Anthropic is in early talks with Microsoft to rent compute powered by Microsoft’s in-house Maia 200 chips to run Claude, according to The Information. The report pushed Microsoft shares up 2% in Thursday trading.
The center of the discussion is Maia 200, Microsoft’s second-generation AI accelerator introduced in January this year. Unlike Nvidia’s more general-purpose GPU lineup, Maia 200 is described as being built with a focus on inference. That makes it less suited to training large new models from scratch, but potentially more attractive for running existing models at better speed, lower power use, and lower cost.
Microsoft pushes Maia 200 beyond internal services
The report says Maia 200 is built on TSMC’s 3-nanometer process. Microsoft has already used Maia chips in internal products including Copilot. A deal with Anthropic would carry a different weight. It would turn Microsoft’s custom AI silicon from an internal tool into a product supporting an outside, high-profile customer.
That matters because Anthropic is already a major buyer of Azure compute. Under an agreement reached in late 2025, the company committed to purchase as much as $30 billion in Azure compute, still centered mainly on Nvidia’s Grace Blackwell architecture. But rising demand for Claude has increased pressure on compute supply, and relying too heavily on one chip vendor comes with both cost and sourcing risk.
Anthropic looks for alternatives to a single chip supplier
Using Maia chips would give Anthropic another path. The source material points to a “multi-cloud + multi-silicon” strategy, aimed at reducing dependence on Nvidia. There is also a cost angle. If Maia 200 can handle large volumes of inference work more efficiently, it could lower the expense of serving Claude at scale. For model providers, that is a practical consideration, not a theoretical one.
The report also notes that renting Microsoft’s current chip capacity could open the door to broader hardware cooperation later on. Anthropic may eventually provide design feedback for Microsoft’s next generation of chips so the hardware aligns more closely with the needs of Claude. At this stage, though, the talks remain preliminary and no final agreement has been announced.
From buying chips to selling compute
For Microsoft, the significance goes beyond one possible customer win. Cloud providers have spent heavily on Nvidia hardware, and those costs cut into margins. If Microsoft can persuade an outside AI company like Anthropic to adopt Maia 200, it would strengthen the commercial case for its own silicon and sharpen its position against Google’s TPU and Amazon’s Trainium offerings.
The broader shift is clear. Microsoft is trying to move from being mainly a buyer of AI chips to becoming a seller of AI compute built on its own hardware. If Anthropic becomes a Maia customer, that transition would take a visible step forward.