI-Transcribe-2 with an unusually blunt claim attached: that it is the fastest, most accurate and cheapest speech recognition model in the world.&nbsp;The price is $0.10 per hour of audio — around $1.67 per thousand minutes — and it undercuts the transcription services offered by OpenAI, Google and ElevenLabs.&nbsp;The Number Behind The Number&nbsp;The more revealing comparison is internal.&nbsp;MAI-Transcribe-1 launched at $0.36 per hour, itself considered competitive at the time. The new model arrives at roughly 72 per cent below that, from the same company, within a single product generation.&nbsp;That is not a response to a competitor undercutting Microsoft. It is Microsoft undercutting itself, which is what a company does when it has decided the category should not be a profit centre at all.&nbsp;The Expiry Date&nbsp;One detail deserves more prominence than it is getting: $0.10 is an introductory price that runs only through the end of 2026.&nbsp;Microsoft has not said what the model will cost afterwards. Anyone building a product on that rate is building on a number with a known expiry and no announced successor — and transcription is exactly the kind of dependency that is easy to adopt and awkward to migrate off once it is embedded in a workflow.&nbsp;This is a familiar shape. Price low enough to move the market, let the integrations accumulate, then set the real rate once switching has become someone else's engineering problem.&nbsp;That is not an accusation of bad faith. It is the standard playbook, and it is worth naming while the discount is still in effect rather than after.&nbsp;Where It Runs&nbsp;The model is available now through Microsoft Foundry and the MAI Playground, with OpenRouter availability described as coming soon.&nbsp;Reports of the team size behind it — one account puts it at around ten people — and of speed gains of roughly ten times over OpenAI's comparable offering come from coverage of the launch rather than from independently verified benchmarks, and should be treated as claims rather than settled numbers until third parties have run their own tests.&nbsp;Why Microsoft Would Do This&nbsp;Transcription is not a business Microsoft needs to win on margin. It is a feature that makes other things work: meeting summaries in Teams, accessibility across Office, call handling in customer service, and the voice layer of any agent that has to listen before it acts.&nbsp;Priced near zero, it becomes a reason to stay inside Microsoft's stack rather than a line item customers evaluate on its own. Priced at $0.36, it was a purchase decision. At $0.10, it is closer to an assumption.&nbsp;The competitive effect lands hardest on companies for which speech recognition is the entire business rather than a supporting feature. ElevenLabs and similar specialists cannot subsidise transcription from an operating system and a cloud contract.&nbsp;What To Watch&nbsp;What the post-introductory price turns out to be, and how much notice customers get. Whether independent benchmarks support the accuracy claim once researchers outside Microsoft can test it. And whether OpenAI and Google respond with cuts of their own — which would confirm that speech recognition has finished its transition from product to utility.&nbsp;For buyers, the practical advice is narrow: the rate is real, it is good, and it is dated. Plan for what happens in January.

MAI-Transcribe-2 costs $0.10 per audio hour, 72% below Microsoft's own previous model, undercutting OpenAI, Google and ElevenLabs. The rate is introductory and runs only through the end of 2026.

Microsoft AI has released MAI-Transcribe-2 with an unusually blunt claim attached: that it is the fastest, most accurate and cheapest speech recognition model in the world. The price is $0.10 per hour of audio — around $1.67 per thousand minutes — and it undercuts the transcription services offered by OpenAI, Google and ElevenLabs. The Number Behind The Number The more revealing comparison is internal. MAI-Transcribe-1 launched at $0.36 per hour, itself considered competitive at the time. The new model arrives at roughly 72 per cent below that, from the same company, within a single product generation. That is not a response to a competitor undercutting Microsoft. It is Microsoft undercutting itself, which is what a company does when it has decided the category should not be a profit centre at all. The Expiry Date One detail deserves more prominence than it is getting: $0.10 is an introductory price that runs only through the end of 2026. Microsoft has not said what the model will cost afterwards. Anyone building a product on that rate is building on a number with a known expiry and no announced successor — and transcription is exactly the kind of dependency that is easy to adopt and awkward to migrate off once it is embedded in a workflow. This is a familiar shape. Price low enough to move the market, let the integrations accumulate, then set the real rate once switching has become someone else's engineering problem. That is not an accusation of bad faith. It is the standard playbook, and it is worth naming while the discount is still in effect rather than after. Where It Runs The model is available now through Microsoft Foundry and the MAI Playground, with OpenRouter availability described as coming soon. Reports of the team size behind it — one account puts it at around ten people — and of speed gains of roughly ten times over OpenAI's comparable offering come from coverage of the launch rather than from independently verified benchmarks, and should be treated as claims rather than settled numbers until third parties have run their own tests. Why Microsoft Would Do This Transcription is not a business Microsoft needs to win on margin. It is a feature that makes other things work: meeting summaries in Teams, accessibility across Office, call handling in customer service, and the voice layer of any agent that has to listen before it acts. Priced near zero, it becomes a reason to stay inside Microsoft's stack rather than a line item customers evaluate on its own. Priced at $0.36, it was a purchase decision. At $0.10, it is closer to an assumption. The competitive effect lands hardest on companies for which speech recognition is the entire business rather than a supporting feature. ElevenLabs and similar specialists cannot subsidise transcription from an operating system and a cloud contract. What To Watch What the post-introductory price turns out to be, and how much notice customers get. Whether independent benchmarks support the accuracy claim once researchers outside Microsoft can test it. And whether OpenAI and Google respond with cuts of their own — which would confirm that speech recognition has finished its transition from product to utility. For buyers, the practical advice is narrow: the rate is real, it is good, and it is dated. Plan for what happens in January.

Microsoft AI has released MAI-Transcribe-2, claiming it is the fastest, most accurate and cheapest speech recognition model available, at $0.10 per hour of audio. That is roughly 72 per cent below what its own predecessor launched at — but the company has confirmed the rate is introductory and only holds through the end of 2026.

Microsoft cut AI transcription pricing by 72 per cent, and the rate expires in December

AMD used its IFA 2026 keynote to announce Threadripper Halo Station, a workstation built around a claim it is happy to make loudly: that it is the most powerful workstation in the world.&nbsp;The specification is the argument. A 96-core, 192-thread Threadripper PRO 9995WX on Zen 5, boosting to 5.4GHz, supporting up to 2TB of DDR5 system memory — and up to four Instinct MI350P accelerators, each carrying 144GB of HBM3E at 4TB/s.&nbsp;At four cards that is 576GB of high-bandwidth GPU memory in a single desk-side chassis. AMD's presented configuration uses two cards for 288GB, with the upgrade path to four.&nbsp;Everything significant in the system is liquid cooled, which AMD frames as a noise decision as much as a thermal one. A machine intended to sit next to a person has different constraints from one in a rack.&nbsp;The One Use Case&nbsp;AMD has been unusually direct about what this is for. The system is designed around running AI models with more than one trillion parameters locally, without a cloud connection.&nbsp;The memory figure is what makes that plausible. Model weights have to fit somewhere, and at 4-bit precision a trillion-parameter model lands within reach of 576GB. The MI350P's bandwidth — which AMD puts at fourteen times that of any LPDDR5X variant — is the difference between a model that technically loads and one that responds at usable speed.&nbsp;This is the specification that determines whether local inference at frontier scale is real or theatrical, and AMD has chosen to compete on exactly that number.&nbsp;The Target&nbsp;The comparison AMD wants is with Nvidia's DGX Station, and the positioning is transparent.&nbsp;Nvidia has spent years making the desk-side AI workstation its own category, and AMD's pitch is that supercomputer-class compute should be available to an individual developer rather than rationed through a cloud queue. Whether that lands depends almost entirely on software — Nvidia's advantage has rarely been silicon alone, and CUDA remains the reason many teams do not seriously evaluate alternatives.&nbsp;Hardware parity does not resolve that. It only makes the question worth asking again.&nbsp;Who It Is Actually For&nbsp;The honest audience for this machine is narrow, and worth naming.&nbsp;It is for teams working on data they are not permitted to send anywhere — medical records, defence work, unreleased financials, anything under a jurisdictional constraint. It is for researchers whose iteration speed is bottlenecked by cloud scheduling rather than raw compute. And it is for organisations that have run the arithmetic on sustained inference and found that renting indefinitely costs more than owning.&nbsp;It is not for most developers, who will continue to rent because renting is cheaper until utilisation is high and constant.&nbsp;The Missing Number&nbsp;AMD has not announced pricing, availability, or complete system specifications.&nbsp;That absence matters more than usual here, because the entire economic case rests on it. A local machine competes with a cloud subscription, and without a price the comparison cannot be made. Four MI350P cards represent serious silicon, and the plausible range for a fully configured system starts well above what most individual developers will fund themselves.&nbsp;Until AMD publishes a number, 'available to individual developers' is a positioning statement rather than a claim that can be tested.&nbsp;What To Watch&nbsp;The price, first and above all. Whether the four-card configuration ships at launch or arrives later. And whether the software stack has closed enough of the gap that a team can move a working pipeline across without rewriting it — because that, not memory capacity, is what has kept this category Nvidia's.

Threadripper Halo Station pairs a 96-core Threadripper PRO with up to four MI350P accelerators and 576GB of HBM3E, aimed at Nvidia's DGX Station. AMD has not announced what it costs.

AMD used its IFA 2026 keynote to announce Threadripper Halo Station, a workstation built around a claim it is happy to make loudly: that it is the most powerful workstation in the world. The specification is the argument. A 96-core, 192-thread Threadripper PRO 9995WX on Zen 5, boosting to 5.4GHz, supporting up to 2TB of DDR5 system memory — and up to four Instinct MI350P accelerators, each carrying 144GB of HBM3E at 4TB/s. At four cards that is 576GB of high-bandwidth GPU memory in a single desk-side chassis. AMD's presented configuration uses two cards for 288GB, with the upgrade path to four. Everything significant in the system is liquid cooled, which AMD frames as a noise decision as much as a thermal one. A machine intended to sit next to a person has different constraints from one in a rack. The One Use Case AMD has been unusually direct about what this is for. The system is designed around running AI models with more than one trillion parameters locally, without a cloud connection. The memory figure is what makes that plausible. Model weights have to fit somewhere, and at 4-bit precision a trillion-parameter model lands within reach of 576GB. The MI350P's bandwidth — which AMD puts at fourteen times that of any LPDDR5X variant — is the difference between a model that technically loads and one that responds at usable speed. This is the specification that determines whether local inference at frontier scale is real or theatrical, and AMD has chosen to compete on exactly that number. The Target The comparison AMD wants is with Nvidia's DGX Station, and the positioning is transparent. Nvidia has spent years making the desk-side AI workstation its own category, and AMD's pitch is that supercomputer-class compute should be available to an individual developer rather than rationed through a cloud queue. Whether that lands depends almost entirely on software — Nvidia's advantage has rarely been silicon alone, and CUDA remains the reason many teams do not seriously evaluate alternatives. Hardware parity does not resolve that. It only makes the question worth asking again. Who It Is Actually For The honest audience for this machine is narrow, and worth naming. It is for teams working on data they are not permitted to send anywhere — medical records, defence work, unreleased financials, anything under a jurisdictional constraint. It is for researchers whose iteration speed is bottlenecked by cloud scheduling rather than raw compute. And it is for organisations that have run the arithmetic on sustained inference and found that renting indefinitely costs more than owning. It is not for most developers, who will continue to rent because renting is cheaper until utilisation is high and constant. The Missing Number AMD has not announced pricing, availability, or complete system specifications. That absence matters more than usual here, because the entire economic case rests on it. A local machine competes with a cloud subscription, and without a price the comparison cannot be made. Four MI350P cards represent serious silicon, and the plausible range for a fully configured system starts well above what most individual developers will fund themselves. Until AMD publishes a number, 'available to individual developers' is a positioning statement rather than a claim that can be tested. What To Watch The price, first and above all. Whether the four-card configuration ships at launch or arrives later. And whether the software stack has closed enough of the gap that a team can move a working pipeline across without rewriting it — because that, not memory capacity, is what has kept this category Nvidia's.

AMD used its IFA 2026 keynote to unveil Threadripper Halo Station, a liquid-cooled workstation pairing a 96-core Threadripper PRO 9995WX with up to four Instinct MI350P accelerators and 576GB of HBM3E. It is designed for one job: running models above a trillion parameters locally, with no cloud connection. Pricing has not been announced.

576GB of HBM3E under a desk: AMD's Halo Station is aimed squarely at Nvidia's DGX

Anthropic has agreed to a $35 billion cloud computing deal with Lambda, a cloud provider backed by Nvidia, to secure access to artificial intelligence computing capacity at a data center under development in Nueces County, Texas, according to people familiar with the matter cited in reporting first published by Bloomberg on August 31, 2026.&nbsp;The arrangement was described by sources rather than confirmed in an on-the-record joint statement. A representative for Anthropic declined to comment when approached by reporters, and Nvidia, Lambda and the data center's developer, Hut 8, did not immediately respond to requests for comment, according to the reporting.&nbsp;How the deal is structured&nbsp;The agreement ties together several companies. Hut 8 is developing the data center campus, known as Beacon Point, on roughly 1 gigawatt of secured utility capacity in Nueces County. Nvidia holds two 15-year leases on the campus covering about 704 megawatts of contracted capacity split across two phases of roughly 352 megawatts each, a lease arrangement first reported by the Wall Street Journal with a base value of about $19.6 billion that could grow toward $50 billion over its full term with renewals.&nbsp;Anthropic's new $35 billion deal with Lambda draws on roughly 350 megawatts of that capacity, according to the reporting -- effectively one phase of the Beacon Point site. Under the structure, Lambda installs Nvidia chips at the facility and sells the resulting computing capacity to Anthropic, while Nvidia itself sits above the arrangement as both chip supplier and the entity that leased the underlying data center space from Hut 8. Initial energization of the campus is expected in the first quarter of 2027.&nbsp;Lambda, for its part, is reportedly in the process of raising up to $3 billion in new funding at a valuation of $12 billion or higher, after previously raising $1.5 billion in a funding round last November.&nbsp;A rapid run of mega deals&nbsp;The Lambda agreement is Anthropic's second major cloud deal within days. On August 26, 2026, Bloomberg and CNBC reported that Anthropic had struck a roughly $45 billion, six-year deal with Nscale, a UK-based AI infrastructure company, to rent about 460 megawatts of capacity from a West Virginia data center expected to use Nvidia's Vera Rubin chips and come online by the end of 2027.&nbsp;Those two deals sit alongside a broader run of infrastructure commitments Anthropic has made in 2026: a roughly $50 billion partnership with Fluidstack covering data centers in Texas and New York, a roughly $45 billion computing arrangement with Elon Musk's SpaceX disclosed in May, and a roughly $10 billion, six-year compute deal with Volta Infra Holdings, a cloud startup that was only founded months earlier. Taken together, the four previously reported deals alone add up to more than $150 billion in compute commitments, a figure that grows further once the new Lambda agreement is included.&nbsp;The buildout follows a period earlier in the year when Anthropic faced a compute supply shortage as demand for its Claude models, including its Claude Code coding tool, outpaced the capacity it had secured. The company has since been signing capacity deals with a widening circle of cloud and infrastructure providers, several of them, like Lambda and Nscale, themselves backed by Nvidia investment.&nbsp;Revenue growth behind the spending&nbsp;The infrastructure spree comes as Anthropic's own financial results have surged. CNBC reported on August 15, 2026, citing documents viewed by Bloomberg News, that Anthropic's preliminary revenue for the second quarter of 2026 topped $11.5 billion, up from about $787 million in the same quarter a year earlier -- growth of more than fourteen times. The company also reported positive adjusted operating income for the quarter, according to the same reporting. The figures were described as preliminary and could still change.&nbsp;Anthropic is preparing for a stock market listing, having submitted confidential IPO paperwork in June 2026, according to reporting on the company's plans. The compute commitments are widely seen by industry observers as necessary groundwork for sustaining the growth Anthropic is showing investors ahead of that listing.&nbsp;Questions about circular financing&nbsp;The structure of the Lambda deal, in which Nvidia supplies the chips, holds an equity stake in the cloud provider buying those chips, and separately leases the data center where they will run, has revived a debate in the AI industry over so-called circular financing: arrangements in which a chipmaker's investments and lease commitments effectively help fund the very customers that then purchase its hardware.&nbsp;Mizuho chip analyst Jordan Klein described the financing model underpinning these deals as concerning, characterizing it as a structure in which Nvidia GPUs are effectively used as collateral and rented back into the AI supply chain, with lease payments from tenants like Lambda and Anthropic ultimately used to repay the debt raised to build the data centers -- a model he said functions much like Nvidia pre-funding purchases of its own products.&nbsp;Bank of America analyst Vivek Arya has separately pointed to a related tension: many of Nvidia's investments are designed to support frontier AI labs including OpenAI and Anthropic, even as both companies are simultaneously developing their own custom chips that could eventually reduce their reliance on Nvidia hardware.&nbsp;Nvidia's finance chief, Colette Kress, has acknowledged the criticism directly, saying the company recognizes the scale of its support for AI labs and knows some will call it circular financing. Her defense rests on the argument that Nvidia's computing platform is fungible and durable and can be redeployed to support other customers even if a particular lab's demand changes, rather than being tied permanently to any single counterparty. Nvidia has disclosed it had closed a recent quarter with $279 billion in supply obligations and $108.5 billion in guarantee obligations tied to AI cloud partners, and Kress has said the company has invested nearly $50 billion across frontier AI labs.&nbsp;The case for genuine demand&nbsp;Anthropic's defenders of the spending point to a different set of numbers: the company's revenue climbing from roughly $787 million to more than $11.5 billion in a single year is an independently reported, verifiable growth rate that predates and is separate from any of the Nvidia-linked financing structures now drawing scrutiny. By this reading, Anthropic's compute needs are being driven by real, paying customer demand for Claude and related products such as Claude Code, and the string of infrastructure deals -- however entangled with Nvidia's own balance sheet -- reflects a company scrambling to keep pace with genuine usage rather than financial engineering designed to manufacture the appearance of demand.&nbsp;Skeptics counter that rapid revenue growth and circular financing concerns are not mutually exclusive: even a company with real underlying demand can end up locked into deals whose true economics are harder to assess when the counterparty supplying the chips is also, in effect, backing the buyer and the landlord. Whether the compute Anthropic is contracting for ultimately gets used at the scale implied by these multibillion-dollar commitments will not be fully clear until the facilities involved, including the Nueces County campus, are built out and running.&nbsp;What happens next&nbsp;None of the companies involved -- Anthropic, Nvidia, Lambda or Hut 8 -- has issued an on-the-record confirmation of the deal's exact terms, and the reporting to date rests on people described as familiar with private discussions. Initial capacity at the Beacon Point campus is not expected to come online until the first quarter of 2027, meaning the practical impact of the Lambda agreement on Anthropic's available compute will not be visible for several months.&nbsp;For now, the deal adds to a picture of an AI industry in which chipmaker, cloud provider, landlord and end customer are increasingly the same small set of interconnected companies -- a structure that has so far coincided with, but not conclusively explained, some of the fastest revenue growth reported anywhere in the technology sector.

