'Failed AI, AWS copy': Zuckerberg is destroying his own empire and investors refuse to see it

There is a version of Mark Zuckerberg's career that looks like visionary boldness: the founder who saw mobile before Google, who built a social graph that connected three billion people, who made a $19 billion bet on WhatsApp when everyone thought he was crazy. Then there is the version that has been playing out in real time since 2021 — a CEO who keeps announcing enormous bets on ideas that either collapse into expensive failures or turn out to be things he cannot actually execute. The Metaverse. Meta AI. And now, Meta Compute: the plan to build an Amazon Web Services rival. Each one announced with confidence. Each one, so far, stumbling.

The Metaverse: $90 Billion On Cartoon Zoom Calls

In late 2021, Zuckerberg renamed the company from Facebook to Meta and declared that the future was the Metaverse — a persistent virtual world where people would work, socialise, and live through VR headsets and digital avatars. Billions poured into Reality Labs, the division responsible for making this happen. The flagship product, Horizon Worlds, launched to widespread mockery. Critics compared it to low-quality cartoon-like virtual meeting rooms — a Zoom call with worse graphics and no obvious reason to use it. The numbers tell the story clearly: Reality Labs has lost approximately $90 billion in total since 2020. Annual losses went from $10 billion in 2021 to $13 billion in 2022, $16 billion in 2023, $17.7 billion in 2024, and $19.2 billion in 2025. In 2026's first quarter alone, the division lost over $4 billion. Despite all of this, the Metaverse generates roughly 1 percent of Meta's total revenue.

Meta AI: Exploiting Users, Banning Creators, Getting Facts Wrong

When the Metaverse bet started looking like a catastrophe, Zuckerberg pivoted hard to artificial intelligence. Meta AI was positioned as the company's answer to ChatGPT and Google Gemini — a conversational AI assistant embedded across Instagram, WhatsApp, Facebook Messenger, and the Quest headsets. The reality has been a different story. Meta's AI moderation systems have been falsely banning thousands of Instagram creators and small businesses, flagging ordinary family photos as child exploitation material, and removing accounts that represent years of work and primary income streams for their owners. A Change.org petition against wrongful Meta AI bans has gathered nearly 56,000 signatures. Meta's own Oversight Board confirmed in June 2026 that the company's AI enforcement lacks due process and transparency — meaning bans are made by automated systems with no human review and no meaningful appeals process. Meanwhile, AI-generated bot accounts with blue tick verification badges continue to thrive on the same platform where real human creators are being removed.

Now He Wants To Build An AWS

With the Metaverse burning money and Meta AI generating headlines for all the wrong reasons, Zuckerberg has announced his latest pivot: Meta Compute, an internal initiative to sell spare AI computing capacity to outside customers and compete directly with Amazon Web Services, Microsoft Azure, and Google Cloud. At Meta's shareholder meeting in May 2026, Zuckerberg said entering cloud computing was ‘definitely on the table,’ noting that companies were approaching Meta ‘almost every week’ to buy access to its AI models or spare computing power. The initiative is being led by Meta's infrastructure chief Santosh Janardhan, Daniel Gross of Meta Superintelligence Labs, and Meta President Dina Powell McCormick.

The Problem: Meta Has No Spare Compute

The immediate problem with Meta's cloud ambitions was identified by analyst Richard Windsor almost as soon as the plan leaked: Meta does not actually have spare compute to sell. Companies like CoreWeave and Nebius exist specifically because Meta — and other major AI companies — need to buy capacity from outside suppliers to meet their own AI workload demands. If Meta is buying compute from CoreWeave and Nebius today, it cannot simultaneously be renting surplus compute to paying customers. Building a real cloud business would require Meta to dramatically increase its data centre and chip capacity — which is exactly what Wolfe Research found when they ran the numbers, concluding that a genuine cloud business would push Meta's 2027 capital expenditure to $200 billion, up from previous estimates of $160 billion, and would likely force the company to raise capital before any cloud revenue materialised.

The Copycat Problem

The deeper issue is not just whether Meta Compute can work. It is the pattern it fits into. Facebook was built partly by copying features from rivals. Instagram Stories was a direct copy of Snapchat Stories. Reels was built to compete with TikTok after TikTok ate into Facebook's youth audience. WhatsApp's payment features mirror WeChat. The difference is that copying a product feature is relatively cheap and reversible. Copying an entire business model — building a cloud infrastructure giant from scratch to compete with companies that have been doing it for fifteen to twenty years — is a different order of magnitude entirely. AWS, Azure, and Google Cloud have decades of enterprise relationships, global infrastructure, and developer ecosystems that Meta does not have. The question the market is increasingly asking is whether Zuckerberg's instinct to respond to every competitive threat by building a replica is the right strategy — or whether it is a habit that has started to cost considerably more than it returns.