Analysis
Anthropic has hired Amir Salek, the executive who founded and ran Google's custom TPU chip program until 2022, as the AI lab lays the groundwork for building its own semiconductors, Bloomberg reported. Salek delivered the first seven generations of Google's TPU chips during his tenure and will report to James Bradbury on Anthropic's compute team.
The hire is a significant signal because Salek's specific expertise -- taking a custom AI accelerator from architecture through seven successive shipping generations -- is among the rarest skill sets in the industry, held by a small number of people who have actually run a hyperscale custom-silicon program at Google, Amazon or Microsoft. Anthropic recruiting the person who built that exact program at Google, rather than a chip-industry generalist, suggests real intent to build a multi-generation program rather than a single one-off chip.
The move follows a familiar competitive pattern: every major AI lab now has some form of custom-silicon ambition. Google has run TPUs internally for a decade, Amazon has Trainium and Inferentia, Microsoft has Maia, and Meta has its own MTIA chips -- all built to reduce dependence on Nvidia's GPUs and to tailor chip architecture more precisely to each company's own model workloads. OpenAI has separately pursued custom-chip partnerships. Anthropic hiring Salek is the clearest evidence yet that it intends to build its own program rather than remain purely a GPU and cloud-compute customer indefinitely.
“The move follows a familiar competitive pattern: every major AI lab now has some form of custom-silicon ambition.”
- Amir Salek -- founded and led Google's TPU program through seven generations, now joining Anthropic's compute team
- Google, Amazon, Microsoft, Meta -- all run internal custom-AI-silicon programs Anthropic's effort would need to eventually compete with or complement
- Fractile -- the UK chip startup Anthropic separately committed roughly $250 million to for an initial chip order, showing a parallel build-and-buy compute strategy
The practical timeline for any custom-silicon effort is long -- typically three to five years from a serious hiring push to a shipping chip at meaningful scale, based on how long it took Google, Amazon and Microsoft's own programs to mature. Anthropic signing this hire now, while also raising capital toward a mega-IPO and scaling revenue toward a reported $100-120 billion annualized run rate by year-end, suggests the company is trying to build compute independence in parallel with its public-market ambitions rather than sequentially after them.
The counterweight is that a single senior hire, however credentialed, does not guarantee a successful chip program -- custom silicon requires sustained capital, a dedicated engineering organization, and foundry relationships that take years to build, and several well-funded custom-chip efforts across the industry have quietly stalled or been scaled back after initial announcements generated more attention than the eventual product.
What to watch is whether Anthropic discloses a specific chip architecture, foundry partner, or timeline in the coming quarters, which would signal the program has moved from a hiring announcement to an actual product roadmap.
