Anthropic is in early-stage talks with Microsoft to lease the Maia 200, an in-house AI inference accelerator. If finalized, Anthropic would become one of the first external clients for the chip. Built on TSMC's 3nm process, the Maia 200 packs 216GB of HBM3e memory and 7TB/s bandwidth, delivering 30% more tokens per dollar compared to existing systems. Designed exclusively for inference (not training), the chip was deployed in Azure data centers in January 2026 to lower the operating costs of Microsoft's Copilot assistant.
Betting on Every Custom Silicon
Anthropic's multi-chip strategy is among the most aggressive in AI. Beyond Nvidia GPUs, the company has long leased Google's TPUs and Amazon's Trainium chips. In early May, news emerged that Anthropic is also in talks with UK-based chip startup Fractile to purchase inference chips. Founded in 2022 by University of Oxford PhD Walter Goodwin, Fractile claims its "in-memory compute" architecture can speed up large language model inference by 100x while cutting operating costs by 90%, with commercial deployment expected by 2027 at the earliest. Adding Microsoft's Maia 200, Anthropic now has four independent custom chip supply chains within reach. CEO Dario Amodei recently admitted the company is "struggling with compute capacity." With annualized revenue exceeding $30 billion, diversifying suppliers is not just about cost savings—it's about ensuring production isn't bottlenecked by a single vendor.
A $330 Billion Bet
Behind this multi-chip strategy lies staggering cloud spending. Anthropic has made long-term compute commitments totaling over $330 billion to three cloud giants: Google Cloud $200 billion (five years), AWS over $100 billion (ten years, with access to up to 5GW compute capacity), and Azure $30 billion. In 2026 alone, compute spending is projected to exceed $20 billion—triple the 2025 figure. Using each cloud provider's custom chips not only provides hardware customization advantages but also allows Anthropic to benefit from cost subsidies, with every percentage point saved equating to hundreds of millions of dollars.
The Microsoft-Anthropic relationship has deepened rapidly. Late last year, Microsoft and Nvidia each invested $5 billion in Anthropic. In early 2025, Microsoft made Claude the default model across Copilot, with an annual procurement deal worth about $500 million. To fuel Anthropic's surging compute needs, Microsoft has since November 2025 been reallocating more of its existing Nvidia servers to Anthropic and building entirely new server clusters for the firm, while selectively deferring the requests of some smaller cloud customers.