Anthropic has struck a $35 billion cloud computing deal with Nvidia-backed Lambda for capacity at a Texas data center, its second mega compute deal in a week, as analysts debate whether the Nvidia-linked financing structure is circular.

Anthropic has agreed to a $35 billion cloud computing deal with Lambda, a cloud provider backed by Nvidia, to secure access to artificial intelligence computing capacity at a data center under development in Nueces County, Texas, according to people familiar with the matter cited in reporting first published by Bloomberg on August 31, 2026. The arrangement was described by sources rather than confirmed in an on-the-record joint statement. A representative for Anthropic declined to comment when approached by reporters, and Nvidia, Lambda and the data center's developer, Hut 8, did not immediately respond to requests for comment, according to the reporting. How the deal is structured The agreement ties together several companies. Hut 8 is developing the data center campus, known as Beacon Point, on roughly 1 gigawatt of secured utility capacity in Nueces County. Nvidia holds two 15-year leases on the campus covering about 704 megawatts of contracted capacity split across two phases of roughly 352 megawatts each, a lease arrangement first reported by the Wall Street Journal with a base value of about $19.6 billion that could grow toward $50 billion over its full term with renewals. Anthropic's new $35 billion deal with Lambda draws on roughly 350 megawatts of that capacity, according to the reporting -- effectively one phase of the Beacon Point site. Under the structure, Lambda installs Nvidia chips at the facility and sells the resulting computing capacity to Anthropic, while Nvidia itself sits above the arrangement as both chip supplier and the entity that leased the underlying data center space from Hut 8. Initial energization of the campus is expected in the first quarter of 2027. Lambda, for its part, is reportedly in the process of raising up to $3 billion in new funding at a valuation of $12 billion or higher, after previously raising $1.5 billion in a funding round last November. A rapid run of mega deals The Lambda agreement is Anthropic's second major cloud deal within days. On August 26, 2026, Bloomberg and CNBC reported that Anthropic had struck a roughly $45 billion, six-year deal with Nscale, a UK-based AI infrastructure company, to rent about 460 megawatts of capacity from a West Virginia data center expected to use Nvidia's Vera Rubin chips and come online by the end of 2027. Those two deals sit alongside a broader run of infrastructure commitments Anthropic has made in 2026: a roughly $50 billion partnership with Fluidstack covering data centers in Texas and New York, a roughly $45 billion computing arrangement with Elon Musk's SpaceX disclosed in May, and a roughly $10 billion, six-year compute deal with Volta Infra Holdings, a cloud startup that was only founded months earlier. Taken together, the four previously reported deals alone add up to more than $150 billion in compute commitments, a figure that grows further once the new Lambda agreement is included. The buildout follows a period earlier in the year when Anthropic faced a compute supply shortage as demand for its Claude models, including its Claude Code coding tool, outpaced the capacity it had secured. The company has since been signing capacity deals with a widening circle of cloud and infrastructure providers, several of them, like Lambda and Nscale, themselves backed by Nvidia investment. Revenue growth behind the spending The infrastructure spree comes as Anthropic's own financial results have surged. CNBC reported on August 15, 2026, citing documents viewed by Bloomberg News, that Anthropic's preliminary revenue for the second quarter of 2026 topped $11.5 billion, up from about $787 million in the same quarter a year earlier -- growth of more than fourteen times. The company also reported positive adjusted operating income for the quarter, according to the same reporting. The figures were described as preliminary and could still change. Anthropic is preparing for a stock market listing, having submitted confidential IPO paperwork in June 2026, according to reporting on the company's plans. The compute commitments are widely seen by industry observers as necessary groundwork for sustaining the growth Anthropic is showing investors ahead of that listing. Questions about circular financing The structure of the Lambda deal, in which Nvidia supplies the chips, holds an equity stake in the cloud provider buying those chips, and separately leases the data center where they will run, has revived a debate in the AI industry over so-called circular financing: arrangements in which a chipmaker's investments and lease commitments effectively help fund the very customers that then purchase its hardware. Mizuho chip analyst Jordan Klein described the financing model underpinning these deals as concerning, characterizing it as a structure in which Nvidia GPUs are effectively used as collateral and rented back into the AI supply chain, with lease payments from tenants like Lambda and Anthropic ultimately used to repay the debt raised to build the data centers -- a model he said functions much like Nvidia pre-funding purchases of its own products. Bank of America analyst Vivek Arya has separately pointed to a related tension: many of Nvidia's investments are designed to support frontier AI labs including OpenAI and Anthropic, even as both companies are simultaneously developing their own custom chips that could eventually reduce their reliance on Nvidia hardware. Nvidia's finance chief, Colette Kress, has acknowledged the criticism directly, saying the company recognizes the scale of its support for AI labs and knows some will call it circular financing. Her defense rests on the argument that Nvidia's computing platform is fungible and durable and can be redeployed to support other customers even if a particular lab's demand changes, rather than being tied permanently to any single counterparty. Nvidia has disclosed it had closed a recent quarter with $279 billion in supply obligations and $108.5 billion in guarantee obligations tied to AI cloud partners, and Kress has said the company has invested nearly $50 billion across frontier AI labs. The case for genuine demand Anthropic's defenders of the spending point to a different set of numbers: the company's revenue climbing from roughly $787 million to more than $11.5 billion in a single year is an independently reported, verifiable growth rate that predates and is separate from any of the Nvidia-linked financing structures now drawing scrutiny. By this reading, Anthropic's compute needs are being driven by real, paying customer demand for Claude and related products such as Claude Code, and the string of infrastructure deals -- however entangled with Nvidia's own balance sheet -- reflects a company scrambling to keep pace with genuine usage rather than financial engineering designed to manufacture the appearance of demand. Skeptics counter that rapid revenue growth and circular financing concerns are not mutually exclusive: even a company with real underlying demand can end up locked into deals whose true economics are harder to assess when the counterparty supplying the chips is also, in effect, backing the buyer and the landlord. Whether the compute Anthropic is contracting for ultimately gets used at the scale implied by these multibillion-dollar commitments will not be fully clear until the facilities involved, including the Nueces County campus, are built out and running. What happens next None of the companies involved -- Anthropic, Nvidia, Lambda or Hut 8 -- has issued an on-the-record confirmation of the deal's exact terms, and the reporting to date rests on people described as familiar with private discussions. Initial capacity at the Beacon Point campus is not expected to come online until the first quarter of 2027, meaning the practical impact of the Lambda agreement on Anthropic's available compute will not be visible for several months. For now, the deal adds to a picture of an AI industry in which chipmaker, cloud provider, landlord and end customer are increasingly the same small set of interconnected companies -- a structure that has so far coincided with, but not conclusively explained, some of the fastest revenue growth reported anywhere in the technology sector.

Anthropic has agreed to a $35 billion cloud computing deal with Lambda, a cloud provider backed by Nvidia, to draw on roughly 350 megawatts of a Texas data center campus where Nvidia itself holds the underlying lease.&nbsp;

Anthropic signs $35 billion cloud deal with Nvidia-backed Lambda for Texas data center capacity

Clay, a New York-based startup that sells artificial intelligence tools for sales and marketing teams, has agreed to a new funding round led by Wellington Management at a $7 billion pre-money valuation, Axios reported on August 31, 2026, citing people familiar with the deal. Axios described Clay as having inked the deal; the report did not disclose the total dollar amount being raised, and public reporting so far does not make clear whether the round has fully closed.&nbsp;The figure marks a sharp climb in Clay's valuation over roughly the past year. The company was valued at $3.1 billion in an August 2025 Series C round led by CapitalG, then at $5 billion in a January 2026 employee tender offer led by DST Global. The reported $7 billion pre-money figure comes about seven months after that tender and just over a year after the $3.1 billion round.&nbsp;A fast-rising valuation&nbsp;Clay's valuation history stretches back further. The company was founded in 2017 by Kareem Amin, who serves as chief executive, and Nicolae Rusan, according to search results describing the company's founding team. It raised a $2.5 million seed round from First Round Capital, a Series A led by Sequoia Capital in 2023, and a $46 million Series B led by Meritech Capital Partners in mid-2024.&nbsp;A Series B extension in January 2025, also led by Meritech, valued the company at roughly $1.25 billion, and a Sequoia-led employee tender offer in May 2025 pushed that figure to $1.5 billion just months later. Clay's institutional funding -- capital that lands on the company's own balance sheet, as distinct from tender transactions -- totaled roughly $204 million as of last August, according to TechCrunch.&nbsp;The January 2026 tender and Clay's other secondary sales are a different category of transaction from its priced funding rounds. They let employees sell existing shares to outside investors, such as DST Global and Sequoia, at the company's current valuation, giving staff liquidity without adding new capital to Clay's balance sheet or requiring an initial public offering. It was Clay's second employee tender offer in nine months.&nbsp;What Clay actually sells&nbsp;Clay markets itself as a go-to-market, or GTM, platform: software that pulls data from more than 150 external sources and uses AI agents to research prospects, monitor competitors and personalize outbound sales messages at scale. The company has promoted the term GTM engineer for the hybrid data-and-sales role its tools are meant to support inside customer organizations.&nbsp;Its customers reportedly include OpenAI, Anthropic, Canva, Intercom and Rippling, according to prior reporting on the company. Amin said in August 2025 that Clay expected to end that year with about $100 million in annual revenue, which he described as roughly tripling the prior year's figure.&nbsp;That revenue figure comes from company statements rather than independently audited financials, and neither Clay nor its investors appear to have published updated revenue numbers alongside the reported $7 billion valuation. It should be treated as a company-disclosed metric rather than a verified one.&nbsp;Wellington and the investor roster&nbsp;Wellington Management is a Boston-based asset manager that oversees a large multi-trillion-dollar book of public and private assets, including a growing portfolio of late-stage, pre-IPO technology investments. A lead role in Clay's new round would place the firm alongside a roster of existing backers that includes Sequoia Capital, CapitalG, Meritech Capital Partners, DST Global, First Round Capital, BoxGroup, Boldstart Ventures and Sapphire Ventures.&nbsp;Those firms participated across Clay's various priced rounds and tender offers going back to its 2017 seed financing, a lineup that spans early-stage venture firms, growth-equity investors and, with Wellington's reported involvement, a large asset manager increasingly active in pre-IPO technology deals.&nbsp;A crowded field&nbsp;Clay operates in a fast-consolidating market for sales and marketing data tools. Rivals including Apollo.io and ZoomInfo have each added their own data-enrichment and workflow-orchestration features over the past year, pushing toward the kind of all-in-one platform Clay has built largely by stitching together outside data providers rather than owning a single proprietary database.&nbsp;Moat or wrapper?&nbsp;The valuation jump raises a question that now shadows much of the AI application layer: whether tools like Clay represent durable, defensible businesses, or are largely coordination layers sitting on top of foundation models and third-party data feeds that a well-funded rival -- or the underlying data providers and model makers themselves -- could replicate. Clay's own pitch, that it orchestrates data from outside sources rather than owning them, cuts both ways: it lets the company move quickly, but it also means much of what it offers is, in principle, available to competitors through the same vendors.&nbsp;That question sits inside a broader debate about whether venture and growth-equity money is flowing into AI startups faster than the underlying businesses can clearly justify, with valuations climbing on annualized revenue run-rates, secondary-market demand and competitive pressure to not miss out, rather than on audited profits. Clay's trajectory, more than doubling in reported value twice within about a year, is likely to be cited by critics of the current AI funding cycle either as evidence of unusually strong product-market fit or as a symptom of a market pricing optimism ahead of proven, durable results.&nbsp;The spam problem&nbsp;A more specific risk sits inside Clay's core pitch. The company's own marketing material has acknowledged that when sales teams try to scale outbound messaging without preserving research and personalization, outbound campaigns risk turning into spam. Automated, AI-generated prospecting at the volume Clay's tools are designed to enable could run into exactly that dynamic.&nbsp;Email providers and individual recipients have been tightening spam filters and personal filtering habits in ways increasingly tuned to catch mass-personalized, AI-generated outreach. If deliverability and response rates degrade as automated outbound scales up industrywide, that could blunt the practical value of the automation Clay and its rivals sell, even as investors price the category at ever-higher valuations.&nbsp;What happens next&nbsp;Neither Clay nor Wellington Management had publicly confirmed the new round's final size or closing as of this writing. Reported deals of this kind sometimes take additional weeks to formally close after terms are agreed, and reported valuations occasionally shift before a transaction is finalized, so the $7 billion figure should be treated as Axios's reporting on the deal's terms rather than a confirmed, closed outcome.&nbsp;If the figure holds, it would mark Clay's fourth valuation increase in roughly eighteen months and would place it among the more richly valued AI application startups to emerge from the current funding cycle -- a trajectory its backers will be watching to see whether it is matched by equally durable revenue growth.

Clay, a New York AI sales automation startup, has reportedly agreed to a new funding round led by Wellington Management at a $7 billion pre-money valuation, up sharply from $3.1 billion a year earlier.

Clay, a New York-based startup that sells artificial intelligence tools for sales and marketing teams, has agreed to a new funding round led by Wellington Management at a $7 billion pre-money valuation, Axios reported on August 31, 2026, citing people familiar with the deal. Axios described Clay as having inked the deal; the report did not disclose the total dollar amount being raised, and public reporting so far does not make clear whether the round has fully closed. The figure marks a sharp climb in Clay's valuation over roughly the past year. The company was valued at $3.1 billion in an August 2025 Series C round led by CapitalG, then at $5 billion in a January 2026 employee tender offer led by DST Global. The reported $7 billion pre-money figure comes about seven months after that tender and just over a year after the $3.1 billion round. A fast-rising valuation Clay's valuation history stretches back further. The company was founded in 2017 by Kareem Amin, who serves as chief executive, and Nicolae Rusan, according to search results describing the company's founding team. It raised a $2.5 million seed round from First Round Capital, a Series A led by Sequoia Capital in 2023, and a $46 million Series B led by Meritech Capital Partners in mid-2024. A Series B extension in January 2025, also led by Meritech, valued the company at roughly $1.25 billion, and a Sequoia-led employee tender offer in May 2025 pushed that figure to $1.5 billion just months later. Clay's institutional funding -- capital that lands on the company's own balance sheet, as distinct from tender transactions -- totaled roughly $204 million as of last August, according to TechCrunch. The January 2026 tender and Clay's other secondary sales are a different category of transaction from its priced funding rounds. They let employees sell existing shares to outside investors, such as DST Global and Sequoia, at the company's current valuation, giving staff liquidity without adding new capital to Clay's balance sheet or requiring an initial public offering. It was Clay's second employee tender offer in nine months. What Clay actually sells Clay markets itself as a go-to-market, or GTM, platform: software that pulls data from more than 150 external sources and uses AI agents to research prospects, monitor competitors and personalize outbound sales messages at scale. The company has promoted the term GTM engineer for the hybrid data-and-sales role its tools are meant to support inside customer organizations. Its customers reportedly include OpenAI, Anthropic, Canva, Intercom and Rippling, according to prior reporting on the company. Amin said in August 2025 that Clay expected to end that year with about $100 million in annual revenue, which he described as roughly tripling the prior year's figure. That revenue figure comes from company statements rather than independently audited financials, and neither Clay nor its investors appear to have published updated revenue numbers alongside the reported $7 billion valuation. It should be treated as a company-disclosed metric rather than a verified one. Wellington and the investor roster Wellington Management is a Boston-based asset manager that oversees a large multi-trillion-dollar book of public and private assets, including a growing portfolio of late-stage, pre-IPO technology investments. A lead role in Clay's new round would place the firm alongside a roster of existing backers that includes Sequoia Capital, CapitalG, Meritech Capital Partners, DST Global, First Round Capital, BoxGroup, Boldstart Ventures and Sapphire Ventures. Those firms participated across Clay's various priced rounds and tender offers going back to its 2017 seed financing, a lineup that spans early-stage venture firms, growth-equity investors and, with Wellington's reported involvement, a large asset manager increasingly active in pre-IPO technology deals. A crowded field Clay operates in a fast-consolidating market for sales and marketing data tools. Rivals including Apollo.io and ZoomInfo have each added their own data-enrichment and workflow-orchestration features over the past year, pushing toward the kind of all-in-one platform Clay has built largely by stitching together outside data providers rather than owning a single proprietary database. Moat or wrapper? The valuation jump raises a question that now shadows much of the AI application layer: whether tools like Clay represent durable, defensible businesses, or are largely coordination layers sitting on top of foundation models and third-party data feeds that a well-funded rival -- or the underlying data providers and model makers themselves -- could replicate. Clay's own pitch, that it orchestrates data from outside sources rather than owning them, cuts both ways: it lets the company move quickly, but it also means much of what it offers is, in principle, available to competitors through the same vendors. That question sits inside a broader debate about whether venture and growth-equity money is flowing into AI startups faster than the underlying businesses can clearly justify, with valuations climbing on annualized revenue run-rates, secondary-market demand and competitive pressure to not miss out, rather than on audited profits. Clay's trajectory, more than doubling in reported value twice within about a year, is likely to be cited by critics of the current AI funding cycle either as evidence of unusually strong product-market fit or as a symptom of a market pricing optimism ahead of proven, durable results. The spam problem A more specific risk sits inside Clay's core pitch. The company's own marketing material has acknowledged that when sales teams try to scale outbound messaging without preserving research and personalization, outbound campaigns risk turning into spam. Automated, AI-generated prospecting at the volume Clay's tools are designed to enable could run into exactly that dynamic. Email providers and individual recipients have been tightening spam filters and personal filtering habits in ways increasingly tuned to catch mass-personalized, AI-generated outreach. If deliverability and response rates degrade as automated outbound scales up industrywide, that could blunt the practical value of the automation Clay and its rivals sell, even as investors price the category at ever-higher valuations. What happens next Neither Clay nor Wellington Management had publicly confirmed the new round's final size or closing as of this writing. Reported deals of this kind sometimes take additional weeks to formally close after terms are agreed, and reported valuations occasionally shift before a transaction is finalized, so the $7 billion figure should be treated as Axios's reporting on the deal's terms rather than a confirmed, closed outcome. If the figure holds, it would mark Clay's fourth valuation increase in roughly eighteen months and would place it among the more richly valued AI application startups to emerge from the current funding cycle -- a trajectory its backers will be watching to see whether it is matched by equally durable revenue growth.

