Anthropic is no longer just an AI model company. It is now in the chip business. The maker of Claude has quietly assembled one of the most aggressive custom silicon strategies of any AI startup — combining Google TPUs, AWS Trainium, and a landmark $21 billion deal with Broadcom to design and manufacture purpose-built AI accelerators. The move signals that Anthropic, valued at roughly $60 billion, is treating compute as a strategic asset rather than a commodity it rents from Nvidia.
What Anthropic Is Actually Building
Reuters reported this week, citing three sources, that Anthropic is weighing the possibility of designing its own AI chips entirely — going beyond partnerships to develop in-house silicon optimized for training and running Claude models. The plans are described as preliminary and the company could still opt to only purchase chips rather than design them. But the direction is clear. Anthropic has already expanded its partnership with Google Cloud and Broadcom to develop custom tensor processing units designed specifically for Claude workloads — one of the most significant commitments any frontier AI lab has made to custom silicon. Separately, Broadcom announced a $21 billion deal to supply Anthropic with nearly one million AI chips, delivering Google-designed TPU v7p chips in fully assembled rack-level AI systems ready for deployment in Anthropic data centers.
Why Anthropic Is Doing This
The economics are straightforward. Running Claude at the scale Anthropic now operates requires enormous compute — and renting Nvidia GPUs at market rates is extraordinarily expensive. Anthropic CFO Krishna Rao framed it plainly: the company is making its most significant compute commitment to date to keep pace with unprecedented growth. Annualised revenue crossed $30 billion in April 2026, up from approximately $9 billion at the end of 2025. The number of business customers spending more than $1 million annually doubled from 500 to over 1,000 in under two months. At that growth rate, the economics of custom silicon compound quickly. Google has demonstrated internally that TPUs can deliver significantly better performance per dollar than GPUs for transformer-based workloads at massive scale. Building on custom silicon rather than renting Nvidia capacity could save hundreds of millions of dollars annually.
The Three-Chip Strategy
Anthropic is not betting on a single chip architecture. Claude is currently trained and deployed across three platforms simultaneously: AWS Trainium chips on Amazon Bedrock, Google TPUs on Google Cloud Vertex AI, and Nvidia GPUs across general workloads. This multi-architecture approach lets Anthropic match specific workloads to the chips best suited for them — training runs on the infrastructure that handles them most efficiently, inference on whatever delivers the best cost-per-token at scale. Claude is the only frontier AI model currently available across all three major cloud platforms: AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure Foundry. The Broadcom deal adds a fourth dimension — purpose-built chips designed specifically around Claude model architectures rather than adapted from general-purpose designs.
The Broader Industry Shift
Anthropic is not alone. Every major AI company with sufficient scale is moving in the same direction. Amazon has been deploying Trainium2 chips for Anthropic workloads through Project Rainier. Google has been designing its own TPUs since 2016. Meta designed the MTIA chip for inference workloads. Microsoft launched its Maia 100 AI accelerator. JPMorgan projected in a June 2025 research note that chips from Google, Amazon, Meta, and OpenAI will make up 45% of the AI chip market by 2028, compared with 37% in 2024. What makes Anthropic unusual is that it is a startup — not a hyperscaler — making this move. A company founded four years ago investing in custom silicon at this scale signals extraordinary confidence in its long-term trajectory and a calculation that the cost advantages compound over time faster than the capital commitment.
What This Means for Nvidia
Nvidia is unlikely to lose AI hardware dominance overnight. Developers consistently prefer Nvidia GPUs because of the software stack — CUDA remains deeply embedded in the AI development ecosystem. But the margin risk is real. Every custom silicon program from a major AI company reduces the volume of Nvidia GPUs those companies purchase at market rates. Futurum Group analyst David Nicholson described it as death by a thousand cuts — multiple custom silicon accelerators gradually degrading the premium pricing Nvidia can command. Notably, Anthropic simultaneously signed a deal to purchase $30 billion of computing capacity from Microsoft Azure, which runs primarily on Nvidia Grace Blackwell and Vera Rubin GPUs. Anthropic is not replacing Nvidia — it is diversifying around it.
Timeline
Industry analysts estimate a custom chip program typically requires 18 to 24 months from design finalization to volume production. If Anthropic and Broadcom began design work in early 2025 — partnerships of this nature often precede their public announcement by months — the first Anthropic-optimized chips could be operational by late 2026 or early 2027. Broadcom has committed to supplying not just chips but networking and other components for next-generation AI racks through 2031, suggesting this is a decade-long infrastructure bet rather than a short-term procurement decision. Anthropic has committed to siting the vast majority of new compute in the United States — an extension of its November 2025 pledge to invest $50 billion in American computing infrastructure.
What It Means for Claude Users
For developers and businesses using Claude via API, custom silicon translates directly into lower inference costs and better availability. Purpose-built chips optimized for Claude model architectures deliver better performance per dollar than general-purpose GPUs for the specific operations Claude runs. As those chips come online, Anthropic gains pricing flexibility it currently lacks when renting capacity at market rates. The company has already cut API pricing by 67% since launch — from $15/$75 to $5/$25 per million tokens for Opus. Custom silicon is the infrastructure foundation that makes further price reductions sustainable at scale.