By Hamza Faisal

Anthropic has publicly confirmed it is assembling an in-house chip design team that will co-design custom silicon to run its Claude models faster and more efficiently. The company told Business Insider this week that it is "putting together an in-house chip team," ending months of speculation that followed reports as far back as April that it was exploring its own AI processors.

The confirmation lands Anthropic in the same club as every other frontier AI lab: Google has its TPUs, Meta its MTIA chips, Microsoft its Maia accelerators, Amazon its Trainium hardware, and OpenAI co-designed the 'Jalapeño' inference chip with Broadcom, which debuted in June. No one wants to be the only lab still renting its most important asset.

A multi-chip approach, not a replacement

An Anthropic spokesperson told Data Center Dynamics that the custom-silicon effort is "the latest step in our multi-chip approach," and that the company will continue to deploy a "diversified hardware stack." That means Nvidia GPUs, Google TPUs, Amazon Trainium chips and AMD hardware all stay in the mix — today's Claude runs on all of them, and Anthropic has also locked in multi-gigawatt TPU capacity through deals with Google and Broadcom announced in April.

The logic is familiar: custom chips can be shaped around the exact operations a company's models use most, cutting the cost of every request and loosening dependence on tight Nvidia GPU supplies that are expected to persist through 2027. Industry analysts estimate designing a custom AI chip can cost roughly $500 million before a single wafer is manufactured — money only frontier labs can justify.

The team is stacked with chip veterans

Anthropic has been quietly hiring for the effort since at least August, with listings for silicon engineers and hardware system architects paying up to $485,000. One notable detail from job boards: a "Research Engineer, Chip Design RL" role — teaching AI models to design chips — was listed at $500,000 to $850,000, well above the $320,000 to $485,000 range for the silicon engineers actually designing the company's first ASIC.

The hardware push is led by Clive Chan, who joined Anthropic in June after two and a half years as the second hardware hire on OpenAI's custom chip program, and earlier work on Tesla's Dojo supercomputer. He is joined by Amir Salek, a former Google chip executive who worked on the first seven generations of Tensor Processing Units. The company declined to say how large the team is or when a chip might ship.

The race for compute independence

The move underscores how central raw compute has become to AI strategy. As Morgan Stanley noted this week, the constraint on AI growth is shifting from simply obtaining chips to securing electricity, land and money for data centers — and companies that control their own silicon get more of all three to go around.

For Anthropic, which is reportedly heading toward one of the largest tech IPOs in history, custom silicon is also a Wall Street story: it promises a path to lower per-token inference costs and a hedge against the pricing power of hardware vendors. First OpenAI, now Anthropic — the chip is becoming the company.

By Hamza Faisal