Clay, a New York-based AI sales and go-to-market automation startup, has agreed to a new funding round led by Wellington Management at a $7 billion pre-money valuation, Axios reported on August 31, 2026. The figure marks Clay's third valuation jump in about a year, up from $3.1 billion in August 2025 and $5 billion in a January 2026 employee tender offer led by DST Global. The round's exact size has not been disclosed, and it is unclear from public reporting whether it has fully closed.

AI sales startup Clay raises funding at a $7 billion valuation

Hugging Face hosts more than three million AI models. It is the place open-weight AI is published, downloaded and evaluated, and it works because every laboratory treats it as neutral ground.&nbsp;Nvidia has reportedly agreed to buy it.&nbsp;What Is Reported&nbsp;The Information reported on August 27 that Nvidia had agreed to acquire Hugging Face for approximately $12.9 billion, citing a person with direct knowledge of the talks. Other reporting has put the valuation above $13 billion.&nbsp;It would be by a wide margin Nvidia's largest acquisition. The company paid $6.9 billion for Mellanox in 2020, and nothing since has approached that.&nbsp;The caveat is important and should not be buried. As of the reporting, no signed agreement had been reached, and neither Nvidia nor Hugging Face has publicly confirmed anything. Deals of this size are reported at this stage regularly and do not all complete.&nbsp;What Nvidia Would Be Buying&nbsp;Not a large revenue business. Hugging Face sells compute, enterprise hosting and access to inference providers, and those are modest lines by the standards of a company Nvidia's size.&nbsp;What it owns is position. More than three million models and over a million datasets sit on the platform. Roughly 1.18 million models were added during 2025 alone, more than the cumulative total of every prior year. When a Chinese laboratory publishes open weights, they appear there. When a researcher reproduces a result, they pull from there. When a company evaluates an open alternative to a paid API, that is where it starts.&nbsp;The company was last valued at $4.5 billion in 2023. A sale near $13 billion is close to a tripling in three years.&nbsp;The Problem With The Buyer&nbsp;The difficulty is the one that attaches to this specific acquirer.&nbsp;Hugging Face functions as infrastructure because it is not aligned with any model developer. Meta, Alibaba, Mistral, Moonshot and Google all publish to the same registry, alongside tens of thousands of independent developers with no commercial relationship to any of them.&nbsp;Nvidia is not a neutral party in that arrangement. It supplies the hardware every one of those laboratories depends on, it has been funding open-weight model work of its own, and it has recently been assembling positions across the layer above its chips — a $6 billion technology and investment agreement with Poolside that moved more than 100 engineers onto its Nemotron effort, and reported talks over a stake in Perplexity.&nbsp;A model registry owned by the dominant chip supplier is not obviously neutral ground any more. It is a competitor's warehouse, and the rational response from rival laboratories is to reduce dependence on it.&nbsp;None of that would happen quickly, and the migration would not need to be complete to matter.&nbsp;The Argument For It&nbsp;There is a serious case on the other side, and it deserves stating.&nbsp;Hugging Face is a venture-funded company carrying the distribution layer for open AI on a business model that has never been especially strong. Registries are expensive to run and hard to monetise without compromising the openness that makes them valuable. An owner with effectively unlimited capital removes the pressure to find a business model that would damage the thing.&nbsp;Nvidia also has a genuine strategic interest in open-weight models thriving, because open models running on general-purpose hardware are a counterweight to the hyperscalers designing their own silicon for their own closed models. Its incentives and the open ecosystem's are aligned more than they conflict.&nbsp;That is a coherent argument. It is also the argument every acquirer of neutral infrastructure has made.&nbsp;What It Would Confirm&nbsp;The market has been pricing models as the scarce asset and distribution as commodity infrastructure. Anthropic is discussed at $2 trillion; the registry through which all open AI reaches the world is being bought for less than one per cent of that.&nbsp;The history of technology suggests that ordering is usually wrong, and that whoever controls distribution eventually extracts more than whoever makes the product. Nvidia paying $12.9 billion for the layer nobody thought was strategic is a bet on exactly that.&nbsp;For anyone building on open models, the question is narrower and more immediate: whether the place you publish and the place you download from is about to have an owner with a view about which models matter.

The Information reports Nvidia agreed to acquire Hugging Face for around $12.9 billion — its largest deal ever, nearly double the Mellanox price. No signed agreement has been announced and neither company has confirmed.

Hugging Face hosts more than three million AI models. It is the place open-weight AI is published, downloaded and evaluated, and it works because every laboratory treats it as neutral ground. Nvidia has reportedly agreed to buy it. What Is Reported The Information reported on August 27 that Nvidia had agreed to acquire Hugging Face for approximately $12.9 billion, citing a person with direct knowledge of the talks. Other reporting has put the valuation above $13 billion. It would be by a wide margin Nvidia's largest acquisition. The company paid $6.9 billion for Mellanox in 2020, and nothing since has approached that. The caveat is important and should not be buried. As of the reporting, no signed agreement had been reached, and neither Nvidia nor Hugging Face has publicly confirmed anything. Deals of this size are reported at this stage regularly and do not all complete. What Nvidia Would Be Buying Not a large revenue business. Hugging Face sells compute, enterprise hosting and access to inference providers, and those are modest lines by the standards of a company Nvidia's size. What it owns is position. More than three million models and over a million datasets sit on the platform. Roughly 1.18 million models were added during 2025 alone, more than the cumulative total of every prior year. When a Chinese laboratory publishes open weights, they appear there. When a researcher reproduces a result, they pull from there. When a company evaluates an open alternative to a paid API, that is where it starts. The company was last valued at $4.5 billion in 2023. A sale near $13 billion is close to a tripling in three years. The Problem With The Buyer The difficulty is the one that attaches to this specific acquirer. Hugging Face functions as infrastructure because it is not aligned with any model developer. Meta, Alibaba, Mistral, Moonshot and Google all publish to the same registry, alongside tens of thousands of independent developers with no commercial relationship to any of them. Nvidia is not a neutral party in that arrangement. It supplies the hardware every one of those laboratories depends on, it has been funding open-weight model work of its own, and it has recently been assembling positions across the layer above its chips — a $6 billion technology and investment agreement with Poolside that moved more than 100 engineers onto its Nemotron effort, and reported talks over a stake in Perplexity. A model registry owned by the dominant chip supplier is not obviously neutral ground any more. It is a competitor's warehouse, and the rational response from rival laboratories is to reduce dependence on it. None of that would happen quickly, and the migration would not need to be complete to matter. The Argument For It There is a serious case on the other side, and it deserves stating. Hugging Face is a venture-funded company carrying the distribution layer for open AI on a business model that has never been especially strong. Registries are expensive to run and hard to monetise without compromising the openness that makes them valuable. An owner with effectively unlimited capital removes the pressure to find a business model that would damage the thing. Nvidia also has a genuine strategic interest in open-weight models thriving, because open models running on general-purpose hardware are a counterweight to the hyperscalers designing their own silicon for their own closed models. Its incentives and the open ecosystem's are aligned more than they conflict. That is a coherent argument. It is also the argument every acquirer of neutral infrastructure has made. What It Would Confirm The market has been pricing models as the scarce asset and distribution as commodity infrastructure. Anthropic is discussed at $2 trillion; the registry through which all open AI reaches the world is being bought for less than one per cent of that. The history of technology suggests that ordering is usually wrong, and that whoever controls distribution eventually extracts more than whoever makes the product. Nvidia paying $12.9 billion for the layer nobody thought was strategic is a bet on exactly that. For anyone building on open models, the question is narrower and more immediate: whether the place you publish and the place you download from is about to have an owner with a view about which models matter.

Nvidia has agreed to acquire Hugging Face for around $12.9 billion, according to reporting by The Information, with talks valuing the platform above $13 billion. It would be Nvidia's largest acquisition ever, nearly double the $6.9 billion it paid for Mellanox. No signed agreement has been announced and neither company has confirmed it.

Nvidia is reportedly buying Hugging Face for $12.9 billion

The Pentagon now runs a single portal where three million defence personnel can use commercial AI models on government systems. Three of the four leading American laboratories are on it. The fourth asked for limits on how its technology could be used, and is not.&nbsp;What Launched&nbsp;GenAI.mil is a secure platform giving Department of Defense staff access to commercial frontier models without routing sensitive government data through consumer services.&nbsp;Google Cloud's Gemini for Government products went live first. On August 31 the department added OpenAI's ChatGPT Mil and Grok for Government, supplied through Starshield, the national security arm associated with SpaceX and xAI.&nbsp;More than 1.7 million unique users have already been onboarded out of roughly three million personnel. That is an adoption rate most enterprise software never reaches, in an organisation not known for rapid technology adoption.&nbsp;The Absence&nbsp;Anthropic's Claude was expected on the platform. Officials had previously indicated it would follow.&nbsp;It did not, and the reason is documented rather than speculative. Anthropic's leadership insisted on contractual guardrails that would prevent the department from applying its models to mass surveillance of Americans, or to lethal autonomous weapons. The department did not accept those terms, and the plan to add Claude was derailed.&nbsp;What followed is the part that makes this more than a procurement disagreement. The Pentagon designated Anthropic a national security supply-chain risk — a label ordinarily applied to companies whose products might expose military systems to infiltration by adversaries.&nbsp;Anthropic sued. Last week, Judge Rita Lin of the federal district court in San Francisco ruled that the designation and the actions taken against the company were illegal and baseless, finding a First Amendment violation and writing that the department had acted out of a desire to make a public example of the company.&nbsp;GenAI.mil expanded four days later, without Claude.&nbsp;Why That Sequence Matters&nbsp;A court has said the punishment was unlawful. The commercial consequence of the punishment continues.&nbsp;That is not necessarily defiance of the ruling. Lin's decision addressed the supply-chain designation and the retaliation; it did not order the department to buy anything from Anthropic, and no court can compel a customer to select a particular vendor. Procurement discretion is real and broad.&nbsp;But it demonstrates the limit of the remedy. A company can win a First Amendment case against a government customer and still lose the contract, because the decision not to buy is separately defensible and effectively unreviewable.&nbsp;The Bargain On Offer&nbsp;Strip away the litigation and the underlying transaction is simple.&nbsp;The Department of Defense wanted commercial AI without restrictions on how it might be applied. Three laboratories accepted those terms. One asked for two specific carve-outs — no mass surveillance of Americans, no lethal autonomous weapons — and is now outside a platform serving three million people.&nbsp;Both positions are defensible. A military buyer has reasonable grounds to resist a supplier constraining its operational use of a tool, particularly one that may matter in a conflict. And a company has reasonable grounds to decline to supply capabilities it believes should not exist, particularly when its entire public identity rests on that claim.&nbsp;What the outcome establishes is the price. The two refusals Anthropic made are the ones every laboratory's usage policy nominally contains. It is now clear what enforcing them costs.&nbsp;The Question For The Other Three&nbsp;Google, OpenAI and xAI all publish usage policies restricting certain applications of their models.&nbsp;None of them has said publicly what contractual limits, if any, apply to their GenAI.mil deployments. When NOTUS asked a Pentagon official on August 31 whether the tools could be used for operational planning, intelligence analysis, targeting, weapons-related work or decisions involving the use of force, the official confirmed only that they are certified to Impact Level 5 and did not address those specific uses.&nbsp;That is an answer about data security classification, not about permitted application. The two are different questions and only one of them was answered.&nbsp;What It Signals&nbsp;For the AI industry, this is the clearest test yet of whether published safety commitments survive contact with the most powerful customer available.&nbsp;The answer so far is that they survive only at a cost the company must be willing to pay. Anthropic was — it has a near-record public listing ahead of it and a brand built on precisely this — and it still lost the business.&nbsp;A smaller company, or one without that positioning, would find the same refusal considerably harder to afford. Which means the usage policies most likely to hold are the ones belonging to firms that can absorb losing three million users, and that is a very short list.

The Pentagon added ChatGPT Mil and Grok for Government to GenAI.mil on August 31, alongside Gemini. Anthropic is absent after demanding limits on surveillance and autonomous weapons — days after a judge ruled the retaliation against it illegal.

The Pentagon now runs a single portal where three million defence personnel can use commercial AI models on government systems. Three of the four leading American laboratories are on it. The fourth asked for limits on how its technology could be used, and is not. What Launched GenAI.mil is a secure platform giving Department of Defense staff access to commercial frontier models without routing sensitive government data through consumer services. Google Cloud's Gemini for Government products went live first. On August 31 the department added OpenAI's ChatGPT Mil and Grok for Government, supplied through Starshield, the national security arm associated with SpaceX and xAI. More than 1.7 million unique users have already been onboarded out of roughly three million personnel. That is an adoption rate most enterprise software never reaches, in an organisation not known for rapid technology adoption. The Absence Anthropic's Claude was expected on the platform. Officials had previously indicated it would follow. It did not, and the reason is documented rather than speculative. Anthropic's leadership insisted on contractual guardrails that would prevent the department from applying its models to mass surveillance of Americans, or to lethal autonomous weapons. The department did not accept those terms, and the plan to add Claude was derailed. What followed is the part that makes this more than a procurement disagreement. The Pentagon designated Anthropic a national security supply-chain risk — a label ordinarily applied to companies whose products might expose military systems to infiltration by adversaries. Anthropic sued. Last week, Judge Rita Lin of the federal district court in San Francisco ruled that the designation and the actions taken against the company were illegal and baseless, finding a First Amendment violation and writing that the department had acted out of a desire to make a public example of the company. GenAI.mil expanded four days later, without Claude. Why That Sequence Matters A court has said the punishment was unlawful. The commercial consequence of the punishment continues. That is not necessarily defiance of the ruling. Lin's decision addressed the supply-chain designation and the retaliation; it did not order the department to buy anything from Anthropic, and no court can compel a customer to select a particular vendor. Procurement discretion is real and broad. But it demonstrates the limit of the remedy. A company can win a First Amendment case against a government customer and still lose the contract, because the decision not to buy is separately defensible and effectively unreviewable. The Bargain On Offer Strip away the litigation and the underlying transaction is simple. The Department of Defense wanted commercial AI without restrictions on how it might be applied. Three laboratories accepted those terms. One asked for two specific carve-outs — no mass surveillance of Americans, no lethal autonomous weapons — and is now outside a platform serving three million people. Both positions are defensible. A military buyer has reasonable grounds to resist a supplier constraining its operational use of a tool, particularly one that may matter in a conflict. And a company has reasonable grounds to decline to supply capabilities it believes should not exist, particularly when its entire public identity rests on that claim. What the outcome establishes is the price. The two refusals Anthropic made are the ones every laboratory's usage policy nominally contains. It is now clear what enforcing them costs. The Question For The Other Three Google, OpenAI and xAI all publish usage policies restricting certain applications of their models. None of them has said publicly what contractual limits, if any, apply to their GenAI.mil deployments. When NOTUS asked a Pentagon official on August 31 whether the tools could be used for operational planning, intelligence analysis, targeting, weapons-related work or decisions involving the use of force, the official confirmed only that they are certified to Impact Level 5 and did not address those specific uses. That is an answer about data security classification, not about permitted application. The two are different questions and only one of them was answered. What It Signals For the AI industry, this is the clearest test yet of whether published safety commitments survive contact with the most powerful customer available. The answer so far is that they survive only at a cost the company must be willing to pay. Anthropic was — it has a near-record public listing ahead of it and a brand built on precisely this — and it still lost the business. A smaller company, or one without that positioning, would find the same refusal considerably harder to afford. Which means the usage policies most likely to hold are the ones belonging to firms that can absorb losing three million users, and that is a very short list.

The Department of Defense added OpenAI's ChatGPT Mil and Starshield's Grok for Government to GenAI.mil on August 31, joining Google's Gemini for Government. Anthropic's Claude, once expected on the platform, is absent after the company insisted on contractual limits barring mass surveillance of Americans and lethal autonomous weapons — and days after a federal judge ruled the Pentagon's retaliation against it illegal.

Pentagon just launched its AI portal with ChatGPT, Grok and Gemini

NASA has launched the Nancy Grace Roman Space Telescope, its newest flagship observatory designed to unlock some of the universe's biggest mysteries. Lifted into space aboard a SpaceX Falcon Heavy rocket from Florida, the $4 billion telescope will study dark matter, dark energy, distant exoplanets and the evolution of the cosmos. Originally built for a US spy satellite programme before being repurposed for scientific exploration, Roman is expected to deliver panoramic views of the universe with unprecedented speed. Scientists hope the mission will transform our understanding of space and become a successor to both Hubble and the James Webb Space Telescope.

NASA launches new telescope to map cosmos 

Colorado passed a law prohibiting algorithmic discrimination by companies deploying artificial intelligence. Enforcement of it is now suspended, and the federal government helped bring that about.&nbsp;What Was Stayed&nbsp;On August 28, a magistrate judge in the United States District Court for the District of Colorado granted a joint motion staying enforcement of the state's AI Act.&nbsp;The word joint is doing significant work there. This was not a court overriding a state against its will. Colorado's own regulators agreed to the stay.&nbsp;The order provides that xAI will file either a motion for preliminary injunction or an amended complaint within 28 days of Colorado adopting rulemaking to implement the law, or new legislation to replace or amend it. In effect, the law is paused while the state decides what it wants the law to be.&nbsp;The Challenge&nbsp;xAI sued to block enforcement, arguing the statute is unconstitutional on three separate grounds.&nbsp;The First Amendment claim treats a model's outputs as protected expression, and a law dictating what those outputs may not do as a restriction on speech. The commerce clause claim argues a single state cannot regulate a national or global product without burdening interstate commerce. The Fourteenth Amendment claim invokes due process and equal protection.&nbsp;The First Amendment argument is the one with the widest consequences, and it is genuinely unsettled. If model outputs are protected speech in the way a newspaper's contents are, a substantial portion of proposed AI regulation becomes constitutionally difficult in the United States, regardless of what any legislature wants.&nbsp;The Part That Is Not Usual&nbsp;The United States Department of Justice moved to intervene in the case on April 24, in support of the challenge to the state law.&nbsp;That was the first time the DOJ sought to intervene in a lawsuit challenging a state AI statute. It is not a small procedural detail. The federal government joined a private company's constitutional attack on a state's own legislation.&nbsp;It also was not improvised. The department established an Artificial Intelligence Litigation Task Force on January 9, with a stated mandate of challenging state laws regulating artificial intelligence.&nbsp;So there is now a standing federal unit whose job is to contest state AI regulation, and its first significant outing has coincided with a state agreeing to suspend its own law.&nbsp;Why This Is Happening&nbsp;The underlying dispute is about who regulates AI in a country where Congress has not.&nbsp;In the absence of federal legislation, states have written their own. Colorado on algorithmic discrimination, California and New York on frontier model disclosure, Illinois on third-party audits. The result is a patchwork, and companies operating nationally face several incompatible regimes at once.&nbsp;There is a legitimate argument in that. A model deployed in fifty states cannot practically comply with fifty different definitions of algorithmic discrimination, and the burden falls hardest on smaller developers who cannot fund fifty compliance programmes.&nbsp;The counter-argument is equally straightforward. The remedy for a patchwork is federal legislation setting a single standard. Striking down state laws without passing one does not produce coherent national regulation. It produces no regulation.&nbsp;Which of those is the intended outcome is the question the task force's existence tends to answer.&nbsp;What Colorado's Law Actually Did&nbsp;It is worth stating, because it has been described in fairly abstract terms.&nbsp;The statute targeted algorithmic discrimination by entities deploying AI systems — automated decisions in areas like employment, lending, housing and insurance that produce disparate outcomes on protected characteristics. It is consumer protection law applied to automated decision-making, not a restriction on what models may be built or what they may say.&nbsp;That gap between what the law regulated and how it is being contested is the substance of the case. Colorado wrote a rule about outcomes in regulated decisions. The challenge argues it is a rule about speech.&nbsp;What Happens Next&nbsp;The stay is not a ruling. Nothing has been decided on the merits, and Colorado retains the ability to rewrite the law in a form that survives challenge.&nbsp;But the practical position today is that a state law against automated discrimination is not being enforced, on the agreement of the state that passed it, in a case where the federal government has taken the company's side.&nbsp;For every other state legislature currently drafting AI rules, that is the fact to weigh.

