Anthropic is in early talks to rent servers powered by Microsoft-designed artificial intelligence chips, a possible deal that would give Microsoft’s custom silicon effort a high-profile customer as cloud providers try to reduce dependence on Nvidia hardware.
The discussions were reported by The Information and cited by Reuters. They are still preliminary and may not result in an agreement, Reuters reported, citing two people who spoke with executives involved in the talks. Microsoft told Reuters it does not comment on rumor or speculation. Anthropic did not immediately respond to Reuters’ request for comment.
The talks center on Microsoft’s Maia chips, part of the company’s effort to build more of the hardware needed to run large AI systems inside its own cloud. Microsoft introduced its second-generation Maia 200 accelerator in January, describing it as a chip built for AI inference, the stage when trained models generate responses for users. The company said Maia 200 is manufactured on Taiwan Semiconductor Manufacturing Co.’s 3-nanometer process and includes 216GB of HBM3e memory and 272MB of on-chip SRAM.
A deal with Anthropic would be important because Claude has become one of the main competitors to OpenAI’s ChatGPT. Anthropic needs large amounts of computing capacity to serve customers and improve its models. Microsoft needs outside proof that its own AI chips can support one of the leading frontier AI companies.
The cloud market has been moving in that direction. Google and Amazon built custom processors first for internal workloads and then made them available to AI customers. Microsoft has been trying to follow the same path while still buying Nvidia processors and running a large Azure cloud business.
Nvidia remains the center of the AI infrastructure market. Its GPUs are widely used for training and running large models, and demand has remained strong as AI developers, cloud companies and enterprise customers scale their systems. Microsoft, Amazon and Google have all pushed custom chips to reduce costs, improve control over supply and tailor hardware more closely to their own cloud software.
Anthropic has become one of the most important customers in that fight. In November, Microsoft, Anthropic and Nvidia announced a cloud infrastructure deal in which Anthropic committed to buying $30 billion in computing capacity from Microsoft Azure. Microsoft also agreed to invest up to $5 billion in Anthropic, while Nvidia said it would invest up to $10 billion in the startup.
That deal came as Microsoft’s long relationship with OpenAI became less exclusive. AP reported at the time that Microsoft had moved farther from its earlier cloud arrangement with OpenAI as OpenAI sought additional computing partnerships with Oracle, SoftBank and other data center and chip providers. Microsoft still holds a significant stake in OpenAI, but its AI strategy now includes more visible ties with Anthropic.
The possible Maia talks add another layer to that shift. Anthropic already works with Amazon and Google. Reuters reported in April that Google-parent Alphabet committed $10 billion in cash to Anthropic at a $350 billion valuation, with another $30 billion tied to performance targets. That report followed Amazon’s plan to invest up to $25 billion in the company.
Those deals show how fast Anthropic has become an infrastructure prize. The company is not only selling Claude to users and businesses. It is also shaping the capital-spending plans of the world’s largest cloud providers.
Microsoft’s Maia 200 was designed for inference rather than initial model training. That distinction matters. Training large AI models requires enormous compute over long runs. Inference happens every time a user asks a model to answer a question, write code, summarize a document or perform a task. As AI products gain users, inference can become a major cost center.
Microsoft has described Maia 200 as a chip intended to improve the economics of token generation. The company said the chip was built with native FP8 and FP4 tensor cores, a redesigned memory system and high-bandwidth memory intended to keep large models running efficiently.
Reuters reported that Maia 200 uses an older and slower generation of high-bandwidth memory than Nvidia’s upcoming Vera Rubin chips, while also carrying a large amount of SRAM that could help with high-volume chatbot workloads.
That is where Anthropic’s possible role becomes useful to Microsoft. Internal claims about cost and performance matter less than actual use by a demanding outside customer. Running Claude workloads on Maia hardware would provide a test of whether Microsoft can turn its chip program into a broader Azure product.
The talks are also a sign of how AI infrastructure has become a bargaining market. Model developers want capacity, lower costs and flexibility. Cloud companies want long-term customers. Chipmakers want their hardware embedded into major AI systems. No single supplier wants to be locked out of the next wave of spending.
Anthropic’s recent growth has increased that pressure. Reuters reported in April that the company’s annual run-rate revenue had passed $30 billion, up from about $9 billion at the end of 2025. The same report said Anthropic had signed several major compute deals and planned $50 billion in U.S. data center investment.
Those numbers explain why cloud providers are competing aggressively. A leading AI lab can consume tens of billions of dollars in compute, creating revenue for cloud platforms while giving those platforms influence over how AI systems are built and deployed.
The Microsoft talks remain early. No contract has been announced. No chip deployment schedule has been confirmed. The companies may decide not to proceed.
Still, the discussions show how quickly Microsoft’s AI hardware project has moved from internal infrastructure to potential external business. Maia started as a way to make Microsoft’s own AI workloads more efficient. A major Anthropic deal would push it closer to the model used by Amazon and Google, where custom chips become part of the cloud product catalog.
The outcome will matter beyond one supplier contract. AI companies are racing to cut the cost of serving models. Cloud companies are racing to make their infrastructure harder to replace. Nvidia is defending its lead against customers that are also becoming competitors in silicon.