Anthropic wants to lease Microsoft Maia 200 chip clusters! Carrying a $330 billion bill, it's not betting all its compute on Nvidia

Anthropic is in talks with Microsoft to lease its self-developed AI inference chip, Maia 200. If an agreement is reached, Anthropic would become one of the first major external customers. Anthropic has already made over $330 billion in computing commitments to the three major US cloud giants, while also leasing Google TPU and Amazon Trainium, and is currently in discussions with the UK startup Fractile. This multi-vendor strategy stands in stark contrast to OpenAI’s path of relying entirely on Nvidia. (Previous coverage: Anthropic rumored to be considering procurement of Microsoft's self-developed AI chip "Maia 200"! Breaking Nvidia's computing monopoly) (Background: AMD announces investment of over $10 billion in Taiwan to secure advanced packaging capacity) TSMC 3nm process, 216GB HBM3e memory, 7TB/s bandwidth, and 30% more tokens produced per dollar. These are the paper specifications of the Maia 200 inference accelerator, which Microsoft began deploying in its Azure data centers in January this year. Designed specifically for AI inference (rather than training), its primary task is to lower the operating costs of the Copilot AI assistant. Anthropic is now considering adding this chip to its computing toolkit, though discussions remain in the early stages and an agreement is not guaranteed. Anthropic’s multi-chip strategy is the most aggressive in the AI industry. In addition to Nvidia GPUs, Anthropic has long leased Google’s TPU and Amazon’s Trainium self-developed chips, and in early May, it was reported that the company was in talks with the UK chip startup Fractile to procure inference chips. Fractile was founded in 2022 by Oxford PhD Walter Goodwin and claims its "compute-in-memory" architecture can make large language model inference 100 times faster and reduce operating costs by 90%, with commercialization expected as early as 2027. Including Microsoft’s Maia 200, Anthropic is effectively reaching out to four independent self-developed chip supply chains simultaneously. CEO Dario Amodei recently admitted that the company is "facing difficulties in terms of computing power." With annualized revenue surpassing $30 billion, diversifying suppliers is not just about saving money, but also ensuring that production is not bottlenecked by a single vendor. Behind the multi-chip strategy lies astronomical cloud spending; Anthropic has made long-term computing commitments totaling over $330 billion to the three major cloud giants: - Google Cloud: $200 billion (five years) - AWS: Over $100 billion (ten years, in exchange for up to 5GW of computing capacity) - Azure: $30 billion In 2026 alone, computing expenditure is expected to exceed $20 billion, triple that of 2025. Using self-developed chips from various providers not only brings customized hardware advantages but also alleviates billing pressure through cost subsidies from cloud vendors, where every percentage point of cost savings translates to hundreds of millions of dollars. The relationship between Microsoft and Anthropic is warming up at an unusual speed, primarily because Microsoft and Nvidia each invested $5 billion in Anthropic late last year. Earlier this year, Microsoft enabled the Claude model by default across Copilot, with annual model procurement totaling approximately $500 million. To satisfy Anthropic’s surging demand for computing power, Microsoft has been increasing resource allocation since November 2025, diverting more existing Nvidia servers to Anthropic while building new server clusters for them, with the needs of some smaller cloud customers being "selectively" pushed back. What is the Microsoft Maia 200 chip? Maia 200 is Microsoft’s self-developed AI inference accelerator, based on TSMC’s 3nm process with 216GB HBM3e memory. Deployed in Azure in January 2026, it produces 30% more tokens per dollar than existing systems and is designed for inference rather than training. Why is Anthropic leasing self-developed chips from multiple vendors simultaneously? Anthropic has made over $330 billion in computing commitments to the three major cloud providers and expects to spend over $20 billion on computing in 2026. Diversifying the use of self-developed chips from various companies not only secures cost subsidies but also avoids being bottlenecked by a single supplier for computing power.