Reuters: Anthropic weighs Microsoft's Maia AI chips

Reuters reports that Anthropic is in early talks to rent servers powered by Microsoft-designed AI chips. If the talks become an agreement, Microsoft gets something its chip strategy badly needs: a major external AI customer that is not OpenAI.

The report is not confirmed by the companies. Reuters says the information comes from The Information, citing two people who spoke to executives involved in the discussions. Microsoft said it does not comment on rumour or speculation. Anthropic did not immediately respond to Reuters' request for comment.

Still, the story matters for CIOs and boards. Not because one possible server deal changes the market overnight, but because it shows where enterprise AI is moving: from model selection to infrastructure strategy. When a model provider like Anthropic considers Microsoft chips in addition to existing relationships with Amazon and Google, the important question becomes less “which model is best?” and more “which capacity chain can we safely build on?”

Maia could move from internal silicon to cloud product

Microsoft introduced Maia 200 in January as an in-house AI accelerator built for inference, meaning the production workload of serving model responses. Microsoft says Maia 200 is built on TSMC's 3 nanometer process, carries 216 GB of HBM3e memory and 272 MB of on-chip SRAM, and is designed for FP4 and FP8 workloads. Microsoft also said it would use Maia 200 in Microsoft Foundry, Microsoft 365 Copilot and internal model workflows.

The Reuters report makes that strategy more consequential. If Anthropic becomes a customer, Maia moves from an internal cost and performance lever into a possible commercial alternative for other model providers. That is the same playbook Alphabet and Amazon have used with TPUs and Trainium: build first for internal AI demand, then rent capacity to AI companies looking for alternatives to Nvidia.

For Microsoft, this is also about reducing concentration around OpenAI. Reuters notes that Microsoft has deepened ties with Anthropic in recent months, including by integrating Claude models into products such as Copilot, while its long-standing OpenAI relationship has loosened somewhat. That does not mean OpenAI is out. It means Microsoft is building more routes into the AI economy.

The real issue is supplier risk

For enterprises, the practical lesson is simple: an AI platform is not just an API subscription. It is a supply chain for models, chips, regions, memory, power, networking, logging and contracts.

If Maia becomes a real option for Anthropic, customers get another variable to assess. Latency, price, data location, reliability and access to capacity may differ depending on where the same model is served. Claude on AWS, Google or Microsoft could carry different operational terms, performance profiles and risk exposure.

That belongs in procurement now. AI contracts should not stop at data processing terms and model names. They should ask what infrastructure is used, where traffic terminates, which regions are available, how capacity shortages are handled, and what happens if the supplier moves workloads between chip types or cloud environments.

CFOs should read this as a pricing signal. Scarce and expensive Nvidia GPUs have been one of the key forces behind high AI service costs. More custom silicon could increase competition, but it can also create new forms of lock-in. Cheap capacity is worth less if it ties the company to one model platform, one control plane or one cloud architecture that is hard to exit.

CISOs should read it as a governance issue. When models move across infrastructure, logging, access control, audit, security domains and incident response move with them. AI agents that touch code, documents or customer data need traceability regardless of which chip served the model.

Do not buy AI as if infrastructure is invisible

This is still a reported discussion, not a signed deal. The right response is not panic. It is sharper due diligence.

Boards should ask three questions now:

• Which critical AI services already depend on one model provider, one cloud or one chip supply chain?
• Do our contracts give us visibility into where and how AI workloads are served?
• Do we have an exit plan if price, capacity, regulatory requirements or security requirements change?

The AI market is not becoming less complex. It is becoming more physical. Chips, data centers and regions behind the models will increasingly decide price, access and control. The Reuters report on Anthropic and Microsoft is a small data point, but it points to a larger shift: AI providers are now building strategy around capacity, not just model quality.

Sources and media

• Primary source: Reuters, “Anthropic in talks to use Microsoft's AI chips, The Information reports”, published May 21, 2026. https://www.reuters.com/technology/anthropic-talks-use-microsofts-ai-chips-information-reports-2026-05-21/
• Background: Microsoft, “Maia 200: The AI accelerator built for inference”, published January 26, 2026. https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/
• Reuters says the report is based on The Information, that talks are early and may not result in an agreement. Microsoft declined to comment on rumour or speculation. Anthropic did not immediately respond to Reuters' request.
• Thumbnail: OpenAI Image 2 / hogby.ai.