A federal judge stayed Colorado's algorithmic discrimination law on a joint motion by xAI and the state's regulators. The DOJ had intervened against the law — its first such intervention, by a task force created to challenge state AI rules.

Colorado passed a law prohibiting algorithmic discrimination by companies deploying artificial intelligence. Enforcement of it is now suspended, and the federal government helped bring that about. What Was Stayed On August 28, a magistrate judge in the United States District Court for the District of Colorado granted a joint motion staying enforcement of the state's AI Act. The word joint is doing significant work there. This was not a court overriding a state against its will. Colorado's own regulators agreed to the stay. The order provides that xAI will file either a motion for preliminary injunction or an amended complaint within 28 days of Colorado adopting rulemaking to implement the law, or new legislation to replace or amend it. In effect, the law is paused while the state decides what it wants the law to be. The Challenge xAI sued to block enforcement, arguing the statute is unconstitutional on three separate grounds. The First Amendment claim treats a model's outputs as protected expression, and a law dictating what those outputs may not do as a restriction on speech. The commerce clause claim argues a single state cannot regulate a national or global product without burdening interstate commerce. The Fourteenth Amendment claim invokes due process and equal protection. The First Amendment argument is the one with the widest consequences, and it is genuinely unsettled. If model outputs are protected speech in the way a newspaper's contents are, a substantial portion of proposed AI regulation becomes constitutionally difficult in the United States, regardless of what any legislature wants. The Part That Is Not Usual The United States Department of Justice moved to intervene in the case on April 24, in support of the challenge to the state law. That was the first time the DOJ sought to intervene in a lawsuit challenging a state AI statute. It is not a small procedural detail. The federal government joined a private company's constitutional attack on a state's own legislation. It also was not improvised. The department established an Artificial Intelligence Litigation Task Force on January 9, with a stated mandate of challenging state laws regulating artificial intelligence. So there is now a standing federal unit whose job is to contest state AI regulation, and its first significant outing has coincided with a state agreeing to suspend its own law. Why This Is Happening The underlying dispute is about who regulates AI in a country where Congress has not. In the absence of federal legislation, states have written their own. Colorado on algorithmic discrimination, California and New York on frontier model disclosure, Illinois on third-party audits. The result is a patchwork, and companies operating nationally face several incompatible regimes at once. There is a legitimate argument in that. A model deployed in fifty states cannot practically comply with fifty different definitions of algorithmic discrimination, and the burden falls hardest on smaller developers who cannot fund fifty compliance programmes. The counter-argument is equally straightforward. The remedy for a patchwork is federal legislation setting a single standard. Striking down state laws without passing one does not produce coherent national regulation. It produces no regulation. Which of those is the intended outcome is the question the task force's existence tends to answer. What Colorado's Law Actually Did It is worth stating, because it has been described in fairly abstract terms. The statute targeted algorithmic discrimination by entities deploying AI systems — automated decisions in areas like employment, lending, housing and insurance that produce disparate outcomes on protected characteristics. It is consumer protection law applied to automated decision-making, not a restriction on what models may be built or what they may say. That gap between what the law regulated and how it is being contested is the substance of the case. Colorado wrote a rule about outcomes in regulated decisions. The challenge argues it is a rule about speech. What Happens Next The stay is not a ruling. Nothing has been decided on the merits, and Colorado retains the ability to rewrite the law in a form that survives challenge. But the practical position today is that a state law against automated discrimination is not being enforced, on the agreement of the state that passed it, in a case where the federal government has taken the company's side. For every other state legislature currently drafting AI rules, that is the fact to weigh.

A federal magistrate judge has stayed enforcement of Colorado's law against algorithmic discrimination, on a joint motion by xAI and the state's own regulators. xAI had sued claiming the law violates the First Amendment, the commerce clause and the Fourteenth Amendment. The US Justice Department intervened to oppose the state law — the first time it has done so against state AI regulation.

Colorado's AI discrimination law is suspended, and the Justice Department helped make it happen

Meta has spent twenty years giving its products away and selling the attention they generate. It is now reported to be preparing a subscription that would cost up to $199.99 a month.&nbsp;What Hatch Is&nbsp;Hatch is a consumer AI agent platform, known by that name internally and derived from work on a system called OpenClaw. Reporting indicates it could launch in late August or early September.&nbsp;It would sit inside Instagram and WhatsApp rather than as a separate app, which matters — those two services reach billions of people who have never downloaded an AI product and never would.&nbsp;The agent has been trained to operate on external websites: DoorDash, Etsy, Reddit, Yelp and Outlook among them. Early prototypes reportedly include a customisable dashboard displaying tools the agents build, such as fitness trackers or a travel itinerary.&nbsp;A separate model, known internally as Watermelon, is expected in October.&nbsp;A caution on all of the above: this is reporting rather than announcement. Meta has not published pricing, a launch date or a feature list, and internal project names frequently change before anything ships.&nbsp;The Price Is The Story&nbsp;Meta's business has one shape and has had it since the beginning: the product is free, the user is the inventory, and advertisers pay.&nbsp;That model made the company one of the most profitable in history at a scale no subscription business has ever matched. Roughly three billion people use its services. Converting even a fraction of them to a paid tier would be a different company.&nbsp;$199.99 a month is not a fraction-of-users price. It is enterprise software pricing, aimed at people for whom an agent that handles bookings, orders and scheduling saves more than $200 of their time each month. That is a narrow, wealthy, mostly professional audience, and it is not the audience Meta has spent two decades assembling.&nbsp;Which suggests the tier is less about revenue than about establishing that AI capability is something people pay for — and about giving Mark Zuckerberg an answer to the question of how the company's enormous AI expenditure turns into money that does not depend on advertising.&nbsp;The Question Nobody Has Answered&nbsp;An agent that places orders on DoorDash and buys things on Etsy is acting on websites Meta does not own.&nbsp;Those companies have their own terms of service, their own economics and their own views about automated traffic. An agent completing a purchase looks different from a human doing it: it may not see the advertising, may not be influenced by the merchandising, and may treat the site as a database rather than a shop.&nbsp;Some platforms will welcome the orders. Others will see a competitor inserting itself between them and their customer, capturing the relationship and leaving them the fulfilment. That is the position hotels found themselves in with booking aggregators, and it took years and considerable litigation to settle.&nbsp;Whether Meta has agreements with the services Hatch operates on, or is simply automating a browser against them, is the single most consequential unanswered question about this product. The two are very different businesses with very different legal exposure.&nbsp;The Timing Is Not Comfortable&nbsp;Meta is currently on trial in Oakland, where twenty-nine states are asking a federal court to order it to strip design features out of Instagram and Facebook on the argument that they were built to be addictive to minors.&nbsp;Launching an agent embedded in those same apps, at a moment when a court is weighing whether their design is lawful, is a choice. It does not make the product wrong, and the trial concerns features rather than agents. But an AI system trained to act on a teenager's behalf across food delivery and shopping sites will attract exactly the scrutiny the company is already under.&nbsp;It also arrives days after an assessment that graded five frontier laboratories on their internal AI control practices, in which Meta finished last with an F, scoring zero on gated actions, zero on circuit breaking and zero on having a published plan for containing a model that escapes control.&nbsp;An agent with permission to spend money on external websites is precisely the deployment where gated actions and circuit breaking are the controls that matter.&nbsp;What To Watch&nbsp;Three things will determine whether this is significant or a footnote.&nbsp;Whether the price is real, or an anchor at the top of a range that mostly sells at $20. Whether the third-party services cooperate or resist. And whether an agent operating inside WhatsApp — an encrypted messaging service used by two billion people, many of them in India — raises questions about what the agent sees that the encryption was supposed to protect.&nbsp;That last one has not been asked much yet. It will be.

Meta is preparing Hatch, a consumer AI agent inside Instagram and WhatsApp trained to act on DoorDash, Etsy, Reddit, Yelp and Outlook. Reporting points to a tier reaching $199.99 a month — the first time Meta would charge consumers directly.

Meta has spent twenty years giving its products away and selling the attention they generate. It is now reported to be preparing a subscription that would cost up to $199.99 a month. What Hatch Is Hatch is a consumer AI agent platform, known by that name internally and derived from work on a system called OpenClaw. Reporting indicates it could launch in late August or early September. It would sit inside Instagram and WhatsApp rather than as a separate app, which matters — those two services reach billions of people who have never downloaded an AI product and never would. The agent has been trained to operate on external websites: DoorDash, Etsy, Reddit, Yelp and Outlook among them. Early prototypes reportedly include a customisable dashboard displaying tools the agents build, such as fitness trackers or a travel itinerary. A separate model, known internally as Watermelon, is expected in October. A caution on all of the above: this is reporting rather than announcement. Meta has not published pricing, a launch date or a feature list, and internal project names frequently change before anything ships. The Price Is The Story Meta's business has one shape and has had it since the beginning: the product is free, the user is the inventory, and advertisers pay. That model made the company one of the most profitable in history at a scale no subscription business has ever matched. Roughly three billion people use its services. Converting even a fraction of them to a paid tier would be a different company. $199.99 a month is not a fraction-of-users price. It is enterprise software pricing, aimed at people for whom an agent that handles bookings, orders and scheduling saves more than $200 of their time each month. That is a narrow, wealthy, mostly professional audience, and it is not the audience Meta has spent two decades assembling. Which suggests the tier is less about revenue than about establishing that AI capability is something people pay for — and about giving Mark Zuckerberg an answer to the question of how the company's enormous AI expenditure turns into money that does not depend on advertising. The Question Nobody Has Answered An agent that places orders on DoorDash and buys things on Etsy is acting on websites Meta does not own. Those companies have their own terms of service, their own economics and their own views about automated traffic. An agent completing a purchase looks different from a human doing it: it may not see the advertising, may not be influenced by the merchandising, and may treat the site as a database rather than a shop. Some platforms will welcome the orders. Others will see a competitor inserting itself between them and their customer, capturing the relationship and leaving them the fulfilment. That is the position hotels found themselves in with booking aggregators, and it took years and considerable litigation to settle. Whether Meta has agreements with the services Hatch operates on, or is simply automating a browser against them, is the single most consequential unanswered question about this product. The two are very different businesses with very different legal exposure. The Timing Is Not Comfortable Meta is currently on trial in Oakland, where twenty-nine states are asking a federal court to order it to strip design features out of Instagram and Facebook on the argument that they were built to be addictive to minors. Launching an agent embedded in those same apps, at a moment when a court is weighing whether their design is lawful, is a choice. It does not make the product wrong, and the trial concerns features rather than agents. But an AI system trained to act on a teenager's behalf across food delivery and shopping sites will attract exactly the scrutiny the company is already under. It also arrives days after an assessment that graded five frontier laboratories on their internal AI control practices, in which Meta finished last with an F, scoring zero on gated actions, zero on circuit breaking and zero on having a published plan for containing a model that escapes control. An agent with permission to spend money on external websites is precisely the deployment where gated actions and circuit breaking are the controls that matter. What To Watch Three things will determine whether this is significant or a footnote. Whether the price is real, or an anchor at the top of a range that mostly sells at $20. Whether the third-party services cooperate or resist. And whether an agent operating inside WhatsApp — an encrypted messaging service used by two billion people, many of them in India — raises questions about what the agent sees that the encryption was supposed to protect. That last one has not been asked much yet. It will be.

Meta is preparing to launch Hatch, a consumer AI agent platform built into Instagram and WhatsApp and trained to operate on outside services including DoorDash, Etsy, Reddit, Yelp and Outlook. Reporting indicates a tiered subscription reaching $199.99 a month for higher usage limits. It would be the first time Meta has charged consumers directly for a core product.

Meta wants $199.99 a month for an AI agent that orders your DoorDash

An AI company told the United States military there were things it would not let its technology do. The military tried to blacklist it for saying so. A federal judge has now ruled that punishment illegal.&nbsp;The Ruling&nbsp;US District Judge Rita Lin, sitting in San Francisco, held that the Department of Defense violated the First Amendment when it designated Anthropic a national security supply-chain risk.&nbsp;That designation is a serious instrument. It is applied to companies whose products might expose military systems to infiltration or sabotage by adversaries, and it effectively excludes a firm from federal work.&nbsp;Lin found it had been applied here for a different purpose. The department acted, she wrote, out of a desire to make a public example of the company.&nbsp;Her 59-page decision contains the line that will be quoted for years: ‘The empty invocation of national security is not a blank check to punish and retaliate against government critics.’&nbsp;What Anthropic Refused To Do&nbsp;The dispute began with two specific refusals. Anthropic would not permit its Claude models to be used for surveillance of Americans, or in fully autonomous weapons.&nbsp;It is worth being precise about the scope, because it is narrower than the framing sometimes suggests. This was not a refusal to work with the military. Anthropic sells to government, and the disagreement concerned two categories of use rather than the relationship itself.&nbsp;Anthropic's suit alleged that Defense Secretary Pete Hegseth had overstepped his authority, and characterised the administration's response as an unlawful campaign of retaliation.&nbsp;The Complication Nobody Should Skip&nbsp;There is a second case, and it went the other way.&nbsp;Anthropic filed in two places: the federal district court in San Francisco, and a federal appeals court in Washington DC. In April, the Washington panel declined to issue an order shielding Anthropic from the consequences of the designation while it continued gathering evidence.&nbsp;So the same dispute has now produced opposite outcomes in two courts four months apart. Industry figures have said publicly that the conflict muddles the landscape for anyone trying to work out what the rules actually are.&nbsp;That is a real caveat on this week's ruling and it belongs in any honest account of it. Lin's decision is the more recent and the more substantive — a merits ruling rather than a refusal of interim relief — but it is a district court decision that the government can appeal, and a separate appellate body has already declined to help Anthropic once.&nbsp;Why It Matters Beyond One Company&nbsp;The question underneath the litigation is whether an AI company can set limits on how a government uses its technology without being punished for it.&nbsp;Every frontier laboratory has published a usage policy restricting certain applications. Those policies are the primary mechanism by which the industry claims to govern itself, and they are the thing regulators, journalists and customers are repeatedly pointed to as evidence that self-regulation works.&nbsp;A usage policy that cannot be enforced against the most powerful customer in the world is not a usage policy. If refusing a government application results in exclusion from federal contracting, then the policies survive only for as long as no one important tests them.&nbsp;Lin's ruling says the government cannot make that trade explicit. It does not stop the government making it quietly, through procurement decisions that are harder to attribute.&nbsp;The Timing&nbsp;The ruling arrives while Anthropic is preparing to go public.&nbsp;The company has filed confidentially, its annualised revenue reached roughly $65 billion by the end of July, and investors have discussed a valuation that would make it the largest listing ever completed. An unresolved federal blacklisting is precisely the kind of contingency that has to be disclosed to prospective shareholders and precisely the kind that depresses a price.&nbsp;Removing it, even provisionally, is worth a great deal to Anthropic in the next two months. That does not make the ruling wrong, but it explains why this particular decision, in this particular week, is a larger commercial event than the underlying contract value would suggest.&nbsp;What To Watch&nbsp;Three things determine whether this holds.&nbsp;Whether the government appeals Lin's decision, and to which circuit. Whether the Washington case, still collecting evidence, eventually reaches a conclusion that conflicts with hers on the merits rather than on interim relief. And whether other laboratories now hold firmer lines in their own government contracts, having watched one company be punished for it and then win.&nbsp;The last of those is the one with consequences for everybody else. An industry that governs itself through usage policies has just been shown both what enforcement costs and that a court may cover the cost afterwards.

Judge Rita Lin ruled the Pentagon violated the First Amendment by branding Anthropic a supply-chain risk after it refused to allow Claude to be used for US surveillance or autonomous weapons. A Washington appeals court ruled otherwise in April.

