Anthropic in talks to rent Maia chips
- Anthropic was reported on May 21 to be in talks to rent Microsoft’s Maia AI chips for inference, adding another compute option beyond Nvidia. - Reuters reported Anthropic is paying SpaceX $1.25 billion a month for data-center capacity, underscoring how expensive model deployment has become. (msn.com) - Microsoft’s Maia 200 was introduced in January 2026 as an inference chip, and Anthropic and Microsoft have not publicly announced a deal. (blogs.microsoft.com)
Anthropic is in talks to use Microsoft’s AI chips for inference work, according to a report by The Information published on May 21 and summarized by Reuters and Bloomberg. The discussions would give the Claude maker another source of compute as demand for AI services rises and model operators try to reduce dependence on Nvidia hardware. Microsoft and Anthropic did not immediately respond to Reuters requests for comment, and Reuters said it could not independently verify the report. (msn.com) The reported talks land as Anthropic’s infrastructure bill has come into sharper view. (blogs.microsoft.com) Reuters reported on May 20 that Anthropic was nearing its first quarterly operating profit and had agreed to pay SpaceX $1.25 billion a month for computing power, a figure also reflected in Wall Street Journal coverage of a multi-year data-center capacity deal. That number has become a reference point for how much frontier-model inference and training can cost at current scale. ### What exactly is Anthropic said to be discussing with Microsoft? Reuters said The Information reported Anthropic was in talks to use Microsoft’s AI chips for inference tasks, citing two people who spoke to executives involved in the discussions. (streetinsider.com) Bloomberg, citing the same report, said the startup was seeking to boost computing power and meet demand for its services. Neither report described signed terms, deployment timing or deal size. Microsoft’s Maia line is part of a broader effort by large cloud companies to build alternatives to Nvidia-based systems. Bloomberg noted Microsoft, Alphabet’s Google and Amazon have all been designing their own chips to reduce reliance on Nvidia semiconductors. (msn.com) ### Why would Maia matter for Claude inference? Microsoft introduced Maia 200 on January 26 and described it as an inference accelerator built to improve the economics of AI token generation. In its product post, Microsoft said the chip uses TSMC’s 3-nanometer process, includes 216GB of HBM3e memory and is aimed at running large models more efficiently in Azure data centers. (streetinsider.com) Inference is the part of the AI business that serves answers to users after a model has been trained. If Anthropic adds Maia-powered servers, it would be seeking lower-cost or more available capacity for Claude responses rather than replacing all of its existing infrastructure, based on the description of Maia 200 and the Reuters characterization of the talks as inference-focused. (bloomberg.com) That is an inference from the reported scope of the discussions and Microsoft’s own description of the chip. ### How does this fit with Anthropic’s other compute relationships? Anthropic already sits at the center of a multi-cloud and multi-supplier network. (blogs.microsoft.com) Bloomberg said Microsoft’s chips would give Anthropic another way to run Claude, while noting Microsoft, Google and Amazon are all pursuing custom silicon. Reuters separately reported Anthropic’s massive SpaceX compute commitment. CoreWeave’s latest earnings offer another marker of demand around the same customer set. CoreWeave said on May 7 that its revenue backlog reached $99.4 billion, and CNBC reported the company had 10 clients committed to spending at least $1 billion on its products. (streetinsider.com) CoreWeave has also announced an Anthropic deal to host Claude workloads, while Meta added another $21 billion commitment a day earlier, CNBC reported. ### What does the CoreWeave backlog say about the market around Anthropic? CoreWeave reported first-quarter revenue of $2.08 billion and a $99.4 billion revenue backlog as of March 31. (bloomberg.com) The company said the quarter was its strongest bookings period on record and that it had surpassed 1 gigawatt of active power. CNBC reported CoreWeave’s customer roster now includes major AI companies and that Anthropic joined a platform already used by many leading model providers. That does not prove Maia talks will become a contract, but it shows Anthropic is already spreading workloads across multiple infrastructure partners as demand for Claude grows. (investors.coreweave.com) ### What happens next? May 21 is the date of the report, not of a confirmed contract. The next concrete step would be a public statement from Anthropic or Microsoft, an Azure product disclosure naming Anthropic, or future company filings that show additional chip or cloud commitments. Reuters said both companies had not commented when it published its report. (investors.coreweave.com) (streetinsider.com) (cnbc.com)