Anthropic has hired former Google executive Amir Salek to join its compute team as the AI lab lays the groundwork for an in-house semiconductor initiative, according to a report from Bloomberg.
Salek, who previously led Alphabet's custom silicon efforts and delivered seven generations of Tensor Processing Units, will report directly to Anthropic compute lead James Bradbury. The recruitment is the strongest signal yet that the company's move toward proprietary AI hardware — first signaled by job listings and team-building reports earlier in August — is now accelerating into its execution phase.
Who Is Amir Salek
Salek arrives with one of the most consequential résumés in the chip industry. At Google, he led the custom silicon organization responsible for the Tensor Processing Unit, the accelerator that powered more than a decade of Google's AI workloads and became the template for in-house AI chips across the industry. Seven generations of TPUs were delivered under that organization's watch.
He left Google in 2022 for private equity firm Cerberus Capital Management, where he served as senior managing director. Anthropic is now bringing that combination of chip architecture depth and large-scale operational experience to bear as it actively recruits for its emerging hardware organization.
Why Anthropic Wants Its Own Chips
The Claude maker continues to buy chips from Nvidia, Amazon, and Google, and those relationships remain central to its near-term compute strategy. But according to the Bloomberg report, establishing an internal silicon division is expected to help the company tailor hardware design to its operational requirements while mitigating ongoing supply chain bottlenecks.
The economics are straightforward. Frontier AI labs are among the largest compute buyers on the planet, and every layer of the stack that can be owned rather than rented reduces long-term costs and dependency. The strategy reflects a broader trend among leading AI labs working to decrease dependence on external hardware suppliers as hyperscalers and developers scramble for scarce data center infrastructure.
A Competitive Race to Own the Stack
Anthropic is not alone. Rival OpenAI has taken similar measures by co-developing its custom Jalapeno chip alongside Broadcom, which is scheduled for operational deployment later this year. Google builds its own TPUs and has kept them at the center of its AI infrastructure strategy for more than a decade. Amazon has its Trainium and Inferentia lines feeding its AWS AI services, and Meta has its MTIA accelerator family in production. Microsoft, meanwhile, has been developing its Maia data center chips as it reduces reliance on Nvidia for its Azure fleet.
To sustain its immediate compute demands alongside long-term hardware development, Anthropic has been rapidly securing multi-vendor infrastructure agreements across the supply chain. The company recently outlined an initial $250 million commitment to UK chip firm Fractile, while simultaneously securing additional data center capacity through deals with Riot Platforms and Volta Infra Holdings.
The pattern extends beyond the hyperscalers. Startup Etched built a transformer-specialized chip to challenge general-purpose GPUs, and a wave of well-funded challengers has raised billions on the premise that AI inference workloads demand purpose-built silicon. What was once a niche procurement decision has become a strategic arms race, with each major AI player concluding that competitive advantage increasingly lives at the hardware layer.
The hire also lands at a delicate moment for Anthropic's public market ambitions. The company is preparing a blockbuster initial public offering that could raise as much as $100 billion, and its filings will show both extraordinary revenue growth and the enormous capital intensity behind it. Investors will scrutinize the durability of its compute strategy with unusual rigor, because compute costs sit directly on the path between revenue and any future profitability. Owning more of the hardware stack — or at least demonstrating a credible path to it — is likely to feature prominently in that story. A lab that designs even part of its own silicon can tune power efficiency, memory bandwidth, and scheduling to its own models rather than accepting general-purpose designs built for the whole market. For a company whose flagship Claude models are among the most compute-intensive in production, those marginal gains compound into serious money.
What Comes Next
Building a chip from scratch typically takes years, and even a team led by a veteran of seven TPU generations will not produce custom silicon overnight. The near-term markers to watch are hiring velocity in the silicon team, potential foundry partnerships, and whether Anthropic's first-generation design targets a narrow workload — inference for Claude models, for example — rather than a general-purpose accelerator.
What is already clear is the direction. The era in which AI labs could treat compute as a commodity purchase is ending, and Anthropic has just acquired the expertise to make sure it is not left renting its future from competitors.
Stay Ahead of AI
The race to own the AI hardware stack is reshuffling alliances across the industry. AI Buzz Wire tracks every major chip, compute, and infrastructure move with source-verified reporting — no rumors, no hype.
Read more AI hardware news →