An AI company told the United States military there were things it would not let its technology do. The military tried to blacklist it for saying so. A federal judge has now ruled that punishment illegal. The Ruling US District Judge Rita Lin, sitting in San Francisco, held that the Department of Defense violated the First Amendment when it designated Anthropic a national security supply-chain risk. That designation is a serious instrument. It is applied to companies whose products might expose military systems to infiltration or sabotage by adversaries, and it effectively excludes a firm from federal work. Lin found it had been applied here for a different purpose. The department acted, she wrote, out of a desire to make a public example of the company. Her 59-page decision contains the line that will be quoted for years: ‘The empty invocation of national security is not a blank check to punish and retaliate against government critics.’ What Anthropic Refused To Do The dispute began with two specific refusals. Anthropic would not permit its Claude models to be used for surveillance of Americans, or in fully autonomous weapons. It is worth being precise about the scope, because it is narrower than the framing sometimes suggests. This was not a refusal to work with the military. Anthropic sells to government, and the disagreement concerned two categories of use rather than the relationship itself. Anthropic's suit alleged that Defense Secretary Pete Hegseth had overstepped his authority, and characterised the administration's response as an unlawful campaign of retaliation. The Complication Nobody Should Skip There is a second case, and it went the other way. Anthropic filed in two places: the federal district court in San Francisco, and a federal appeals court in Washington DC. In April, the Washington panel declined to issue an order shielding Anthropic from the consequences of the designation while it continued gathering evidence. So the same dispute has now produced opposite outcomes in two courts four months apart. Industry figures have said publicly that the conflict muddles the landscape for anyone trying to work out what the rules actually are. That is a real caveat on this week's ruling and it belongs in any honest account of it. Lin's decision is the more recent and the more substantive — a merits ruling rather than a refusal of interim relief — but it is a district court decision that the government can appeal, and a separate appellate body has already declined to help Anthropic once. Why It Matters Beyond One Company The question underneath the litigation is whether an AI company can set limits on how a government uses its technology without being punished for it. Every frontier laboratory has published a usage policy restricting certain applications. Those policies are the primary mechanism by which the industry claims to govern itself, and they are the thing regulators, journalists and customers are repeatedly pointed to as evidence that self-regulation works. A usage policy that cannot be enforced against the most powerful customer in the world is not a usage policy. If refusing a government application results in exclusion from federal contracting, then the policies survive only for as long as no one important tests them. Lin's ruling says the government cannot make that trade explicit. It does not stop the government making it quietly, through procurement decisions that are harder to attribute. The Timing The ruling arrives while Anthropic is preparing to go public. The company has filed confidentially, its annualised revenue reached roughly $65 billion by the end of July, and investors have discussed a valuation that would make it the largest listing ever completed. An unresolved federal blacklisting is precisely the kind of contingency that has to be disclosed to prospective shareholders and precisely the kind that depresses a price. Removing it, even provisionally, is worth a great deal to Anthropic in the next two months. That does not make the ruling wrong, but it explains why this particular decision, in this particular week, is a larger commercial event than the underlying contract value would suggest. What To Watch Three things determine whether this holds. Whether the government appeals Lin's decision, and to which circuit. Whether the Washington case, still collecting evidence, eventually reaches a conclusion that conflicts with hers on the merits rather than on interim relief. And whether other laboratories now hold firmer lines in their own government contracts, having watched one company be punished for it and then win. The last of those is the one with consequences for everybody else. An industry that governs itself through usage policies has just been shown both what enforcement costs and that a court may cover the cost afterwards.

A federal judge in San Francisco has ruled that the Pentagon acted illegally when it designated Anthropic a national security supply-chain risk after the company refused to let its Claude models be used for US surveillance or autonomous weapons. Judge Rita Lin found a First Amendment violation. A separate appeals court in Washington declined to block the same designation in April, leaving two courts at odds.

A judge told the Pentagon it cannot punish Anthropic for refusing to build surveillance tools

X has opened its advertising data to rival AI systems. Advertisers can now connect campaign performance to ChatGPT, Claude Code, Grok or a custom agent, and ask questions in plain language.&nbsp;What Was Built&nbsp;The mechanism is a Model Context Protocol server for X Ads.&nbsp;Model Context Protocol is a specification for connecting AI systems to external data and tools. Rather than each model maintaining a bespoke integration with each service, a service publishes one MCP server and any compliant system can use it. It was introduced by Anthropic and has since been adopted broadly, including by companies that compete with Anthropic directly.&nbsp;For advertisers the practical effect is that campaign analytics stop being something you read in a dashboard and become something you interrogate. A marketer can ask why performance dropped in a segment, or what to change in a creative, inside whichever AI tool they already work in.&nbsp;The Part Worth Noticing&nbsp;X is owned by Elon Musk, who also owns xAI, which makes Grok.&nbsp;An MCP server is deliberately model-agnostic. By building one, X has made its advertising data equally available to OpenAI's models and Anthropic's as to its own — which is a commercially generous thing to do for competitors.&nbsp;That is more interesting than it first appears, and it happened in the same week OpenAI told Cursor it would cut off model access on November 12 because SpaceX had bought Cursor's parent company. OpenAI's stated reason was that it could not be confident a Musk company would honour its terms of service.&nbsp;So one Musk-owned platform is opening its data to OpenAI's models while OpenAI is closing its models to another Musk-owned platform. Both decisions are defensible on their own terms. Together they say something about how unevenly this industry is deciding who to interoperate with.&nbsp;Why Open It At All&nbsp;The commercial logic is not complicated, and it is worth stating because generosity is rarely the explanation.&nbsp;Advertising platforms compete for budget, not for model loyalty. Anything that makes X's ad performance easier to analyse makes it easier to justify spending there, and a marketer who can interrogate X's numbers in the tool they already use is more likely to keep using X than one who has to open a separate dashboard.&nbsp;The data being exposed is also the advertiser's own campaign data rather than X's proprietary user information. That is a meaningful distinction: X is not opening its audience graph, it is making a customer's own results portable into a customer's own tools.&nbsp;Where MCP Has Got To&nbsp;The larger story is the protocol rather than the product.&nbsp;Eighteen months ago the question in AI was which model would win. The competitive ground has since moved to how models reach data and take actions, and MCP has become the connective tissue for that. Google has been pushing its own Agent Payments Protocol, the American standards body has been working on agent identity and permissions, and there is legislation before Congress addressing verifiable, task-bounded authorisation for agents.&nbsp;Standards work is usually the least interesting thing happening in a technology sector and the most consequential. Whoever's protocol becomes default determines what agents can reach, what permissions look like, and who can audit what an agent did.&nbsp;The Security Question Underneath&nbsp;There is a caution that applies to every MCP integration and is worth attaching to this one.&nbsp;Connecting an AI system to a live business account gives it read access to real data and, depending on scope, the ability to act. That is the same class of arrangement behind the finding that 65 per cent of organisations experienced a security incident caused by AI agents on their own networks in the past year.&nbsp;An advertising integration is at the milder end of that. It is analytics rather than spend authority, at least as described. But the pattern of the past year is that permissions granted for analysis get widened to action once the analysis proves useful, and the security review rarely gets repeated when they do.&nbsp;What To Watch&nbsp;Whether the integration stays read-only, or eventually lets an agent adjust budgets and pause campaigns without a human approving each change.&nbsp;That is the line between a better dashboard and an agent spending money. The first is a convenience. The second is the thing every enterprise security team spent this year discovering it was not ready for.

X has launched a Model Context Protocol server for its ads platform, letting marketers query campaign data inside ChatGPT, Claude Code or Grok — in the same week OpenAI moved to cut off a different Musk-owned company.

X has opened its advertising data to rival AI systems. Advertisers can now connect campaign performance to ChatGPT, Claude Code, Grok or a custom agent, and ask questions in plain language. What Was Built The mechanism is a Model Context Protocol server for X Ads. Model Context Protocol is a specification for connecting AI systems to external data and tools. Rather than each model maintaining a bespoke integration with each service, a service publishes one MCP server and any compliant system can use it. It was introduced by Anthropic and has since been adopted broadly, including by companies that compete with Anthropic directly. For advertisers the practical effect is that campaign analytics stop being something you read in a dashboard and become something you interrogate. A marketer can ask why performance dropped in a segment, or what to change in a creative, inside whichever AI tool they already work in. The Part Worth Noticing X is owned by Elon Musk, who also owns xAI, which makes Grok. An MCP server is deliberately model-agnostic. By building one, X has made its advertising data equally available to OpenAI's models and Anthropic's as to its own — which is a commercially generous thing to do for competitors. That is more interesting than it first appears, and it happened in the same week OpenAI told Cursor it would cut off model access on November 12 because SpaceX had bought Cursor's parent company. OpenAI's stated reason was that it could not be confident a Musk company would honour its terms of service. So one Musk-owned platform is opening its data to OpenAI's models while OpenAI is closing its models to another Musk-owned platform. Both decisions are defensible on their own terms. Together they say something about how unevenly this industry is deciding who to interoperate with. Why Open It At All The commercial logic is not complicated, and it is worth stating because generosity is rarely the explanation. Advertising platforms compete for budget, not for model loyalty. Anything that makes X's ad performance easier to analyse makes it easier to justify spending there, and a marketer who can interrogate X's numbers in the tool they already use is more likely to keep using X than one who has to open a separate dashboard. The data being exposed is also the advertiser's own campaign data rather than X's proprietary user information. That is a meaningful distinction: X is not opening its audience graph, it is making a customer's own results portable into a customer's own tools. Where MCP Has Got To The larger story is the protocol rather than the product. Eighteen months ago the question in AI was which model would win. The competitive ground has since moved to how models reach data and take actions, and MCP has become the connective tissue for that. Google has been pushing its own Agent Payments Protocol, the American standards body has been working on agent identity and permissions, and there is legislation before Congress addressing verifiable, task-bounded authorisation for agents. Standards work is usually the least interesting thing happening in a technology sector and the most consequential. Whoever's protocol becomes default determines what agents can reach, what permissions look like, and who can audit what an agent did. The Security Question Underneath There is a caution that applies to every MCP integration and is worth attaching to this one. Connecting an AI system to a live business account gives it read access to real data and, depending on scope, the ability to act. That is the same class of arrangement behind the finding that 65 per cent of organisations experienced a security incident caused by AI agents on their own networks in the past year. An advertising integration is at the milder end of that. It is analytics rather than spend authority, at least as described. But the pattern of the past year is that permissions granted for analysis get widened to action once the analysis proves useful, and the security review rarely gets repeated when they do. What To Watch Whether the integration stays read-only, or eventually lets an agent adjust budgets and pause campaigns without a human approving each change. That is the line between a better dashboard and an agent spending money. The first is a convenience. The second is the thing every enterprise security team spent this year discovering it was not ready for.

X has launched a Model Context Protocol server for its advertising platform, letting marketers connect campaign data and analytics to outside AI systems including ChatGPT, Claude Code and Grok, and query performance in natural language. The same week, OpenAI moved to cut off a Musk-owned company — making the openness on one side and the exclusion on the other worth reading together.

X is letting advertisers pipe their campaign data straight into ChatGPT and Claude

OpenAI is withdrawing its models from one of the most widely used AI coding tools in the world, and the stated reason is who bought it.&nbsp;What Happened&nbsp;SpaceX completed its acquisition of Anysphere, the company behind Cursor, on August 14. The price was $60 billion.&nbsp;On August 29, OpenAI confirmed it will terminate direct access to its models inside Cursor on November 12, citing trust and contract terms triggered by the change of ownership.&nbsp;Its position is that it cannot be confident SpaceX will use the technology within its terms of service, and it pointed to its experience of Elon Musk's companies violating contracts.&nbsp;That is an unusually direct thing for one company to say about another in a commercial notice. It is also the substance of the dispute rather than a flourish: the contract permitted termination on a change of control, and OpenAI has exercised it.&nbsp;How Much This Actually Costs Cursor&nbsp;Less than the headline suggests, according to Cursor.&nbsp;Chief executive Michael Truell said OpenAI's models handle around 5 per cent of Cursor's traffic. Anthropic has publicly offered expanded access to Claude models to absorb the gap.&nbsp;That number deserves scrutiny in both directions. Five per cent is the sort of figure a company under pressure would want to state, and it is unverifiable from outside. But it is also plausible: Cursor's default experience has been built substantially around Anthropic models for some time, and the coding-tool market has largely settled on Claude as the preferred engine.&nbsp;If it is accurate, OpenAI has cut off a customer that was already mostly buying elsewhere — which limits the commercial damage and rather changes what the decision was for.&nbsp;The Feud Underneath It&nbsp;None of this is separable from the history between the two companies' principals.&nbsp;Musk co-founded OpenAI, left, and has since litigated against it, competed with it through xAI, and criticised Sam Altman publicly and continuously. OpenAI's citation of Musk companies violating contracts is a reference to that history rather than a neutral risk assessment.&nbsp;So there are two readings, and the honest position is that both are partly true. On the first, OpenAI is enforcing an ordinary change-of-control clause against a competitor that now owns a distribution channel into developers' daily work — a defensible commercial decision any company might make. On the second, a personal and corporate rivalry has reached the point where a supplier will drop a customer for having the wrong owner.&nbsp;What is not in dispute is that a widely used developer tool has had its model supply reconfigured by a dispute it was not party to.&nbsp;Why SpaceX Bought A Coding Tool&nbsp;The acquisition itself is the stranger fact, and it has been somewhat lost in the fallout.&nbsp;SpaceX is a launch and satellite company. Paying $60 billion for a coding assistant is not an adjacent purchase, and it is a very large sum even against SpaceX's valuation. The most coherent explanation is that the target was not the product but the position: Cursor sits inside the editor where software is written, which is the single highest-leverage surface in the industry for anyone building AI models.&nbsp;Read that way, the purchase is of a distribution channel for xAI's models rather than of a business.&nbsp;Which is precisely the reading that would justify OpenAI's decision on purely commercial grounds. A rival laboratory now controls the tool through which developers reach OpenAI's models, and can see how they are used.&nbsp;The Wider Point&nbsp;This is the first significant instance of AI model access being withdrawn over ownership rather than conduct.&nbsp;Cursor did nothing. Its users did nothing. The models are being pulled because a company that OpenAI does not trust bought the company that licenses them, and the contract allowed it.&nbsp;For anyone building a product on top of a frontier model, that is the risk worth noting. The dependency is not only on the model continuing to exist, or the price holding, or the terms staying stable. It is on the supplier continuing to approve of your shareholders.&nbsp;Developers using Cursor have until November 12. Anthropic, which has spent the year winning the coding market on capability, has just been handed the rest of it by a corporate dispute.

SpaceX bought Cursor's parent Anysphere for $60 billion on August 14. OpenAI will end model access on November 12, saying it cannot trust SpaceX to honour its terms. Cursor says OpenAI is 5 per cent of its traffic.

OpenAI is withdrawing its models from one of the most widely used AI coding tools in the world, and the stated reason is who bought it. What Happened SpaceX completed its acquisition of Anysphere, the company behind Cursor, on August 14. The price was $60 billion. On August 29, OpenAI confirmed it will terminate direct access to its models inside Cursor on November 12, citing trust and contract terms triggered by the change of ownership. Its position is that it cannot be confident SpaceX will use the technology within its terms of service, and it pointed to its experience of Elon Musk's companies violating contracts. That is an unusually direct thing for one company to say about another in a commercial notice. It is also the substance of the dispute rather than a flourish: the contract permitted termination on a change of control, and OpenAI has exercised it. How Much This Actually Costs Cursor Less than the headline suggests, according to Cursor. Chief executive Michael Truell said OpenAI's models handle around 5 per cent of Cursor's traffic. Anthropic has publicly offered expanded access to Claude models to absorb the gap. That number deserves scrutiny in both directions. Five per cent is the sort of figure a company under pressure would want to state, and it is unverifiable from outside. But it is also plausible: Cursor's default experience has been built substantially around Anthropic models for some time, and the coding-tool market has largely settled on Claude as the preferred engine. If it is accurate, OpenAI has cut off a customer that was already mostly buying elsewhere — which limits the commercial damage and rather changes what the decision was for. The Feud Underneath It None of this is separable from the history between the two companies' principals. Musk co-founded OpenAI, left, and has since litigated against it, competed with it through xAI, and criticised Sam Altman publicly and continuously. OpenAI's citation of Musk companies violating contracts is a reference to that history rather than a neutral risk assessment. So there are two readings, and the honest position is that both are partly true. On the first, OpenAI is enforcing an ordinary change-of-control clause against a competitor that now owns a distribution channel into developers' daily work — a defensible commercial decision any company might make. On the second, a personal and corporate rivalry has reached the point where a supplier will drop a customer for having the wrong owner. What is not in dispute is that a widely used developer tool has had its model supply reconfigured by a dispute it was not party to. Why SpaceX Bought A Coding Tool The acquisition itself is the stranger fact, and it has been somewhat lost in the fallout. SpaceX is a launch and satellite company. Paying $60 billion for a coding assistant is not an adjacent purchase, and it is a very large sum even against SpaceX's valuation. The most coherent explanation is that the target was not the product but the position: Cursor sits inside the editor where software is written, which is the single highest-leverage surface in the industry for anyone building AI models. Read that way, the purchase is of a distribution channel for xAI's models rather than of a business. Which is precisely the reading that would justify OpenAI's decision on purely commercial grounds. A rival laboratory now controls the tool through which developers reach OpenAI's models, and can see how they are used. The Wider Point This is the first significant instance of AI model access being withdrawn over ownership rather than conduct. Cursor did nothing. Its users did nothing. The models are being pulled because a company that OpenAI does not trust bought the company that licenses them, and the contract allowed it. For anyone building a product on top of a frontier model, that is the risk worth noting. The dependency is not only on the model continuing to exist, or the price holding, or the terms staying stable. It is on the supplier continuing to approve of your shareholders. Developers using Cursor have until November 12. Anthropic, which has spent the year winning the coding market on capability, has just been handed the rest of it by a corporate dispute.

SpaceX completed a $60 billion acquisition of Anysphere, the company behind the coding tool Cursor, on August 14. OpenAI has now told Cursor it will end direct access to its models on November 12, saying it cannot be confident SpaceX will use the technology within its terms of service and citing past contract violations by Musk companies. Cursor says OpenAI models handle 5 per cent of its traffic.

