Anthropic confirms it is building custom silicon for Claude, making custom chips the price of entry for every top AI lab

An Anthropic spokesperson told Business Insider the company plans to "co-design hardware and models" so Claude runs faster and more efficiently "at the scale our customers need" 1. Reuters separately confirmed the plans and reported that Anthropic did not provide a timeline for its chips or say whether it intends to manufacture them itself 2.

Designing an advanced AI chip can cost roughly half a billion dollars, according to industry sources cited by Reuters 2. Anthropic's job listing seeks engineers who have "shipped silicon," with salaries reaching $485,000 1. The spokesperson said Anthropic will maintain a "multi-chip approach" with hardware from AWS, Google, Nvidia, and AMD remaining central 1. Nvidia stays in the stack. But the direction is unmistakable.

A decade of divergence, compressed into months

Google began the custom silicon movement in 2015, designing its Tensor Processing Units for the matrix math that neural networks depend on 3. For years, Google was the outlier. Then the pace compressed:

• 2015: Google begins using internally designed TPU processors 3
• 2018: Amazon Web Services announces Inferentia for AI inference 3
• 2020: Amazon Web Services unveils Trainium for AI training 3
• 2023: Meta begins producing its first MTIA chips 4
• June 2026: OpenAI unveils the Jalapeno inference chip, built with Broadcom 1
• August 2026: Anthropic confirms its custom silicon team 1
• September 2026: Meta's next-generation Iris chip enters manufacturing 4

The gap between Google and Amazon was three years. The gap between OpenAI, Anthropic, and Meta is three months. DA Davidson analysts estimated last year that Google's TPU business, combined with its DeepMind AI group, would be worth roughly $900 billion 3. That is the upside of going in-house. The half-billion-dollar design cost is the buy-in.

Why "co-design" is the strategic tell

The specific language matters. "Co-design hardware and models" means the chip architecture and the model architecture are developed together, each shaped to the other 1. An Nvidia GPU serves every customer's workload. A custom chip serves one model family. Every design decision, from memory bandwidth to interconnect topology, can be matched to a known computational pattern rather than a general one. That is a qualitatively different optimization, not just a cheaper version of the same component.

The advantage compounds at scale. When inference costs are measured across billions of queries, even modest efficiency gains from architecture-specific silicon translate into real savings. Forrester vice president and principal analyst Mike Gualtieri stated the logic plainly: "You can't become an AI titan if you are dependent on another company for chips" 5.

Nvidia's customers are building the exit ramp

Every lab going in-house still buys Nvidia GPUs and plans to keep doing so 1. Anthropic has separately committed to multiple gigawatts of Google TPUs 3. But OpenAI's Jalapeno targets inference specifically, and Meta's MTIA program covers training and recommendation workloads 4. The inference layer, where models serve end users and costs scale with usage, is where custom silicon delivers the fastest return.

This is producing a structural divide. Labs with the capital to design custom silicon gain a cost advantage at the inference layer that grows as their models scale. Labs without that capital face a widening gap they cannot close through model quality alone. AI startup Mistral's CEO has said the French company is considering its own chips 1. Whether smaller labs can afford to follow will determine whether the frontier stays open to new entrants or concentrates around the few labs wealthy enough to build their own hardware.