Anthropic hires the father of Google's TPU to lead its own chips

Anthropic's hardware push has a face on it now — and it's the person who built Google's answer to Nvidia.

Bloomberg reports that Amir Salek, the executive who created and led Google's TPU program across its first seven generations during nine years there, has joined Anthropic to work on custom silicon under compute chief James Bradbury. Salek led Nvidia's system-on-a-chip design team before Google, then spent time at Cerberus Capital Management after leaving in 2022 — chip architecture plus capital-markets experience, which is roughly the profile you want when a frontier lab decides to pour billions into its own silicon. He lands at a moment when the effort is already real: Anthropic stood up an internal Custom Silicon group this month, and in June it hired Clive Chan, an early member of OpenAI's secretive "Jalapéno" chip project.

The money explains the urgency. Anthropic reportedly expects to spend about $19 billion on compute this year while running gross margins near 40% against a 77% target it has told investors to expect ahead of a planned Nasdaq listing. Closing that gap means cutting the Nvidia toll — and owning the model-hardware boundary is how Google did exactly that with TPUs. Anthropic is hedging while it builds: it still buys Nvidia and cloud-partner silicon, made an initial $250 million commitment to UK inference-chip startup Fractile, and has quietly lined up data-center and power deals with partners including Riot Platforms and Volta Infra.

We flagged the direction earlier this month — Silicon rush: Anthropic goes in-house, AMD buys Taalas — but hiring the TPU's creator converts a strategy memo into a headcount. The open question is timeline: custom silicon programs run three to five years, and Anthropic needs margin relief before its IPO math stops working. Salek's first public milestone will tell us whether this is a serious program or an insurance policy.

What to watch: whether Anthropic discloses tape-out or deployment targets for its first internal chip, and whether OpenAI's Jalapéno team answers with a hiring wave of its own.

Every major lab is now trying to become Nvidia's biggest customer and its first real competitor at the same time — which side breaks first? Tell us in the comments.

Sources: Bloomberg · AI Era (Xin Zhi Yuan) · Ars Technica · Tech Times