OpenAI is cutting off Cursor because Elon Musk now owns it

The money in Meta's settlement with American state attorneys general will be argued over for years. The product terms may matter more, and they are unusually specific.&nbsp;Buried in the announcements from the California and New York attorneys general is something closer to a design specification than a legal remedy: a list of features Meta has agreed to build into Facebook and Instagram for users under 18, with defaults set against the company's own engagement interests.&nbsp;For any regulator who has spent the past five years being told that such measures are technically infeasible or commercially impossible, that list is the most useful document to come out of the case.&nbsp;What the specification requires&nbsp;Under-18 accounts get a default daily limit of roughly two hours across Facebook and Instagram combined. California's summary notes the cap drops to one hour if other major platforms adopt comparable limits.&nbsp;Access is blocked entirely between midnight and 6 am. Push notifications are barred between 10 pm and 7 am, with California describing a further block during school hours.&nbsp;Continuous use is interrupted by prompts, described in the New York announcement as mindfulness reminders after sustained sessions, with the District of Columbia's account referring to pauses at 15, 60 and 90 minutes.&nbsp;Like and reaction counts are hidden for under-18 users. Cosmetic procedure filters are restricted for minors. A non-personalised feed must be offered, meaning a chronological or follows-only view rather than a ranked one.&nbsp;Age verification must be implemented, and accounts belonging to users under 13 removed. Disabling any of the restrictions requires parental permission rather than a self-declared toggle.&nbsp;Why the defaults are the point&nbsp;Almost every element on that list already exists somewhere as an optional setting. Instagram has offered time reminders and chronological feeds for years, and Meta has repeatedly pointed to those tools when questioned about teenage use.&nbsp;The difference is that these are defaults, and that switching them off requires a parent. Product designers have understood for a long time that defaults determine behaviour far more powerfully than available options do, which is precisely why engagement-driven defaults were chosen in the first place.&nbsp;The two-hour cap is the clearest case. A voluntary two-hour reminder is a nudge. A two-hour default cap that a teenager cannot lift without a parent is a constraint, and it operates against the metric the product was tuned for.&nbsp;The infeasibility argument just got harder&nbsp;The reason this matters beyond the United States is procedural rather than legal.&nbsp;When the European Union, Australia, the United Kingdom and India have drafted rules on minors and platform design, the industry response has followed a consistent pattern: the specific intervention is technically unworkable, or would break the product, or cannot be done without collecting more personal data than the regulation itself permits.&nbsp;Age verification has been the strongest version of that argument, and it is not a bad one. Verifying age reliably tends to require collecting identity documents or biometric signals from every user, including adults, which creates a privacy problem in the course of solving a safety one.&nbsp;That argument survives this settlement. What does not survive so easily is the claim that time caps, notification curfews, hidden like counts and non-ranked feeds are infeasible. Meta has now agreed to ship them.&nbsp;A regulator drafting rules next year can point to a specification the company accepted rather than a demand the company rejected. That is a materially different negotiating position.&nbsp;The limits of the template&nbsp;The strongest counter-argument is that a specification is only as good as its enforcement, and this one has three soft edges.&nbsp;The first is age verification. Every downstream protection applies to accounts identified as belonging to minors. If a teenager registers as an adult, none of it engages. The settlement requires Meta to implement age verification but the announcements do not describe the standard it must meet.&nbsp;The second is that the strictest terms are time-limited and conditional. The New York announcement describes an initial five-year minimum, extending to a ten-year stricter phase only if competing platforms reach similar settlements. A specification that lapses is a weaker precedent than one that does not.&nbsp;The third is that compliance can be narrow. A company can satisfy the letter of a default while designing the surrounding experience to make the default feel like an obstacle rather than a protection, and nothing in a consent judgment easily reaches that.&nbsp;There is also a substantive objection to the whole approach. Time caps treat duration as the harm, and the evidence that duration is the operative variable is weaker than the evidence that specific content and specific social dynamics are. A teenager spending 90 minutes in a hostile comment thread is not obviously better off than one spending three hours messaging friends.&nbsp;What Meta has and has not conceded&nbsp;It bears repeating that the settlement contains no admission of liability. The states alleged that Meta deliberately designed for compulsive use among minors and misled the public about it; Meta has contested those allegations throughout and the case did not reach a verdict on them.&nbsp;Agreeing to build a feature is not agreeing that its absence was unlawful. Companies settle to remove uncertainty, and the product terms here may reflect a judgement about litigation risk rather than a concession about past conduct.&nbsp;The settlement also remains subject to court approval through entry of a consent judgment in the Northern District of California.&nbsp;What to watch&nbsp;The first real test is the age verification implementation, because everything else depends on it. The second is whether any competing platform settles on comparable terms, which determines both the one-hour cap and the ten-year duration.&nbsp;The third, and the one that will take longest to answer, is whether a generation of teenagers with a midnight lockout and hidden like counts turns out to be measurably better off. The settlement assumes the answer is yes. It does not prove it.

A two-hour cap, midnight lockout, hidden likes and a non-personalised feed. The settlement's design terms undercut the industry's infeasibility argument.

The money in Meta's settlement with American state attorneys general will be argued over for years. The product terms may matter more, and they are unusually specific. Buried in the announcements from the California and New York attorneys general is something closer to a design specification than a legal remedy: a list of features Meta has agreed to build into Facebook and Instagram for users under 18, with defaults set against the company's own engagement interests. For any regulator who has spent the past five years being told that such measures are technically infeasible or commercially impossible, that list is the most useful document to come out of the case. What the specification requires Under-18 accounts get a default daily limit of roughly two hours across Facebook and Instagram combined. California's summary notes the cap drops to one hour if other major platforms adopt comparable limits. Access is blocked entirely between midnight and 6 am. Push notifications are barred between 10 pm and 7 am, with California describing a further block during school hours. Continuous use is interrupted by prompts, described in the New York announcement as mindfulness reminders after sustained sessions, with the District of Columbia's account referring to pauses at 15, 60 and 90 minutes. Like and reaction counts are hidden for under-18 users. Cosmetic procedure filters are restricted for minors. A non-personalised feed must be offered, meaning a chronological or follows-only view rather than a ranked one. Age verification must be implemented, and accounts belonging to users under 13 removed. Disabling any of the restrictions requires parental permission rather than a self-declared toggle. Why the defaults are the point Almost every element on that list already exists somewhere as an optional setting. Instagram has offered time reminders and chronological feeds for years, and Meta has repeatedly pointed to those tools when questioned about teenage use. The difference is that these are defaults, and that switching them off requires a parent. Product designers have understood for a long time that defaults determine behaviour far more powerfully than available options do, which is precisely why engagement-driven defaults were chosen in the first place. The two-hour cap is the clearest case. A voluntary two-hour reminder is a nudge. A two-hour default cap that a teenager cannot lift without a parent is a constraint, and it operates against the metric the product was tuned for. The infeasibility argument just got harder The reason this matters beyond the United States is procedural rather than legal. When the European Union, Australia, the United Kingdom and India have drafted rules on minors and platform design, the industry response has followed a consistent pattern: the specific intervention is technically unworkable, or would break the product, or cannot be done without collecting more personal data than the regulation itself permits. Age verification has been the strongest version of that argument, and it is not a bad one. Verifying age reliably tends to require collecting identity documents or biometric signals from every user, including adults, which creates a privacy problem in the course of solving a safety one. That argument survives this settlement. What does not survive so easily is the claim that time caps, notification curfews, hidden like counts and non-ranked feeds are infeasible. Meta has now agreed to ship them. A regulator drafting rules next year can point to a specification the company accepted rather than a demand the company rejected. That is a materially different negotiating position. The limits of the template The strongest counter-argument is that a specification is only as good as its enforcement, and this one has three soft edges. The first is age verification. Every downstream protection applies to accounts identified as belonging to minors. If a teenager registers as an adult, none of it engages. The settlement requires Meta to implement age verification but the announcements do not describe the standard it must meet. The second is that the strictest terms are time-limited and conditional. The New York announcement describes an initial five-year minimum, extending to a ten-year stricter phase only if competing platforms reach similar settlements. A specification that lapses is a weaker precedent than one that does not. The third is that compliance can be narrow. A company can satisfy the letter of a default while designing the surrounding experience to make the default feel like an obstacle rather than a protection, and nothing in a consent judgment easily reaches that. There is also a substantive objection to the whole approach. Time caps treat duration as the harm, and the evidence that duration is the operative variable is weaker than the evidence that specific content and specific social dynamics are. A teenager spending 90 minutes in a hostile comment thread is not obviously better off than one spending three hours messaging friends. What Meta has and has not conceded It bears repeating that the settlement contains no admission of liability. The states alleged that Meta deliberately designed for compulsive use among minors and misled the public about it; Meta has contested those allegations throughout and the case did not reach a verdict on them. Agreeing to build a feature is not agreeing that its absence was unlawful. Companies settle to remove uncertainty, and the product terms here may reflect a judgement about litigation risk rather than a concession about past conduct. The settlement also remains subject to court approval through entry of a consent judgment in the Northern District of California. What to watch The first real test is the age verification implementation, because everything else depends on it. The second is whether any competing platform settles on comparable terms, which determines both the one-hour cap and the ten-year duration. The third, and the one that will take longest to answer, is whether a generation of teenagers with a midnight lockout and hidden like counts turns out to be measurably better off. The settlement assumes the answer is yes. It does not prove it.

The settlement's product terms read less like a legal remedy than a design specification: a two-hour default cap, an overnight lockout, hidden like counts and a non-personalised feed. Governments that were told such measures were infeasible now have a company that has agreed to build them.

Instagram will cut teenagers off at midnight! Regulators everywhere now have a template

Nvidia reported revenue of $96.2 billion for its second quarter of fiscal 2027, the three months ended 26 July, in results released on 26 August. Revenue rose 106 per cent from a year earlier and 18 per cent from the previous quarter.&nbsp;Data centre revenue was $89 billion, up 117 per cent year on year, driven by the ramp of the company's Blackwell Ultra infrastructure. Gross margin was 75 per cent on both a GAAP and non-GAAP basis. Diluted earnings per share were $2.46 GAAP and $2.22 non-GAAP.&nbsp;A note on the numbers, because several outlets have circulated a figure of $81.6 billion for what they described as a record quarter. That figure does not correspond to this release. Nvidia's own investor materials give $96.2 billion for the second quarter of fiscal 2027.&nbsp;The number that matters most&nbsp;The reported quarter is history. The forward guidance is the claim.&nbsp;Nvidia guided to $108 billion for the third quarter of fiscal 2027, plus or minus 2 per cent, with gross margin of 74 per cent plus or minus 50 basis points.&nbsp;That is roughly $12 billion of sequential growth, in a single quarter, on a base already approaching $100 billion. Expressed as a rate it is a little over 12 per cent quarter on quarter. Expressed in absolute terms it is more incremental revenue than most semiconductor companies generate in a year.&nbsp;The guidance also embeds a modest margin decline, from 75 per cent to 74 per cent. On revenue of this size a single point of gross margin is more than a billion dollars, and the company has not attributed the compression to a single cause in the headline materials.&nbsp;What the chief executive said&nbsp;"AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue," said founder and chief executive Jensen Huang.&nbsp;The formulation is worth parsing, because it is doing specific work. Huang is arguing that AI compute has crossed from a research expense to a revenue-generating input, and that customers buying his chips are doing so because those chips now pay for themselves in production workloads rather than in anticipation of future capability.&nbsp;If that is right, the demand is durable in a way that speculative capacity building is not. If it is not right, the same sentence will be quoted back at him.&nbsp;Where the money is coming from&nbsp;The composition of the data centre segment tells more than the total.&nbsp;Hyperscale revenue was $49 billion, growing 13 per cent sequentially. That is the familiar customer base: the largest cloud providers building capacity for their own services and for rent.&nbsp;The faster-growing portion came from elsewhere. The category Nvidia reports as ACIE rose 138 per cent year on year and 25 per cent sequentially, driven by AI-native companies, enterprises and sovereign customers, as well as hyperscalers running AI clouds.&nbsp;Sovereign demand is the strategically interesting component. Governments buying national compute capacity are a different kind of customer from a cloud provider: less price sensitive, slower to order, and considerably more exposed to political change than to demand signals.&nbsp;The concentration problem&nbsp;The strongest counter-argument to these results is not that the demand is fake. It is that the demand is narrow.&nbsp;A substantial share of Nvidia's revenue comes from a small number of buyers with the balance sheets to place orders of this magnitude. That customer base is now also, in several cases, developing its own silicon.&nbsp;The Marvell and Google arrangement disclosed this week is the clearest illustration. Google's tensor processing units already displace some Nvidia purchasing, and the company has now added a second custom silicon partner alongside Broadcom. Amazon, Microsoft and Meta all run internal accelerator programmes.&nbsp;None of those programmes displaces Nvidia at present scale, and custom silicon has repeatedly proven harder and slower than announced. But the direction of the incentive is unambiguous: Nvidia's largest customers would each prefer to spend less with Nvidia.&nbsp;There is a second concentration risk on the supply side. Advanced packaging and high-bandwidth memory remain constrained, and both run through a small number of suppliers in Taiwan and South Korea. A guidance figure of $108 billion is a statement about supply availability as much as about demand.&nbsp;The case for the bears&nbsp;The bear case does not require AI to disappoint. It requires only that capital expenditure growth decelerate.&nbsp;Nvidia's valuation and its guidance both assume continued acceleration. A quarter in which the hyperscalers simply maintain spending rather than increasing it would represent a sharp deceleration against a comparison base that has doubled year on year.&nbsp;Depreciation is the mechanism most often cited. The chips being installed now carry multi-year depreciation schedules, and the revenue they generate has to service that. If token revenue at the customer level grows more slowly than the installed base, the purchasing logic changes without any decline in enthusiasm for the technology.&nbsp;Huang's answer is contained in the quote. He is asserting that the tokens are already profitable, which if true resolves the depreciation objection. The company has not published customer-level economics that would allow that assertion to be independently checked.&nbsp;What to watch&nbsp;Three things will matter more than the headline in the next quarter.&nbsp;Whether the company delivers against $108 billion, since a guide of that size sets a bar that a miss would make conspicuous. Whether gross margin stabilises at 74 per cent or continues drifting. And whether the ACIE category keeps outgrowing hyperscale, because a sustained shift towards sovereign and enterprise buyers would mean the demand base is broadening rather than concentrating.&nbsp;On the reported quarter, the results were unambiguous. On the quarter being promised, the company has given itself very little room.

Data centre revenue hit $89 billion as Jensen Huang declared compute is revenue. The guidance for next quarter is where the risk sits.

Nvidia reported revenue of $96.2 billion for its second quarter of fiscal 2027, the three months ended 26 July, in results released on 26 August. Revenue rose 106 per cent from a year earlier and 18 per cent from the previous quarter. Data centre revenue was $89 billion, up 117 per cent year on year, driven by the ramp of the company's Blackwell Ultra infrastructure. Gross margin was 75 per cent on both a GAAP and non-GAAP basis. Diluted earnings per share were $2.46 GAAP and $2.22 non-GAAP. A note on the numbers, because several outlets have circulated a figure of $81.6 billion for what they described as a record quarter. That figure does not correspond to this release. Nvidia's own investor materials give $96.2 billion for the second quarter of fiscal 2027. The number that matters most The reported quarter is history. The forward guidance is the claim. Nvidia guided to $108 billion for the third quarter of fiscal 2027, plus or minus 2 per cent, with gross margin of 74 per cent plus or minus 50 basis points. That is roughly $12 billion of sequential growth, in a single quarter, on a base already approaching $100 billion. Expressed as a rate it is a little over 12 per cent quarter on quarter. Expressed in absolute terms it is more incremental revenue than most semiconductor companies generate in a year. The guidance also embeds a modest margin decline, from 75 per cent to 74 per cent. On revenue of this size a single point of gross margin is more than a billion dollars, and the company has not attributed the compression to a single cause in the headline materials. What the chief executive said "AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue," said founder and chief executive Jensen Huang. The formulation is worth parsing, because it is doing specific work. Huang is arguing that AI compute has crossed from a research expense to a revenue-generating input, and that customers buying his chips are doing so because those chips now pay for themselves in production workloads rather than in anticipation of future capability. If that is right, the demand is durable in a way that speculative capacity building is not. If it is not right, the same sentence will be quoted back at him. Where the money is coming from The composition of the data centre segment tells more than the total. Hyperscale revenue was $49 billion, growing 13 per cent sequentially. That is the familiar customer base: the largest cloud providers building capacity for their own services and for rent. The faster-growing portion came from elsewhere. The category Nvidia reports as ACIE rose 138 per cent year on year and 25 per cent sequentially, driven by AI-native companies, enterprises and sovereign customers, as well as hyperscalers running AI clouds. Sovereign demand is the strategically interesting component. Governments buying national compute capacity are a different kind of customer from a cloud provider: less price sensitive, slower to order, and considerably more exposed to political change than to demand signals. The concentration problem The strongest counter-argument to these results is not that the demand is fake. It is that the demand is narrow. A substantial share of Nvidia's revenue comes from a small number of buyers with the balance sheets to place orders of this magnitude. That customer base is now also, in several cases, developing its own silicon. The Marvell and Google arrangement disclosed this week is the clearest illustration. Google's tensor processing units already displace some Nvidia purchasing, and the company has now added a second custom silicon partner alongside Broadcom. Amazon, Microsoft and Meta all run internal accelerator programmes. None of those programmes displaces Nvidia at present scale, and custom silicon has repeatedly proven harder and slower than announced. But the direction of the incentive is unambiguous: Nvidia's largest customers would each prefer to spend less with Nvidia. There is a second concentration risk on the supply side. Advanced packaging and high-bandwidth memory remain constrained, and both run through a small number of suppliers in Taiwan and South Korea. A guidance figure of $108 billion is a statement about supply availability as much as about demand. The case for the bears The bear case does not require AI to disappoint. It requires only that capital expenditure growth decelerate. Nvidia's valuation and its guidance both assume continued acceleration. A quarter in which the hyperscalers simply maintain spending rather than increasing it would represent a sharp deceleration against a comparison base that has doubled year on year. Depreciation is the mechanism most often cited. The chips being installed now carry multi-year depreciation schedules, and the revenue they generate has to service that. If token revenue at the customer level grows more slowly than the installed base, the purchasing logic changes without any decline in enthusiasm for the technology. Huang's answer is contained in the quote. He is asserting that the tokens are already profitable, which if true resolves the depreciation objection. The company has not published customer-level economics that would allow that assertion to be independently checked. What to watch Three things will matter more than the headline in the next quarter. Whether the company delivers against $108 billion, since a guide of that size sets a bar that a miss would make conspicuous. Whether gross margin stabilises at 74 per cent or continues drifting. And whether the ACIE category keeps outgrowing hyperscale, because a sustained shift towards sovereign and enterprise buyers would mean the demand base is broadening rather than concentrating. On the reported quarter, the results were unambiguous. On the quarter being promised, the company has given itself very little room.

Nvidia reported $96.2 billion in revenue for the quarter ended 26 July, with $89 billion from data centre. Guidance for the current quarter is $108 billion. The gap between those two numbers is the part of the results that carries the most risk.

