Anthropic Builds a Chip Team as AI Demand Surges

Anthropic has verified that it is assembling an internal team tasked with creating bespoke chips for its Claude AI models. This marks the first occasion the company has openly recognized a hardware initiative that had circulated as rumor since the spring.

The confirmation, reported by Ars Technica, followed a company statement and a job listing for a new "custom silicon team." Reuters reported in April that Anthropic was weighing whether to design its own AI chips, and the company has now said it is going ahead. A spokesperson framed the decision as an addition rather than a break, calling it "the latest step in our multi-chip approach" and saying Anthropic would keep using a "diversified hardware stack." That stack currently leans on Amazon Web Services, Google, Nvidia, and AMD.

The idea is to develop chips and models in tandem, allowing Claude to operate more quickly and at lower cost as customer usage keeps climbing. Anthropic is hiring engineers who work across both hardware and software, and the posted salary range runs from $320,000 to $485,000. The listing asks for people who have personally shipped finished semiconductor designs and are comfortable making high-stakes calls without a large organization behind them. Anthropic has not named a manufacturing partner, has not given a timeline, and has not said whether it plans to build the chips itself. Reporting from The Information earlier this summer described Samsung as a potential manufacturing partner, but those talks appear to be exploratory, not settled.

Claude and models like it are now baked into the software you already use, from writing assistants to image tools to the AI features creeping into editing apps. The cost of running those models is falling fast, and custom silicon is a big reason why. Companies design chips tuned to the exact math their models run most often, which squeezes out efficiency that general-purpose hardware can't match. Apple followed this path with its M-series processors, and Google did the same with its Tensor Processing Units. When the per-query cost drops, the AI features get cheaper to offer.

Anthropic is late to a race everyone else already entered. Google has its TPUs, Amazon has Trainium, and OpenAI unveiled a custom inference chip built with Broadcom in June. Meta is reportedly moving its next-generation accelerator toward production, and even smaller players have said they are considering their own silicon. The pull here is less about escaping Nvidia and more about arithmetic. At the scale these companies now operate, designing a chip that shaves cost off every single query pays for itself, even though industry estimates put the design cost of an advanced AI chip near half a billion dollars. Anthropic did not wait for its own team to get custom hardware either. It expanded a deal with Google and Broadcom in April for roughly 3.5 gigawatts of next-generation TPU capacity coming online in 2027, on top of capacity already arriving under an earlier Google Cloud agreement.

The catch worth keeping in mind is how early this is. Chip programs take years, and Anthropic is at the hiring stage, not the shipping stage. A single design flaw can cost months and millions to fix. The company hired Clive Chan, who previously worked on OpenAI's chip effort, and Chan wrote on X that he joined because he was "deeply impressed with the team's talent, values, and ambition." Whether that talent produces working silicon, and how much it changes the price of the AI tools you use, is a question for 2027 and beyond, not today.