Microsoft’s Maia 200 chips targeted by Anthropic in talks to scale compute
- Anthropic and Microsoft were in talks on May 21 for Anthropic to rent Azure servers using Maia 200 chips as Claude’s compute needs climbed. (cnbc.com) - Maia 200 is Microsoft’s in-house inference chip, introduced on January 26 with 30% better performance per dollar than the latest hardware in its fleet. (blogs.microsoft.com) - Next, investors will watch whether Microsoft opens Maia capacity beyond internal use and whether Anthropic confirms an Azure deployment agreement. (cnbc.com)
Anthropic is discussing a deal to use Microsoft’s Maia 200 chips, according to a May 21 CNBC report that cited people familiar with the talks. The arrangement under discussion would have Anthropic rent Azure servers powered by Microsoft-designed silicon as demand for Claude rises. (cnbc.com) The talks would deepen a relationship that already includes Microsoft’s $5 billion investment in Anthropic and Anthropic’s commitment to spend at least $30 billion on Azure. (blogs.microsoft.com) For Microsoft, the talks would amount to an early external test of a chip family it has so far kept inside its own data centers. For Anthropic, they would add another compute source at a time when frontier-model companies are spreading workloads across multiple clouds and chip platforms. (cnbc.com) CNBC said the discussions were still in early stages and could change. ### Why does Anthropic need another chip supplier now? Anthropic has been expanding its compute partnerships as usage of Claude and its developer tools grows, according to CNBC and follow-up coverage citing the same report. The company has already lined up access to Amazon Web Services’ Trainium chips and has also said it plans to use Google tensor processing units, leaving Microsoft’s Maia as a potential fourth major compute path alongside Nvidia-based infrastructure. (cnbc.com) The Information reported that Anthropic was considering Microsoft-powered servers for more complex workloads, according to Reuters and other outlets that summarized the report. That framing matters because frontier-model companies increasingly split training, inference and internal tooling across different hardware depending on cost, availability and software maturity, according to the companies’ own infrastructure announcements. (cnbc.com) ### What exactly is Maia 200? Microsoft introduced Maia 200 on January 26 as an inference accelerator built on TSMC’s 3-nanometer process. Scott Guthrie, Microsoft’s executive vice president for Cloud + AI, wrote that the chip includes 216GB of HBM3e memory, 7 TB/s of bandwidth and native FP8 and FP4 tensor cores. (cnbc.com) Microsoft said Maia 200 delivers 30% better performance per dollar than the latest generation hardware in its fleet and is designed to serve multiple models, including GPT-5.2 models from OpenAI. The company said the chip is deployed in its U.S. Central region near Des Moines, Iowa, with U.S. West 3 near Phoenix, Arizona, next in line. (thestar.com.my) ### Why would this matter for Microsoft beyond one customer? Microsoft has not generally made Maia 200 available to outside customers, CNBC reported. That means any Anthropic deployment would be notable not only because of Anthropic’s scale, but because it would show Microsoft is willing to expose first-party silicon to a major model developer outside its own internal stack. (blogs.microsoft.com) The broader backdrop is Microsoft’s push to reduce dependence on Nvidia by building more of its own AI infrastructure. Reuters, citing The Information, said a deal with Anthropic would be a boost for Microsoft’s in-house chip effort as Alphabet and Amazon pursue similar custom-silicon strategies. (blogs.microsoft.com) ### How does this fit with Anthropic’s other cloud ties? Microsoft’s November 2025 investment tied the companies more closely together but did not make Azure Anthropic’s only home. GeekWire reported at the time that Amazon remained Anthropic’s “primary cloud provider and training partner,” even as Microsoft secured broader access to Anthropic models across Foundry, GitHub Copilot, Microsoft 365 Copilot and Copilot Studio. (cnbc.com) That leaves Anthropic with a multi-cloud posture even after the Microsoft investment. If Maia 200 capacity is added, the next question will be whether Anthropic uses the chips for Claude inference, internal model operations or a narrower class of workloads. (thestar.com.my) CNBC said the talks were ongoing as of May 21, and Microsoft had not yet opened Maia 200 as a standard Azure offering. (cnbc.com) (geekwire.com)
