Amir Salek spent four years doing something that looked a lot like walking away from the chip business.
After founding and running Google’s Tensor Processing Unit program from 2013 to 2022 and shipping seven generations of the custom silicon that now powers much of Google’s AI infrastructure, he left to invest in deep-tech startups at Cerberus Capital Management, a press release announced at the time.
That detour is why his next move matters more than a routine hire. Anthropic has brought Salek onto its compute team, where he will report to compute lead James Bradbury, according to Bloomberg.
He is not being poached out of a rival’s office. He is being pulled back into chip design after four years on the sidelines by a company that did not have a hardware team a year ago.
Salek is not Anthropic’s first hardware hire this year
Two months earlier, Anthropic landed Clive Chan, the second engineer ever hired to OpenAI’s custom chip program, who had spent more than two years building the Broadcom-designed inference accelerator OpenAI markets as Jalapeno.
Related: The U.S. just turned the Al race into a loyalty test
Chan announced the move himself in June, framing it as a step up rather than a rescue.
The pattern is what makes the Salek hire read as strategy instead of opportunism. Anthropic publicly confirmed on Aug. 5 that it is standing up an in-house silicon team, with the explicit goal of co-designing chips and Claude models together to cut inference costs by roughly half, according to TechCrunch.
TechCrunch also reported that Anthropic has held early manufacturing talks with Samsung.
Two hires in three months, both pulled from the two companies furthest along in custom AI silicon, is not a coincidence. It is a shopping list.
Raiding Google’s bench
Here is the part most coverage of the Salek hire has skipped past. Google is not a bystander in Anthropic’s chip ambitions. It is the supplier.
Anthropic agreed in October to buy up to a million of Google’s Tensor Processing Units in a deal worth tens of billions of dollars, then expanded that arrangement in April to cover multiple additional gigawatts of TPU capacity through 2027, according to the company’s announcement.
Google is, in other words, simultaneously arming Anthropic with chips after losing the person most responsible for the chip program that those chips came from. That is an unusual position for a supplier to hold with a customer that competes directly against its own Gemini AI models. It complicates the idea that the relationship is simply commercial.
Anthropic still says AWS Trainium, Google TPUs and Nvidia GPUs will remain central to how it scales Claude, and nothing about Salek’s hire changes that in the short run.
For investors in Nvidia, Alphabet and Broadcom, the practical impact today is limited, since custom silicon takes years to design and manufacture at volume.
But each AI lab that builds its own silicon program shrinks the list of frontier customers still fully dependent on Nvidia.
Broadcom, which already designs custom chips for OpenAI and Google, is the most likely beneficiary if Anthropic’s effort advances that far.
The real contest has moved from GPUs to headcount
Bidding wars over AI researchers are old news by now. What’s newer is the fight over hardware architects, the engineers who can take a chip from blueprint to production line. There are far fewer of them than machine learning researchers, and that scarcity is starting to show.
That scarcity has already produced legal fights in other industries. Warner Bros. Discovery sued Amazon in July over an executive who left a fixed-term contract 16 months early, arguing Amazon ran what its complaint called a lawless hiring spree.
Salek’s move to anthropic carries none of that risk. He had already left Google’s payroll years before Anthropic called, which means there is no contract to breach and no lawsuit to file.
That is the quieter lesson here. The clean way to build a chip team is not to raid a competitor mid-contract. It is to identify who already built one of these programs once, wait until they are between roles, and make the offer before anyone else does.
More AI:
- Nvidia just made a move Wall Street wasn’t ready for
- Microsoft just took sides in AI policy fight
- OpenAI just disclosed something genuinely alarming
The custom silicon race
Meta’s in-house Iris chip is due in production by September. OpenAi’s Jalapeno targets the back half of this year. Microsoft has run Maia chips for two years, and Amazon has done the same with Trainium. Anthropic is the last major lab into custom silicon, not the first.
Owning hardware talent is not the same as owning working silicon. Anthropic’s chip, whatever it becomes, is not expected before 2028 at the earliest.
The next test is not who Anthropic hires next. It is whether a lab built on renting other companies’ chips can actually ship one of its own.
As frontier AI labs and hyperscalers double down on in-house silicon, a critical question emerges for Wall Street: Does this mark the beginning of structural market share erosion for Nvidia?
While Nvidia’s total addressable market continues to expand rapidly, every successful custom ASIC deployed for inference or specialized training shifts workloads away from general purpose GPUs, slowly chipping away at the revenue concentration of the market’s leading chipmaker.
Related: Anthropic-powered AI model sends shocking message to employee
