Anthropic wants to rent Microsoft's Maia 200 chip cluster! Carrying a $330 billion bill, they won't put all their computing power into Nvidia.

Anthropic is currently in talks with Microsoft about leasing its in-house AI inference chip, Maia 200. If an agreement is reached, it will become one of the first major external customers. Anthropic has made more than $330 billion in operational commitments to the three major U.S. cloud giants, while also leasing Google TPU and Amazon Trainium, and is even discussing cooperation with UK startup Fractile—its multi-supplier strategy and its complete departure from relying solely on Nvidia’s OpenAI route are starkly different.
(Background: Anthropic is reportedly considering the purchase of Microsoft’s in-house AI chip “Maia 200”! Breaking Nvidia’s computing bottleneck)
(Additional context: Supermicro AMD has announced plans to invest over $10 billion in Taiwan to secure advanced packaging capacity.)

Table of contents

Toggle

• No self-developed chip is left out
• The $330 billion strategy
• Frequently Asked Questions

TSMC’s 3nm process, 216GB HBM3e memory, 7TB/s bandwidth, and 30% more token output per dollar. These are the written specifications of the Maia 200 inference accelerator that Microsoft began deploying in Azure data centers this January. It is designed specifically for AI inference (not training), and its primary mission is to lower the operating costs of the Copilot smart assistant. Now Anthropic is considering adding this chip to its own computing toolset, but negotiations are still in the early stage and an agreement is not guaranteed.

No self-developed chip is left out

Anthropic’s multi-chip strategy is already the most aggressive in the AI industry. Besides Nvidia GPUs, Anthropic has long been renting Google’s TPU and Amazon’s Trainium in-house chips. In early May, reports emerged that it is in talks with UK chip startup Fractile to procure inference chips. Fractile was founded in 2022 by Walter Goodwin, a PhD from Oxford University, and claims that its “memory-computation fusion” architecture can make large language model inference 100 times faster and reduce operating costs by 90%, with commercial use expected as early as 2027.

With Microsoft Maia 200 added into the mix, Anthropic is essentially extending its reach to four independent in-house chip supply chains at the same time. CEO Dario Amodei recently admitted that the company “is facing difficulties with computing power.” With annual revenue already surpassing $30 billion, diversifying suppliers is not only about saving money—it is also about ensuring it won’t get stuck on production capacity by a single vendor.

The $330 billion strategy

Behind the multi-chip strategy lies an astronomical amount of cloud spending. Anthropic has made total long-term computing commitments of more than $330 billion to the three major cloud giants.

• Google Cloud: 200 billion (five years)
• AWS: more than 100 billion (ten years, in exchange for up to 5GW of computing capacity)
• Azure: 30 billion

Just for 2026, projected computing spending is expected to exceed $20 billion—three times that of 2025. Using each company’s in-house chips brings not only hardware customization advantages, but also eases billing pressure through cost subsidies from cloud providers. Every 1 percentage point of cost savings translates into tens of millions of dollars.

The relationship between Microsoft and Anthropic is warming up at a rare speed, mainly because at the end of last year Microsoft and Nvidia each invested $5 billion in Anthropic, and at the beginning of this year Copilot had the Claude model enabled as the default across the board, with annual model procurement of about $500 million. To feed Anthropic’s rapidly rising computing demand, Microsoft has been continuously tilting resources since November 2025—allocating more existing Nvidia servers to Anthropic, and building brand-new server clusters for it at the same time. Meanwhile, the needs of some smaller cloud customers are being “selectively” pushed back.

Frequently Asked Questions

What is the Microsoft Maia 200 chip?

Maia 200 is Microsoft’s in-house AI inference accelerator, based on TSMC’s 3nm process and 216GB HBM3e memory. It will be deployed on Azure in January 2026. Compared with existing systems, it produces 30% more tokens per dollar, and it is designed specifically for inference rather than training.

Why does Anthropic rent self-developed chips from multiple vendors?

Anthropic has made commitments of more than $330 billion in computing to the three major cloud providers, and it is expected to spend more than $20 billion on computing in 2026. Using different in-house chips not only allows it to receive cost subsidies, but also avoids being bottlenecked in compute supply by a single supplier.