Amir Salek spent eight years at Nvidia designing system-on-a-chip products before Google recruited him in 2013 to start its custom silicon effort from scratch. He ran the program for nearly a decade, delivering the first seven generations of Tensor Processing Units, the chips that still anchor Google’s AI infrastructure. Last week, Anthropic persuaded him to do it again, this time in its own building.
Anthropic has hired Salek to join its compute team and lead the design of its own AI chips, according to people familiar with the matter, in the latest sign that the race among AI labs has moved from model quality to hardware. He will report to James Bradbury, Anthropic’s head of compute. Bloomberg first reported the hire on Aug. 22.
Salek left Google in 2022 and spent the intervening years as a senior managing director at Cerberus Capital Management, investing in semiconductor and AI companies. His return to chip engineering gives Anthropic one of the few people in the industry who have built a custom silicon program from zero to scale inside a company that depends on it.
Anthropic began assembling a chip design team last month, according to people familiar with the matter, hiring engineers and architects for roles that did not exist at the company a year ago. The effort is early, and the first Anthropic-designed chip would take years to tape out and reach a data center. But the direction is unambiguous: the lab behind the Claude models wants to own more of its computing stack.
The motivation is cost. Anthropic rents most of its computing power from cloud providers, and its compute bill runs into the billions of dollars a year — estimates from people familiar with the company’s finances put it near $19 billion annually. Nvidia’s accelerators, which power the bulk of the industry’s AI training and inference, carry premium pricing and long lead times. Every model Anthropic ships depends on hardware it does not control.
The hire is one piece of a broader strategy to change that. In April, Anthropic signed a deal with Amazon.com Inc. pledging more than $100 billion over a decade for cloud services, including access to Amazon’s custom Trainium and Inferentia chips. It has committed about $250 million to buy chips from Fractile, a British startup whose products do not yet exist, and signed capacity agreements with infrastructure firms Riot Platforms and Volta Infra Holdings. Google, an investor, supplies TPUs directly, and Anthropic itself rents TPU capacity at scale — capacity that Salek once helped build.
Salek’s arrival does not mean Anthropic is abandoning those suppliers. The company describes its approach as a multi-chip strategy: it will keep buying from Nvidia, AMD and the cloud providers while developing silicon tailored to its own models. The economics of AI favor vertical integration — a chip designed around a specific model can deliver large gains in inference cost per token — and the industry’s largest labs have all reached the same conclusion.
OpenAI, Anthropic’s chief rival, unveiled its own inference chip, code-named Jalapeño, co-developed with Broadcom Inc., and plans to deploy it in its own data centers. Amazon builds Trainium. Google builds TPUs. Microsoft works with custom silicon partners. Anthropic was one of the few frontier labs still renting everything it runs; the Salek hire closes that gap in personnel if not yet in hardware.
The talent itself is the scarce resource. Custom chip programs are led by a small group of people who have done it before, and most of them are inside Google, Nvidia or the major cloud companies. Anthropic’s courtship of Salek, who could have stayed at a fund or joined a startup, is evidence that its pitch — equity in a company heading toward a public listing, research control, the chance to build from zero — is landing. The company has been reported to be weighing an IPO that could value it at up to $2 trillion, which would make its equity the most valuable currency in the AI talent market.
The risks are just as real. Designing a leading-edge chip costs on the order of $500 million and takes years, and the failure rate for new silicon programs is high. Anthropic’s chips, if they ship at all, will arrive long after the current generation of models is built and sold, meaning the payoff depends on a compute demand that is itself uncertain. Analysts said the hire should be read as an insurance policy as much as a build plan.
For the rest of the industry, the signal is simpler: the AI arms race has moved into silicon. Whoever controls the chip controls the unit economics of inference, and the biggest model makers are no longer willing to leave that to their suppliers. Anthropic’s move, coming weeks after OpenAI’s Jalapeño announcement and months of Amazon and Google infrastructure deals, means the two most prominent AI labs are now competing on hardware roadmaps that will not bear fruit for years.
Salek will not publicly describe his plans before he starts, according to people familiar with the matter. But the assignment is clear. Anthropic wants the next generation of chips — whoever designs them — to be shaped by its own needs, and it has hired the person who built the template for doing exactly that at Google.
