Anthropic is building its own chips! Weeks after agreeing to pay Musk $1.25 billion a month for compute

Anthropic has spent this year signing some of the largest compute contracts in the industry, paying other companies for the hardware its models run on. It has now decided to start building the hardware itself.

What Was Confirmed

On August 5, Anthropic confirmed it is assembling an in-house silicon team to design custom chips for its Claude models. The company is recruiting semiconductor engineers, with roles advertised at salaries reaching $485,000.

The stated objective is efficiency at scale: running Claude faster and more cost-effectively than general-purpose hardware allows. The technical case rests on hardware-software co-design — building chip architecture around the specific computational patterns of Claude's attention mechanisms, rather than adapting the model to silicon designed for a broad range of workloads.

Anthropic is not walking away from its suppliers. The company says it will continue to rely on chips from AWS, Google, Nvidia and AMD under a multi-chip strategy, with its own silicon added as a further option rather than a replacement. Reporting from The Information last month indicated Anthropic has been scouting Samsung as a potential manufacturing partner.

The Context That Makes This Notable

Anthropic's compute position has become unusually exposed this year.

The company agreed to pay xAI roughly $1.25 billion a month through May 2029 for exclusive use of the Colossus 1 data centre in Memphis — more than $40 billion over the term — for access to 220,000 Nvidia processors drawing 300 megawatts. Multiple reports describe a clause permitting the compute to be reclaimed if Anthropic's models are found to harm humanity, though its exact contractual status has not been jointly confirmed.

That arrangement solved an immediate capacity problem and created a longer-term dependency: a company whose market position rests on being the careful, safety-focused laboratory now runs a substantial share of its workloads on infrastructure owned by a competitor whose chief executive had previously called it ‘woke’ and ‘evil.’

Designing its own silicon does not undo that. But it is the beginning of a path where Anthropic's compute costs and capabilities are shaped by decisions it controls.

Everyone Else Is Already Doing It

Anthropic is late to this rather than early. Google's TPUs have underpinned DeepMind's systems for years. Amazon has Trainium and Inferentia. Microsoft has its own accelerator programme. Meta has been developing MTIA chips. OpenAI unveiled its Broadcom-built Jalapeño chip in June.

The logic driving all of them is the same. At sufficient scale, the margin a chip vendor earns becomes one of the largest line items in an AI company's cost structure. Nvidia's gross margins on data centre products have been extraordinary throughout this cycle, and every dollar of that margin is a cost borne by its customers. Custom silicon is how the largest buyers stop paying it.

There is a strategic dimension too. Nvidia has increasingly financed its own demand — reportedly in talks to guarantee some $250 billion for OpenAI's Ohio campus, and committing $5 billion to Safe Superintelligence in a deal that returns as Nvidia hardware purchases. Custom silicon is the counterweight to that gravity.

The Difficulty Should Not Be Understated

Designing a competitive AI accelerator is among the hardest problems in engineering. It requires years, enormous capital, access to leading-edge foundry capacity that is already allocated, and a software stack mature enough that researchers will actually use it — the obstacle that has kept CUDA dominant regardless of competing hardware's specifications.

Anthropic is starting this while committed to $40 billion in rented compute, while competing in the steepest AI price war the industry has seen, and while its models are the subject of security disclosures from both the company itself and the UK government. The chip team is a bet on a position several years out.

That it is being made at all indicates how the economics look from inside a frontier laboratory: paying for someone else's silicon is no longer viewed as sustainable, even by a company that just agreed to pay a rival $1.25 billion a month to do exactly that.