Anthropic is in early discussions to use Microsoft’s Maia 200 AI chips to support its Claude models, according to multiple reports, in a deal that demonstrates how the AI sector is diversifying beyond NVIDIA GPUs as hyperscalers create their own in-house chips.
For Microsoft, a chip deal with Anthropic would validate years of investment in developing proprietary AI processors and position the company more directly against Amazon and Google, both of which offer internally developed AI chips to outside customers.
Anthropic’s interest appears driven by necessity. The company has acknowledged growing pressure on its compute resources amid rapid adoption of Claude and Claude Code. CEO Dario Amodei recently described the company as facing compute constraints, an example of how AI developers are competing for limited access to high-performance infrastructure.
“This deal is an illustration of how motivated Anthropic is to get compute anywhere it can get it and Microsoft is eager to get exposure to Anthropic’s growth,” Gil Luria, Managing Director at D.A. Davidson, told Techstrong.ai.
“If Anthropic wasn’t so motivated it wouldn’t sign up for what is—at best—the fourth best chip available for its compute.”
The High Cost of Compute
The pressure to line up robust AI infrastructure is wildly expensive. SpaceX recently disclosed that Anthropic plans to spend $1.25 billion per month through May 2029 for computing power.
As it has grown, Anthropic has relied on NVIDIA hardware for both training and inference workloads. But many AI developers are now looking for alternatives as GPU shortages persist and costs climb.
AI developers’ urgent demand for compute has reshaped the market for custom silicon. Amazon has aggressively promoted Trainium, designed for AI training, and Inferentia, built for inference. Google continues expanding deployment of its TPUs, which are considered one of the most mature alternatives to NVIDIA chips. Anthropic has inked agreements to use both AWS Trainium chips and Google TPUs.
Microsoft, later to market, is now attempting to gain similar success and establish Maia as a credible entrant in the lucrative AI chip market.
Microsoft debuted the Maia 200 in January as the second generation of its in-house AI accelerator effort. Fabricated by TSMC, the chip is deployed inside Microsoft data centers but has not yet been offered through Azure. CEO Satya Nadella claimed during a recent earnings call that the processor delivers more than 30% better tokens per dollar compared with other silicon in Microsoft’s infrastructure fleet.
The company has also touted Maia’s memory architecture and inference efficiency. The chip reportedly includes significant SRAM capacity built to improve performance for chatbot-style workloads handling vast volumes of user requests.
For Microsoft, there’s benefit in vertical integration. Building proprietary chips gives hyperscalers greater control over AI economics while reducing exposure to NVIDIA’s supply constraints and pricing power.
That move to lower costs is crucial as AI infrastructure costs escalate. Microsoft is reportedly on pace to spend as much as $190 billion on AI infrastructure and data center expansion in 2026, while competitors are also investing heavily to secure compute advantages.
Pros and Cons
Cloud providers are no longer simply funding AI startups. They now in some instances bundle financing, infrastructure, chips and cloud commitments into linked partnerships that increase long-term dependency between AI model builders and hyperscalers.
Microsoft’s relationship with Anthropic illustrates that process. The two companies expanded ties last year through a multibillion-dollar investment arrangement tied to large Azure spending commitments. Microsoft has also integrated Anthropic models into some of its AI offerings as its relationship with OpenAI becomes less exclusive.
For Anthropic, adopting Microsoft’s Maia has potential pros and cons. Access to alternative silicon should lower costs and diversify infrastructure supply. But it will also increase dependence on Microsoft, which could create pressures around pricing and control at a later date.
