Anthropic Hired the Man Who Built Google's AI Chips, and It Is Not Subtle About Why
Amir Salek shipped seven generations of Google's TPU. He now joins Anthropic's compute team, while Broadcom lines up more than 60 billion dollars of debt for chip capacity.
Anthropic has hired Amir Salek, the engineer who founded Google’s Tensor Processing Unit programme and delivered its first seven generations of chips before leaving in 2022. Bloomberg reported the move on 21 August. Salek joins Anthropic’s compute team, reporting to James Bradbury, and comes from Cerberus Capital Management, where he was a senior managing director.
You do not hire the founder of a chip programme to negotiate better prices from suppliers. Anthropic currently runs on silicon from three different vendors: Nvidia GPUs, Google TPUs and Amazon’s Trainium. It has already committed roughly 250 million dollars to Fractile, an Oxford spinout building inference chips that will not be production-ready until 2027, and it has been posting hardware job listings. The Salek hire is the clearest signal yet that the company wants chips of its own rather than a better deal on someone else’s.
The money side moved the same week. Broadcom is reportedly in talks to raise more than 60 billion dollars through a special-purpose vehicle that would lease custom AI chips to Anthropic and others, with the total package potentially reaching 100 billion dollars once a junior tranche is included. Apollo and Blackstone are said to be involved. That extends a pact from June that already committed 35 billion dollars towards Anthropic compute, aiming at 20 gigawatts of capacity by 2028.
Two forces are pushing every large lab in this direction. The first is cost: inference, the work a model does every time you send it a prompt, is now the dominant expense, and general-purpose GPUs do a lot of things you do not need if you only ever run your own models. Custom silicon trades flexibility for efficiency. The second is supply. Nvidia allocates its best chips across many customers, and being one of them means your roadmap depends on someone else’s. OpenAI reached the same conclusion and has a Broadcom-designed chip called Jalapeno. Google reached it a decade ago, which is why Salek’s CV is worth this much. A fair caveat: custom chips take years and enormous sums, and the industry has a long list of accelerator projects that never shipped. Nothing here arrives before 2027.
What this means for you: in the short term, nothing you will notice. Longer term, this is one of the clearest signals of where model prices are heading. Labs only build their own hardware when inference volume is large enough and stable enough to justify the up-front cost, which tells you they expect usage to keep climbing rather than plateau. If it works, it is the main lever for making capable models cheaper per request, which eventually reaches you as lower API prices or more generous free tiers. If you follow the industry as an investor or a builder, the more immediate story is that AI infrastructure is now being financed with the kind of debt structures normally reserved for power plants and telecom networks.
Sources
A Third of Everything Uploaded to Apple Music Is Fully AI-Made. Soon You Will Be Able to See Which
Apple told label partners that a Made With AI badge is coming later this year, and that disclosure tags are no longer optional.
