Anthropic in Early Talks to Rent Microsoft’s Artificial Intelligence Chip, No Deal Signed Yet

Anthropic is in early talks to rent server capacity powered by Microsoft's custom Maia 200 artificial intelligence chip, according to two people familiar with the discussions The Information. No deal has been signed, and talks may not lead to one. But if one materializes, Anthropic becomes something notable: the first real customer for a chip Microsoft has been sitting on for five months.

Microsoft announced the Maia 200 in January, calling it a breakthrough inference accelerator built on TSMC's 3-nanometer process Microsoft. The specs are real: 140 billion transistors, 216 gigabytes of HBM3e memory, 272 megabytes of on-chip SRAM, more than 10 petaFLOPS in 4-bit precision. The company deployed it in two Azure regions, US Central near Des Moines and US West 3 near Phoenix. Scott Guthrie, Microsoft's cloud chief, said at the time it would bring 30 percent better performance per dollar than the prior generation.

Five months later, no external customer has been able to rent it. Microsoft has not listed Maia 200 in its public Azure pricing catalog or made it available through standard provisioning CNBC. The chips it announced with such fanfare have effectively been sitting in two data centers, waiting.

The backdrop to these talks is Anthropic's extraordinary growth — and the infrastructure crisis that growth has created. The company grew 80-fold in the first quarter, on an annualized basis, when it had planned for 10-fold CNBC. At Anthropic's developer conference earlier this month, chief executive Dario Amodei laid out the consequence plainly: That is the reason we have had difficulties with compute. The company is working as quickly as possible to add capacity and will, in Amodei's words, pass that compute on to you as soon as we can.

What that looks like in practice is a company stretched across every hardware relationship it can find. Anthropic has a $30 billion spending commitment to Microsoft Azure, announced in November. It has a deal with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, with the first systems expected online in 2027 Anthropic. It struck a deal with Elon Musk's SpaceX for more than 300 megawatts of capacity at the Colossus 1 data center in Memphis CNBC. It runs models on Amazon Web Services Trainium chips. Claude, its flagship model, is the only frontier AI available on all three major cloud platforms: AWS, Google Cloud, and Microsoft Azure Anthropic. That breadth of simultaneous dependence on every major hyperscaler sits uneasily with the independence Anthropic has long presented as central to its safety mission.

For Microsoft, landing Anthropic would give the company its first reference client for a chip it has been trying to position as competitive with Google's TPUs and Amazon's Trainium — the kind of third-party validation that opens doors with other AI companies. It would also give Microsoft a prominent customer at a moment when its long-standing partnership with OpenAI has grown more complicated. In April, the two companies amended their agreement: OpenAI's license became non-exclusive, Microsoft stopped paying revenue share to OpenAI, and OpenAI gained the right to serve its products across any cloud provider Microsoft. Microsoft remains the primary cloud partner, but the tie is looser than it was.

The chip is not without weaknesses. Reuters reported that Maia 200 uses older, slower high-bandwidth memory than Nvidia is shipping with its forthcoming Vera Rubin chips Reuters, a meaningful gap as AI labs push for every incremental performance gain. Even if the deal closes, Maia 200 would almost certainly cover only part of Anthropic's compute needs. The company is not reducing its Nvidia dependency through this; it is patching a gap.

Anthropic's revenue run-rate has passed $30 billion, up from $9 billion at the end of 2025, according to company figures published in April Anthropic. More than 1,000 business customers are now spending more than $1 million annually, a figure that doubled from 500 in under two months. The growth Amodei called extraordinary is also, plainly, an infrastructure crisis if the compute cannot keep pace.

The talks remain early. A person familiar cautioned that they may not lead to a final agreement. Neither company commented for this article. But the fact that they are happening at all says something about where the leverage actually sits in the AI infrastructure food chain: not with the lab raising the most money, and not with the cloud offering the most capacity, but with whoever has compute available when a competitor is desperate for it.