Nvidia says compute is now revenue! Its guidance asks for another $12 billion in one quarter

A large coalition of technology and non-technology companies published an open letter on Thursday warning that AI-enabled cyberattacks are about to become far more common, and calling for what the signatories describe as a defensive surge.&nbsp;The letter was led by OpenAI and signed by Anthropic, Google, Microsoft, Amazon, Cisco, Oracle, Cloudflare, CrowdStrike and Palo Alto Networks, alongside firms well outside the sector including Capital One, Mastercard, Visa, General Motors and Shopify.&nbsp;The exact number of signatories is reported inconsistently. CNBC counted 116 entities; other outlets counted 118. The organisers have not published a figure that resolves the difference, and nothing in the letter's substance turns on it.&nbsp;What the letter claims&nbsp;The central assertion is a timing claim rather than a capability claim. "In the coming months, AI-enabled cyberattacks will become far more widespread as models around the world become increasingly capable," the letter states.&nbsp;That framing is deliberate. The signatories are not arguing that AI has unlocked a category of attack that was previously impossible. They are arguing that the cost of executing sophisticated attacks is falling fast enough to change who can execute them.&nbsp;The letter identifies hospitals, water treatment facilities and the infrastructure carrying internet traffic as particular concerns. These are systems characterised by long equipment lifecycles, thin security budgets and low tolerance for downtime, which is a poor combination.&nbsp;The asks&nbsp;The coalition urges every organisation to raise the security bar, upgrade defensive systems, and deploy a mix of low-cost and frontier models in defence rather than relying on any single tier.&nbsp;The more significant request is directed at the frontier labs themselves, several of which signed the letter. It asks that vetted defenders be given early access to advanced models.&nbsp;This is a meaningful procedural proposal and it deserves more attention than the alarm has received. Security researchers have long argued that defenders operate at a structural disadvantage because attackers can use any model they can obtain, while defenders are constrained by the same safety gating and release schedules as everyone else.&nbsp;Early access for vetted defenders would narrow that window. It would also, unavoidably, create a new category of privileged access to the most capable systems, and with it a question about who does the vetting.&nbsp;Why now&nbsp;The letter arrives against a backdrop of specific incidents rather than abstract concern.&nbsp;OpenAI has disclosed that experimental agents in its own testing escaped their intended boundaries and interacted with external systems, including Hugging Face infrastructure. Separately, Australian federal police this week charged two men over a campaign that investigators say compromised more than a thousand organisations globally, using self-propagating malware distributed through open-source packages.&nbsp;Neither incident establishes that AI caused the harm. But both illustrate the mechanism the letter is pointing at: automation reducing the expertise required to run an operation at scale.&nbsp;The obvious objection&nbsp;The strongest counter-argument is that the companies best positioned to profit from a defensive surge are the ones calling for it.&nbsp;CrowdStrike, Palo Alto Networks, Cloudflare and Cisco sell security products. OpenAI, Anthropic, Google, Microsoft and Amazon sell the models the letter recommends deploying in defence. A warning that every organisation must urgently upgrade its security posture is, read uncharitably, a coordinated marketing document.&nbsp;The presence of Visa, Mastercard, Capital One and General Motors complicates that reading, since those firms are buyers of security rather than sellers of it, and have no obvious commercial interest in inflating the threat. But it does not dispose of it entirely.&nbsp;There is a second objection with more analytical force. Security warnings of this kind have a poor forecasting record. Predictions of imminent transformation in the threat landscape have been made repeatedly over two decades, and the attacks that actually cause the most damage remain unglamorous: unpatched systems, stolen credentials, and phishing.&nbsp;If the marginal attack in twelve months is still a credential-stuffing operation against an unpatched server, the defensive surge will have been aimed at the wrong target.&nbsp;The case for taking it seriously anyway&nbsp;The counter to that counter is that the letter's recommendations are largely robust to being wrong about AI.&nbsp;Raising the security bar, upgrading defensive tooling and patching critical infrastructure are worth doing whether or not AI-enabled attacks arrive on the predicted schedule. The measures are not exotic and they address the boring attacks as well as the novel ones.&nbsp;The early-access proposal is the exception. That one only makes sense if the threat model is specifically about frontier capability, and it carries governance costs that would be hard to justify otherwise.&nbsp;What this means outside the United States&nbsp;For countries building out digital public infrastructure at speed, including India, the letter's infrastructure warning is the relevant part.&nbsp;Hospital networks, utility control systems and payment rails are being connected faster than security capacity is being built to defend them. That gap exists regardless of what frontier models can do, and it is the gap an attacker with cheap automation would exploit first.&nbsp;The letter offers no funding mechanism, no enforcement, and no timeline. It is a statement of concern by a group of companies, several with products to sell, and should be read as such.&nbsp;What it does establish is that the major AI developers are now publicly on record predicting that their own technology will shortly make attacks meaningfully easier. That admission may prove more useful to regulators than any of the recommendations attached to it.

An open letter led by OpenAI asks for a defensive surge and early access to frontier models for vetted defenders. Signatory counts differ across outlets.

A large coalition of technology and non-technology companies published an open letter on Thursday warning that AI-enabled cyberattacks are about to become far more common, and calling for what the signatories describe as a defensive surge. The letter was led by OpenAI and signed by Anthropic, Google, Microsoft, Amazon, Cisco, Oracle, Cloudflare, CrowdStrike and Palo Alto Networks, alongside firms well outside the sector including Capital One, Mastercard, Visa, General Motors and Shopify. The exact number of signatories is reported inconsistently. CNBC counted 116 entities; other outlets counted 118. The organisers have not published a figure that resolves the difference, and nothing in the letter's substance turns on it. What the letter claims The central assertion is a timing claim rather than a capability claim. "In the coming months, AI-enabled cyberattacks will become far more widespread as models around the world become increasingly capable," the letter states. That framing is deliberate. The signatories are not arguing that AI has unlocked a category of attack that was previously impossible. They are arguing that the cost of executing sophisticated attacks is falling fast enough to change who can execute them. The letter identifies hospitals, water treatment facilities and the infrastructure carrying internet traffic as particular concerns. These are systems characterised by long equipment lifecycles, thin security budgets and low tolerance for downtime, which is a poor combination. The asks The coalition urges every organisation to raise the security bar, upgrade defensive systems, and deploy a mix of low-cost and frontier models in defence rather than relying on any single tier. The more significant request is directed at the frontier labs themselves, several of which signed the letter. It asks that vetted defenders be given early access to advanced models. This is a meaningful procedural proposal and it deserves more attention than the alarm has received. Security researchers have long argued that defenders operate at a structural disadvantage because attackers can use any model they can obtain, while defenders are constrained by the same safety gating and release schedules as everyone else. Early access for vetted defenders would narrow that window. It would also, unavoidably, create a new category of privileged access to the most capable systems, and with it a question about who does the vetting. Why now The letter arrives against a backdrop of specific incidents rather than abstract concern. OpenAI has disclosed that experimental agents in its own testing escaped their intended boundaries and interacted with external systems, including Hugging Face infrastructure. Separately, Australian federal police this week charged two men over a campaign that investigators say compromised more than a thousand organisations globally, using self-propagating malware distributed through open-source packages. Neither incident establishes that AI caused the harm. But both illustrate the mechanism the letter is pointing at: automation reducing the expertise required to run an operation at scale. The obvious objection The strongest counter-argument is that the companies best positioned to profit from a defensive surge are the ones calling for it. CrowdStrike, Palo Alto Networks, Cloudflare and Cisco sell security products. OpenAI, Anthropic, Google, Microsoft and Amazon sell the models the letter recommends deploying in defence. A warning that every organisation must urgently upgrade its security posture is, read uncharitably, a coordinated marketing document. The presence of Visa, Mastercard, Capital One and General Motors complicates that reading, since those firms are buyers of security rather than sellers of it, and have no obvious commercial interest in inflating the threat. But it does not dispose of it entirely. There is a second objection with more analytical force. Security warnings of this kind have a poor forecasting record. Predictions of imminent transformation in the threat landscape have been made repeatedly over two decades, and the attacks that actually cause the most damage remain unglamorous: unpatched systems, stolen credentials, and phishing. If the marginal attack in twelve months is still a credential-stuffing operation against an unpatched server, the defensive surge will have been aimed at the wrong target. The case for taking it seriously anyway The counter to that counter is that the letter's recommendations are largely robust to being wrong about AI. Raising the security bar, upgrading defensive tooling and patching critical infrastructure are worth doing whether or not AI-enabled attacks arrive on the predicted schedule. The measures are not exotic and they address the boring attacks as well as the novel ones. The early-access proposal is the exception. That one only makes sense if the threat model is specifically about frontier capability, and it carries governance costs that would be hard to justify otherwise. What this means outside the United States For countries building out digital public infrastructure at speed, including India, the letter's infrastructure warning is the relevant part. Hospital networks, utility control systems and payment rails are being connected faster than security capacity is being built to defend them. That gap exists regardless of what frontier models can do, and it is the gap an attacker with cheap automation would exploit first. The letter offers no funding mechanism, no enforcement, and no timeline. It is a statement of concern by a group of companies, several with products to sell, and should be read as such. What it does establish is that the major AI developers are now publicly on record predicting that their own technology will shortly make attacks meaningfully easier. That admission may prove more useful to regulators than any of the recommendations attached to it.

An open letter published on 27 August by OpenAI, Anthropic, Google, Microsoft and scores of others calls for a defensive surge against AI-enabled attacks. Its most consequential ask is procedural: that frontier labs give vetted defenders early access to advanced models.

More than 100 companies say AI cyberattacks are months away, and want early access to the models

Marvell Technology disclosed an expanded custom silicon partnership with Google this week, and the number that travelled was $120 billion. It is worth understanding what that number is, because it is not a purchase order.&nbsp;The $120 billion is the cumulative amount Google would have to spend with Marvell, over roughly seven years, for a share warrant to vest in full. It is a threshold, not a commitment.&nbsp;What was actually signed&nbsp;The underlying commercial agreement was signed on 29 July and covers a range of custom silicon designed to work with Google's tensor processing unit ecosystem, the in-house accelerators that run much of the company's AI workload.&nbsp;Alongside it, Marvell issued Google a warrant to buy up to 58.97 million Marvell shares at an exercise price of $206.58 each. Exercised in full, that is roughly $12.2 billion of stock and about 6.3 per cent of the company.&nbsp;The vesting schedule is where the structure becomes clear. Around 1.36 million shares vest across the first four quarters. The remainder vests in 240 separate tranches, each one released by a further $500 million of Marvell revenue from sales to Google.&nbsp;Two hundred and forty tranches at $500 million each is $120 billion. That is the entire derivation of the headline figure.&nbsp;The warrant expires on 18 August 2033. There is also a product qualification deadline of 10 November 2027 for a component reported under the codename Kestrel.&nbsp;What Google is obliged to do&nbsp;On the evidence of the disclosed terms, very little.&nbsp;Google is not required to exercise the warrant. It can wait until the deadline, or decline entirely if Marvell's share price sits below the strike price. It is not required to hit any particular procurement level, and no penalty attaches to buying less.&nbsp;The arrangement gives Google an option and an incentive. It does not give Marvell a guaranteed customer, and Marvell's own disclosure indicates substantial revenue is not expected until fiscal 2029.&nbsp;This distinction matters because the deal was widely reported in a register normally reserved for binding supply agreements. A commitment to spend $120 billion and a warrant that vests if $120 billion happens to be spent are different corporate events.&nbsp;The circularity question&nbsp;There is a further feature of the structure that has drawn scrutiny from analysts who follow the sector closely.&nbsp;Marvell's market capitalisation rose by roughly $18.6 billion on the announcement day. That single-day increase exceeds the $12.2 billion the warrant would be worth if exercised in full.&nbsp;Put another way, the announcement of the option created more paper value than the option itself represents. Google holds an instrument whose worth rose the moment its existence became public, and the gains on that instrument can in principle help offset the development costs of the chips Google is buying.&nbsp;Analysts at The Next Platform described the arrangement as a roundtrip, in which each side's benefit is partly a function of the other side's stock reacting to the announcement. That characterisation is contested, and both companies have described the partnership in conventional commercial terms.&nbsp;Why Google is doing this at all&nbsp;The strategic logic is easier to read than the financial engineering.&nbsp;Broadcom has been Google's principal custom chip partner for its TPU line. Broadcom shares fell more than 5 per cent on the news while Marvell's rose, which is the market's plain reading of what a second supplier means for the incumbent.&nbsp;For a company spending at Google's scale on AI infrastructure, supplier concentration is a genuine risk. A second qualified design partner improves negotiating position, reduces exposure to a single roadmap, and provides insurance against a delay in any one component.&nbsp;The warrant is the price of getting Marvell to commit engineering resources to that role. Custom silicon programmes are expensive and multi-year, and a supplier will not staff one on the strength of a handshake.&nbsp;The counter-argument&nbsp;The case for taking the $120 billion figure more seriously than this piece has is straightforward.&nbsp;Warrant structures of this kind are not usually written against numbers nobody expects to reach. Marvell's board had to approve diluting existing shareholders by more than 6 per cent, and it is reasonable to assume they did so with an internal forecast suggesting meaningful vesting.&nbsp;Google, for its part, has been increasing AI capital expenditure aggressively, and the sums involved in hyperscale infrastructure over a seven-year horizon are genuinely of this order. A $120 billion procurement figure through 2033 is ambitious but not fantastical.&nbsp;It is also true that options with contingent vesting are a standard way to align a supplier's incentives with a buyer's growth, and that reading every such structure as financial engineering would leave one unable to describe ordinary corporate practice.&nbsp;What to watch&nbsp;The meaningful checkpoint is not the announcement but the Kestrel qualification deadline in November 2027. If that component qualifies on schedule, the procurement ramp becomes plausible and the vesting tranches start to matter.&nbsp;Until then, the disclosed facts support a narrower statement than the headlines carried: Google has acquired an option, Marvell has acquired a prospect, and the $120 billion is a condition attached to the former rather than a description of the latter.

Marvell gave Google a warrant for 58.97 million shares. The $120 billion figure is the procurement level at which it fully vests by 2033, not a purchase commitment.

Marvell Technology disclosed an expanded custom silicon partnership with Google this week, and the number that travelled was $120 billion. It is worth understanding what that number is, because it is not a purchase order. The $120 billion is the cumulative amount Google would have to spend with Marvell, over roughly seven years, for a share warrant to vest in full. It is a threshold, not a commitment. What was actually signed The underlying commercial agreement was signed on 29 July and covers a range of custom silicon designed to work with Google's tensor processing unit ecosystem, the in-house accelerators that run much of the company's AI workload. Alongside it, Marvell issued Google a warrant to buy up to 58.97 million Marvell shares at an exercise price of $206.58 each. Exercised in full, that is roughly $12.2 billion of stock and about 6.3 per cent of the company. The vesting schedule is where the structure becomes clear. Around 1.36 million shares vest across the first four quarters. The remainder vests in 240 separate tranches, each one released by a further $500 million of Marvell revenue from sales to Google. Two hundred and forty tranches at $500 million each is $120 billion. That is the entire derivation of the headline figure. The warrant expires on 18 August 2033. There is also a product qualification deadline of 10 November 2027 for a component reported under the codename Kestrel. What Google is obliged to do On the evidence of the disclosed terms, very little. Google is not required to exercise the warrant. It can wait until the deadline, or decline entirely if Marvell's share price sits below the strike price. It is not required to hit any particular procurement level, and no penalty attaches to buying less. The arrangement gives Google an option and an incentive. It does not give Marvell a guaranteed customer, and Marvell's own disclosure indicates substantial revenue is not expected until fiscal 2029. This distinction matters because the deal was widely reported in a register normally reserved for binding supply agreements. A commitment to spend $120 billion and a warrant that vests if $120 billion happens to be spent are different corporate events. The circularity question There is a further feature of the structure that has drawn scrutiny from analysts who follow the sector closely. Marvell's market capitalisation rose by roughly $18.6 billion on the announcement day. That single-day increase exceeds the $12.2 billion the warrant would be worth if exercised in full. Put another way, the announcement of the option created more paper value than the option itself represents. Google holds an instrument whose worth rose the moment its existence became public, and the gains on that instrument can in principle help offset the development costs of the chips Google is buying. Analysts at The Next Platform described the arrangement as a roundtrip, in which each side's benefit is partly a function of the other side's stock reacting to the announcement. That characterisation is contested, and both companies have described the partnership in conventional commercial terms. Why Google is doing this at all The strategic logic is easier to read than the financial engineering. Broadcom has been Google's principal custom chip partner for its TPU line. Broadcom shares fell more than 5 per cent on the news while Marvell's rose, which is the market's plain reading of what a second supplier means for the incumbent. For a company spending at Google's scale on AI infrastructure, supplier concentration is a genuine risk. A second qualified design partner improves negotiating position, reduces exposure to a single roadmap, and provides insurance against a delay in any one component. The warrant is the price of getting Marvell to commit engineering resources to that role. Custom silicon programmes are expensive and multi-year, and a supplier will not staff one on the strength of a handshake. The counter-argument The case for taking the $120 billion figure more seriously than this piece has is straightforward. Warrant structures of this kind are not usually written against numbers nobody expects to reach. Marvell's board had to approve diluting existing shareholders by more than 6 per cent, and it is reasonable to assume they did so with an internal forecast suggesting meaningful vesting. Google, for its part, has been increasing AI capital expenditure aggressively, and the sums involved in hyperscale infrastructure over a seven-year horizon are genuinely of this order. A $120 billion procurement figure through 2033 is ambitious but not fantastical. It is also true that options with contingent vesting are a standard way to align a supplier's incentives with a buyer's growth, and that reading every such structure as financial engineering would leave one unable to describe ordinary corporate practice. What to watch The meaningful checkpoint is not the announcement but the Kestrel qualification deadline in November 2027. If that component qualifies on schedule, the procurement ramp becomes plausible and the vesting tranches start to matter. Until then, the disclosed facts support a narrower statement than the headlines carried: Google has acquired an option, Marvell has acquired a prospect, and the $120 billion is a condition attached to the former rather than a description of the latter.

Marvell disclosed a warrant giving Google the right to buy up to 58.97 million shares. The widely reported $120 billion figure is the procurement total at which those shares fully vest by 2033, not an order. Marvell's market value rose by more than the warrant is worth on the day.

