Anthropic has confirmed that it is assembling an in-house chip team to design hardware for Claude, formally validating speculation that had circulated for months.
Claude is currently supported by three major hardware stacks at once: Google’s TPUs, Amazon’s Trainium chips, and Nvidia GPUs. The report says Anthropic has signed a deal with Google for as many as 1 million TPUs. Now the company is moving to add a fourth track of its own.
Job listings exposed the effort before Anthropic confirmed it
Business Insider first spotted the signs in Anthropic’s official job board, and an Anthropic spokesperson confirmed the plan the next day. The openings currently listed are for a silicon engineer and a technical program manager, silicon.
The roles span front-end design, pre-silicon verification, physical design, design for test, analog and mixed-signal work, process and foundry engagement, packaging, and signal integrity. Taken together, they map to nearly the full flow of chip development.
The silicon engineer position offers annual compensation of $320,000 to $485,000. The listing says candidates must have “shipped silicon,” meaning they need hands-on experience taking chips through tape-out and into production rather than only working at the architecture stage.
One line in the job description also stands out: applicants should be able to “make significant decisions without the support structure of a large organization.” That points to a lean group with decision-making authority, not a large conventional chip division.
Clive Chan is identified as the technical lead on the effort
According to the report, the technical center of this initiative is Clive Chan, who joined Anthropic in June 2026. He previously worked at OpenAI and was part of the early team behind its in-house inference chip, Jalapeño.
Anthropic’s spokesperson also said the company will continue to follow a “multi-chip approach.” In practice, that means Anthropic plans to expand compute capacity while still using hardware from other suppliers alongside any chips it designs itself.
That position lines up with earlier reporting that Anthropic had held talks with Samsung and evaluated having Samsung manufacture a custom chip for the company.
Anthropic is joining a wider industry move
Anthropic is not the first AI company to take this route. The report says OpenAI has worked with Broadcom on an in-house chip called Jalapeño built for large language model inference in data centers. Google has long run its own models on TPUs. Meta has already designed and deployed its own chips. Mistral, according to the report, is also studying the same path.
The article frames this as a broader industry decision rather than an isolated move by one company.
Why custom silicon matters: inference cost and margins
The core argument is straightforward. Rented chips are designed for general-purpose use across many workloads. A custom chip can be optimized for a narrower set of demands, in this case Claude’s long-context processing and mixture-of-experts, or MoE, architecture.
Industry estimates cited in the report, based on precedents such as Google TPU and Amazon Trainium, suggest that a chip co-designed around a specific model architecture could reduce inference cost per token by roughly 30% to 50%. The article is careful on this point: those figures are not an official Anthropic commitment. They are industry estimates drawn from prior custom-chip experience, closer to an expectation than a guaranteed financial outcome.
Even at the lower end, the report says, that could still amount to a structural gross-margin shift for a company whose business depends on API usage and whose inference bill makes up a large share of operating expense.
Production may still be years away
The trade-off is time. Based on the industry timeline cited in the report, moving from team formation to mass production would, at the earliest, put a real chip launch in the 2028 to 2030 window.
That makes this less about what hardware runs the current generation of Claude and more about whether Anthropic, three years from now, can avoid being constrained by other companies’ supply schedules and capacity allocation.