Anthropic asks Microsoft to rent its in‑house chip capacity

Anthropic has held early talks with Microsoft about renting servers powered by Microsoft’s in-house Maia AI chips to support Claude, according to a May 21 report from The Information that was also cited by Reuters and Bloomberg. The discussions were described as preliminary, and Reuters said they may not lead to an agreement. Microsoft has not broadly opened Maia to external customers, making Anthropic a potentially important outside user if a deal is reached. Microsoft introduced Maia 200 in January as its latest in-house inference accelerator for Azure. The company said the chip is built on TSMC’s 3-nanometer process and includes 216GB of HBM3e memory, 7 TB/s of memory bandwidth and native FP8 and FP4 support. Microsoft said Maia 200 was designed to improve the economics of AI token generation inside its cloud infrastructure. (money.usnews.com) ### What exactly is Anthropic asking Microsoft for? The reported talks concern rented compute capacity, not an equity investment or a broad strategic partnership. Reuters said Anthropic was discussing renting servers powered by Microsoft-designed chips to meet rising demand for its AI services, while Bloomberg said the startup was seeking more computing power to serve Claude. (blogs.microsoft.com) The Maia angle matters because Microsoft’s custom silicon effort has so far been used mainly in its own data centers. CNBC reported on May 21 that Microsoft had not made Maia 200 chips available to customers, even though the company uses them internally and says they offer efficiency advantages over other silicon. ### Why would Anthropic look outside its existing suppliers? (money.usnews.com) Anthropic has already signed several large compute agreements this year. On May 6, the company said it had agreed to use all of the compute capacity at SpaceX’s Colossus 1 data center, giving it access to more than 300 megawatts of new capacity and more than 220,000 Nvidia GPUs within the month. Anthropic said that added capacity would improve service for Claude Pro, Claude Max and the Claude API. (cnbc.com) Amazon and Google are also part of Anthropic’s compute build-out. On April 20, Anthropic said it had expanded its collaboration with Amazon to secure up to 5 gigawatts of capacity for training and deploying Claude, including nearly 1 gigawatt of Trainium2 and Trainium3 capacity expected online by the end of 2026. Anthropic also said in April that a separate agreement with Google and Broadcom would bring multiple gigawatts of next-generation TPU capacity online starting in 2027. (anthropic.com) ### Why would Microsoft want Anthropic as a customer? Microsoft has spent more than two years building out first-party AI silicon as it tries to reduce dependence on Nvidia. The company said in January that Maia 200 was built specifically for inference workloads, the stage where models generate responses for users in production. Anthropic, whose Claude products are already deployed at large scale, would give Microsoft a high-profile external test case for that hardware. (anthropic.com) DigiTimes said on May 25 that an Anthropic-Microsoft arrangement could help broaden demand for application-specific AI chips across cloud supply chains. That assessment is DigiTimes’ characterization, but the reported talks fit a wider pattern in which cloud providers are trying to turn internal chip programs into rentable infrastructure. ### Is Anthropic building its own chips instead? (blogs.microsoft.com) The reported outreach points the other way. Reuters and Bloomberg both described Anthropic as seeking additional rented capacity rather than unveiling a proprietary chip program of its own. Anthropic’s public announcements this year have centered on securing outside compute from SpaceX, Amazon, Google and Broadcom, not on manufacturing in-house accelerators. (digitimes.com) Krishna Rao, Anthropic’s CFO, said in the company’s April announcement with Google and Broadcom that the partnership continued a “disciplined approach to scaling infrastructure” as customer demand grew. That language, together with the SpaceX and Amazon deals, shows Anthropic publicly framing compute as a supply and capacity question. ### What should readers watch next? (money.usnews.com) The next concrete sign will be whether Microsoft formally offers Maia-powered Azure capacity to outside customers or whether Anthropic discloses another compute agreement. Anthropic has already said additional Amazon Trainium capacity is due by the end of 2026, while its Google-Broadcom TPU expansion is expected to start coming online in 2027. (anthropic.com 1) (anthropic.com 2)