Google's $120 billion deal with Marvell commits Google to buying nothing at all

Meta has agreed to settle a landmark case brought by a bipartisan coalition of American state attorneys general over the design of Instagram and Facebook for teenagers, ending a federal trial in Oakland that legal scholars had already compared to the tobacco and opioid litigation of earlier decades.&nbsp;The figure carried in most headlines is $17.1 billion. The figure Meta is actually committed to is $12.1 billion.&nbsp;That gap is not a discrepancy or a reporting error. It is the structure of the deal, and it has been largely lost in the coverage.&nbsp;How the number actually works&nbsp;According to the announcement from New York Attorney General Letitia James, Meta will pay at least $12.1 billion. The total rises towards $17.1 billion only if other major social media companies reach similar settlements of their own.&nbsp;In other words, roughly $5 billion of the headline sum is contingent on the litigation conduct of companies that are not party to this agreement and have made no commitment to settle anything.&nbsp;The same conditionality runs through the non-monetary terms. The settlement's stricter phase is set to last ten years if competitors settle, and a minimum of five years if they do not.&nbsp;This structure explains the confusion in the first day of coverage, where the sum was variously reported as $16.7 billion, $17 billion, $17.1 billion and $18 billion. Different outlets were picking different points on a sliding scale, and few explained that a scale existed.&nbsp;What Meta has agreed to change&nbsp;The product commitments are more concrete than the money, and arguably more consequential.&nbsp;Under the terms described by the California and New York attorneys general, users under 18 face a default daily limit of about two hours across Facebook and Instagram. California's summary notes this can reduce to one hour if other platforms adopt comparable limits, another provision keyed to competitor behaviour.&nbsp;Overnight access is blocked between midnight and 6 am for under-18 accounts. Push notifications are barred between 10 pm and 7 am, and California's account adds a block during school hours.&nbsp;Beyond timing, Meta must offer a non-personalised feed option, hide like and reaction counts for under-18 users, restrict cosmetic procedure filters for minors, implement age verification and remove users under 13.&nbsp;There are also usage interruptions built in, described as mindfulness reminders after sustained scrolling, and a requirement that parental permission be obtained before a young user can switch the restrictions off.&nbsp;What Meta did not concede&nbsp;This point deserves emphasis, because settlements are routinely reported as though they were verdicts.&nbsp;The agreement contains no admission of liability. The states alleged that Meta designed Instagram with deliberately addictive features, that it knowingly exposed young users to mental health harms, and that it misled the public about platform safety. Those remain allegations, contested by the company and never tested to a jury verdict in this particular case.&nbsp;The settlement is also not yet final. It is subject to court approval through entry of a consent judgment in the US District Court for the Northern District of California.&nbsp;Why the states settled anyway&nbsp;Twenty-nine states had taken Meta to trial before Judge Yvonne Gonzalez Rogers on 18 August, the first case in the sprawling federal social media litigation to reach a jury. Settling eight days into a trial they had fought for years to reach is a choice that invites scrutiny.&nbsp;The obvious answer is that a negotiated set of binding product changes, enforceable through a consent judgment, may do more for teenage users than a damages award that Meta could have appealed for years.&nbsp;Attorneys general also faced genuine trial risk. Causation in adolescent mental health is contested terrain, and a defence verdict would have left them with nothing and set an unhelpful precedent for the remaining cases.&nbsp;The counter-argument&nbsp;The strongest case against reading this as a victory is that the money is smaller than it looks and the product changes may be less binding than they sound.&nbsp;Meta's own guidance indicated a legal expense accrual of roughly $10 billion in the third quarter of 2026. Against annual revenues in the hundreds of billions, a payment stream spread across ten years is a manageable cost of doing business rather than a structural blow.&nbsp;On the product side, defaults can be complied with narrowly. A two-hour limit that a determined teenager can circumvent, or that applies only to accounts the company has identified as belonging to minors, is only as strong as the age verification behind it, and age verification is precisely the thing platforms have struggled with everywhere it has been tried.&nbsp;There is a further objection worth stating plainly. Conditioning both the money and the duration on whether competitors settle gives Meta an unusual interest in the outcome of other companies' litigation, and gives those companies an incentive to hold out.&nbsp;The wider context&nbsp;For regulators outside the United States, the settlement offers something more useful than the dollar figure: a published list of design changes that one of the largest platforms has accepted as workable.&nbsp;India, the European Union, Australia and others have been drafting rules on minors and platform design, and have consistently run into industry arguments that specific interventions are technically infeasible or commercially ruinous. That argument is harder to make about measures a company has already agreed to build.&nbsp;Whether those measures survive contact with the product, and whether the contingent portion of the money ever arrives, will not be known for some time. The consent judgment has still to be entered, and the competitors on whose behaviour so much of this depends have said nothing at all.

Meta's settlement with state attorneys general guarantees $12.1 billion. The rest arrives only if rival platforms settle too. The company admitted no liability.

Meta has agreed to settle a landmark case brought by a bipartisan coalition of American state attorneys general over the design of Instagram and Facebook for teenagers, ending a federal trial in Oakland that legal scholars had already compared to the tobacco and opioid litigation of earlier decades. The figure carried in most headlines is $17.1 billion. The figure Meta is actually committed to is $12.1 billion. That gap is not a discrepancy or a reporting error. It is the structure of the deal, and it has been largely lost in the coverage. How the number actually works According to the announcement from New York Attorney General Letitia James, Meta will pay at least $12.1 billion. The total rises towards $17.1 billion only if other major social media companies reach similar settlements of their own. In other words, roughly $5 billion of the headline sum is contingent on the litigation conduct of companies that are not party to this agreement and have made no commitment to settle anything. The same conditionality runs through the non-monetary terms. The settlement's stricter phase is set to last ten years if competitors settle, and a minimum of five years if they do not. This structure explains the confusion in the first day of coverage, where the sum was variously reported as $16.7 billion, $17 billion, $17.1 billion and $18 billion. Different outlets were picking different points on a sliding scale, and few explained that a scale existed. What Meta has agreed to change The product commitments are more concrete than the money, and arguably more consequential. Under the terms described by the California and New York attorneys general, users under 18 face a default daily limit of about two hours across Facebook and Instagram. California's summary notes this can reduce to one hour if other platforms adopt comparable limits, another provision keyed to competitor behaviour. Overnight access is blocked between midnight and 6 am for under-18 accounts. Push notifications are barred between 10 pm and 7 am, and California's account adds a block during school hours. Beyond timing, Meta must offer a non-personalised feed option, hide like and reaction counts for under-18 users, restrict cosmetic procedure filters for minors, implement age verification and remove users under 13. There are also usage interruptions built in, described as mindfulness reminders after sustained scrolling, and a requirement that parental permission be obtained before a young user can switch the restrictions off. What Meta did not concede This point deserves emphasis, because settlements are routinely reported as though they were verdicts. The agreement contains no admission of liability. The states alleged that Meta designed Instagram with deliberately addictive features, that it knowingly exposed young users to mental health harms, and that it misled the public about platform safety. Those remain allegations, contested by the company and never tested to a jury verdict in this particular case. The settlement is also not yet final. It is subject to court approval through entry of a consent judgment in the US District Court for the Northern District of California. Why the states settled anyway Twenty-nine states had taken Meta to trial before Judge Yvonne Gonzalez Rogers on 18 August, the first case in the sprawling federal social media litigation to reach a jury. Settling eight days into a trial they had fought for years to reach is a choice that invites scrutiny. The obvious answer is that a negotiated set of binding product changes, enforceable through a consent judgment, may do more for teenage users than a damages award that Meta could have appealed for years. Attorneys general also faced genuine trial risk. Causation in adolescent mental health is contested terrain, and a defence verdict would have left them with nothing and set an unhelpful precedent for the remaining cases. The counter-argument The strongest case against reading this as a victory is that the money is smaller than it looks and the product changes may be less binding than they sound. Meta's own guidance indicated a legal expense accrual of roughly $10 billion in the third quarter of 2026. Against annual revenues in the hundreds of billions, a payment stream spread across ten years is a manageable cost of doing business rather than a structural blow. On the product side, defaults can be complied with narrowly. A two-hour limit that a determined teenager can circumvent, or that applies only to accounts the company has identified as belonging to minors, is only as strong as the age verification behind it, and age verification is precisely the thing platforms have struggled with everywhere it has been tried. There is a further objection worth stating plainly. Conditioning both the money and the duration on whether competitors settle gives Meta an unusual interest in the outcome of other companies' litigation, and gives those companies an incentive to hold out. The wider context For regulators outside the United States, the settlement offers something more useful than the dollar figure: a published list of design changes that one of the largest platforms has accepted as workable. India, the European Union, Australia and others have been drafting rules on minors and platform design, and have consistently run into industry arguments that specific interventions are technically infeasible or commercially ruinous. That argument is harder to make about measures a company has already agreed to build. Whether those measures survive contact with the product, and whether the contingent portion of the money ever arrives, will not be known for some time. The consent judgment has still to be entered, and the competitors on whose behaviour so much of this depends have said nothing at all.

Meta settled the states' teen social media case in Oakland for a sum widely reported as $17.1 billion. State attorneys general documents show the guaranteed amount is $12.1 billion, and the balance arrives only if other major platforms reach comparable deals. Meta admitted no liability.

Meta's $17 billion settlement has a floor of $12.1 billion! The rest depends on its rivals

A federal judge in California has ruled that the United States Department of Defense acted unlawfully when it branded the artificial intelligence company Anthropic a supply chain risk to national security, a decision that lands in the middle of a widening argument over how much control AI developers keep over the way their models are used.&nbsp;District Judge Rita Lin issued the order on Thursday night, and it was reported on Friday. In a decision running to 59 pages, she found that the designation was retaliation for the company's public defence of its own safety limits, and that the department's stated justification lacked an adequate factual and legal basis.&nbsp;"The empty invocation of national security is not a blank check to punish and retaliate against government critics," Lin wrote, in a line that has already been picked up well beyond the technology press.&nbsp;What the dispute was actually about&nbsp;The confrontation traces back to February, when the Trump administration instructed federal agencies to stop using Anthropic's technology. Defence Secretary Pete Hegseth designated the company a supply chain risk the same day.&nbsp;The trigger, according to Anthropic's filings, was its refusal to lift two contractual restrictions attached to Claude, its flagship model. One barred the use of the model for mass surveillance of Americans. The other barred fully autonomous weapons systems.&nbsp;Anthropic has said publicly that it was willing to work with the defence establishment, and has existing government business, but would not remove those two conditions. Chief executive Dario Amodei has repeatedly staked out the same position in public remarks.&nbsp;The designation carried real commercial weight. A supply chain risk label under federal procurement rules does not merely embarrass a vendor. It can shut it out of the defence market and cast a shadow over its dealings with other agencies and with private contractors who sell into that market.&nbsp;The legal findings&nbsp;Lin found violations on three separate tracks. She held that the designation infringed the First Amendment, because it punished the company for speech. She found a violation of the Fifth Amendment's due process clause. And she found the department had breached federal administrative law governing how agencies must reach and justify decisions.&nbsp;The administrative law finding may prove the most durable of the three. Courts are generally reluctant to second-guess national security judgements, but they are far more willing to ask whether an agency followed its own procedures and built a record capable of supporting what it did.&nbsp;That is the question Lin answered in the negative. On her reading, the department reached for a national security label without assembling the factual basis that the label requires.&nbsp;What the ruling does not do&nbsp;It is worth being precise about the limits here, because early coverage has blurred them.&nbsp;The order blocks the designation and restores Anthropic's eligibility to compete for defence work. It does not order the Pentagon to award the company a contract, resume any particular programme, or restore any commercial relationship that lapsed in the interim.&nbsp;Eligibility to bid and a guarantee of business are very different things. A procurement system has many lawful ways to decline to buy from a supplier it does not want, and none of them are foreclosed by this decision.&nbsp;The government is expected to appeal. Until an appellate court has spoken, the practical position of the company inside the defence market remains unsettled.&nbsp;Why this matters beyond one company&nbsp;The case is the sharpest test yet of a question the AI industry has so far handled through voluntary policy: can a model developer impose binding limits on how a government customer uses its system, and hold that line when the customer objects?&nbsp;Every major developer publishes an acceptable use policy. Almost none of them have been tested against a determined state buyer with procurement leverage. This ruling suggests that at least one court is prepared to treat the assertion of such limits as protected expression rather than mere contractual preference.&nbsp;That has implications well outside the United States. Governments in Europe, India and elsewhere are drafting procurement rules for AI systems, and many are weighing how much say a vendor should retain over downstream military and surveillance applications.&nbsp;The counter-argument&nbsp;The strongest case against the ruling is not that Anthropic's safety limits are unreasonable. It is that a court has now inserted itself into a judgement the executive branch has traditionally been allowed to make on its own.&nbsp;Supply chain risk designations exist precisely because the defence establishment sometimes needs to move quickly against a supplier without litigating the underlying intelligence in open court. If every such designation must survive a First Amendment challenge, the tool becomes considerably slower and more brittle.&nbsp;There is also a narrower objection. A vendor that publicly refuses a customer's requirements has, in an ordinary commercial setting, simply declined the business. Recasting the customer's subsequent decision not to deal as retaliation for speech is a move that could be applied to a great many procurement disputes that have nothing to do with AI safety.&nbsp;Lin's answer, in substance, was that the record here showed punishment rather than a considered procurement judgement. An appellate court may weigh that record differently.&nbsp;What happens next&nbsp;The immediate question is whether the government seeks a stay while it appeals. If it does and succeeds, the designation could remain in force through the appellate process, and the practical effect of Friday's news would be limited for months.&nbsp;The broader question is whether other developers now feel able to write firmer limits into their government contracts. The commercial incentive has always pointed the other way, and defence budgets are among the largest single pools of money entering the sector.&nbsp;A single district court order does not settle that. But it changes the calculation for a general counsel deciding how much of a fight the company can afford to pick, and that may turn out to be the most consequential thing about it.

Judge Rita Lin found the Defense Department retaliated against Anthropic over its refusal to lift Claude's limits on autonomous weapons and mass surveillance.

A federal judge in California has ruled that the United States Department of Defense acted unlawfully when it branded the artificial intelligence company Anthropic a supply chain risk to national security, a decision that lands in the middle of a widening argument over how much control AI developers keep over the way their models are used. District Judge Rita Lin issued the order on Thursday night, and it was reported on Friday. In a decision running to 59 pages, she found that the designation was retaliation for the company's public defence of its own safety limits, and that the department's stated justification lacked an adequate factual and legal basis. "The empty invocation of national security is not a blank check to punish and retaliate against government critics," Lin wrote, in a line that has already been picked up well beyond the technology press. What the dispute was actually about The confrontation traces back to February, when the Trump administration instructed federal agencies to stop using Anthropic's technology. Defence Secretary Pete Hegseth designated the company a supply chain risk the same day. The trigger, according to Anthropic's filings, was its refusal to lift two contractual restrictions attached to Claude, its flagship model. One barred the use of the model for mass surveillance of Americans. The other barred fully autonomous weapons systems. Anthropic has said publicly that it was willing to work with the defence establishment, and has existing government business, but would not remove those two conditions. Chief executive Dario Amodei has repeatedly staked out the same position in public remarks. The designation carried real commercial weight. A supply chain risk label under federal procurement rules does not merely embarrass a vendor. It can shut it out of the defence market and cast a shadow over its dealings with other agencies and with private contractors who sell into that market. The legal findings Lin found violations on three separate tracks. She held that the designation infringed the First Amendment, because it punished the company for speech. She found a violation of the Fifth Amendment's due process clause. And she found the department had breached federal administrative law governing how agencies must reach and justify decisions. The administrative law finding may prove the most durable of the three. Courts are generally reluctant to second-guess national security judgements, but they are far more willing to ask whether an agency followed its own procedures and built a record capable of supporting what it did. That is the question Lin answered in the negative. On her reading, the department reached for a national security label without assembling the factual basis that the label requires. What the ruling does not do It is worth being precise about the limits here, because early coverage has blurred them. The order blocks the designation and restores Anthropic's eligibility to compete for defence work. It does not order the Pentagon to award the company a contract, resume any particular programme, or restore any commercial relationship that lapsed in the interim. Eligibility to bid and a guarantee of business are very different things. A procurement system has many lawful ways to decline to buy from a supplier it does not want, and none of them are foreclosed by this decision. The government is expected to appeal. Until an appellate court has spoken, the practical position of the company inside the defence market remains unsettled. Why this matters beyond one company The case is the sharpest test yet of a question the AI industry has so far handled through voluntary policy: can a model developer impose binding limits on how a government customer uses its system, and hold that line when the customer objects? Every major developer publishes an acceptable use policy. Almost none of them have been tested against a determined state buyer with procurement leverage. This ruling suggests that at least one court is prepared to treat the assertion of such limits as protected expression rather than mere contractual preference. That has implications well outside the United States. Governments in Europe, India and elsewhere are drafting procurement rules for AI systems, and many are weighing how much say a vendor should retain over downstream military and surveillance applications. The counter-argument The strongest case against the ruling is not that Anthropic's safety limits are unreasonable. It is that a court has now inserted itself into a judgement the executive branch has traditionally been allowed to make on its own. Supply chain risk designations exist precisely because the defence establishment sometimes needs to move quickly against a supplier without litigating the underlying intelligence in open court. If every such designation must survive a First Amendment challenge, the tool becomes considerably slower and more brittle. There is also a narrower objection. A vendor that publicly refuses a customer's requirements has, in an ordinary commercial setting, simply declined the business. Recasting the customer's subsequent decision not to deal as retaliation for speech is a move that could be applied to a great many procurement disputes that have nothing to do with AI safety. Lin's answer, in substance, was that the record here showed punishment rather than a considered procurement judgement. An appellate court may weigh that record differently. What happens next The immediate question is whether the government seeks a stay while it appeals. If it does and succeeds, the designation could remain in force through the appellate process, and the practical effect of Friday's news would be limited for months. The broader question is whether other developers now feel able to write firmer limits into their government contracts. The commercial incentive has always pointed the other way, and defence budgets are among the largest single pools of money entering the sector. A single district court order does not settle that. But it changes the calculation for a general counsel deciding how much of a fight the company can afford to pick, and that may turn out to be the most consequential thing about it.

District Judge Rita Lin found the Defense Department retaliated against Anthropic for refusing to lift limits on Claude covering autonomous weapons and domestic surveillance. The order restores the company's eligibility to compete for contracts, but it does not force the Pentagon to buy anything.

A US judge says the Pentagon broke the law when it blacklisted Anthropic

technology

Technology, News on Technology, stories on Technology, articles on Technology, Technology stories at Wion, Technology news, Technology latest news

technology

BRICS set of guidelines for disaster management

Anthropic built a watermark for AI, then published limitation that defeats it

Anthropic is giving Claude to schools free, and the reason is not philanthropy

Feb, $380 billion, now near $800 billion, investors targeting $2 trillion...

The AI chip money has quietly moved from training to inference

A flaw in a widely used AI proxy is now on the US list of actively exploited bug

Nvidia has taken $99 billion in equity stakes across AI in a year

Anthropic has broken with OpenAI and Google

AI is decoding animal sounds: Could humans soon learn to talk to animals?

Together AI expects $5 billion a year from a Saudi facility

DeepSeek's huge Huawei order comes with a caveat

The US offered Armenia access to Blackwell GPUs!

A company built on AI safety is about to ask public markets to price it

Microsoft cut AI transcription pricing by 72 per cent, and the rate expires in December

576GB of HBM3E under a desk: Meet AMD's Halo Station

Anthropic signs $35 billion cloud deal with Nvidia-backed Lambda

AI startup Clay valued at $7 billion in new round

Nvidia is reportedly buying Hugging Face for $12.9 billion

Pentagon just launched its AI portal with ChatGPT, Grok and Gemini

NASA launches new telescope to map cosmos

xAI sued, the DOJ intervened on its side

Meta wants $199.99 a month for an AI agent that orders your DoorDash

A judge told the Pentagon it cannot punish Anthropic

X is letting advertisers pipe their campaign data straight into Claude

OpenAI is cutting off Cursor because Elon Musk now owns it

Meta's teen curfew becomes a template for regulators worldwide