Anthropic Hires Google's Former Custom-Chip Chief to Build Its Own AI Silicon

Amir Salek ran Google's custom AI-chip program for years and personally delivered the first seven generations of its Tensor Processing Units, the chips that trained many of the largest models in the field.

The hire, first reported by Bloomberg, is the latest public marker of a shift across the frontier AI labs: they are no longer content to rent their compute from Google and Amazon. They want to design it.

Anthropic currently buys chips from all three. It has also signed a UK-based agreement with chip-design firm Fractile for an initial order worth roughly $250 million, with plans to expand, and has struck capacity deals with Riot Platforms and Volta Infra Holdings, per Bloomberg. Those deals cover near-term supply. The Salek hire is for the longer term.

Salek's mandate is to start an in-house silicon business. He joins Anthropic's compute team and will report to James Bradbury, one of the company's co-founders, and Anthropic has begun recruiting for the group, according to Business Insider's reporting on its in-house chip effort. The company is in laying-the-groundwork mode, not shipping mode. Google took years to get TPUs to a scale where they could carry meaningful training and inference loads; Anthropic is starting from scratch, but it is starting with one of the few people in the industry who has done the equivalent at two other employers.

Frontier-model training runs now cost billions of dollars in compute. The suppliers who can deliver that scale, including Amazon's Trainium and Inferentia lines, also set the price. A lab that designs its own silicon ties its model roadmap to its own chip roadmap rather than to a vendor's generation cycle, and it can shape the chip's data paths and memory hierarchy to its specific model architecture.

If the in-house program works, the upside is concrete: a lab can price inference at cost rather than at a vendor's markup, ship new model generations on its own clock, and tune the chip's matrix-multiplication units and on-chip memory to the specific shape of its own architectures. None of that is guaranteed. Custom silicon programs routinely run years behind schedule, and Anthropic's effort is starting without a public design, a named foundry partner, or a timeline.

OpenAI has effectively made the same bet, with a different approach. The company unveiled a chip codenamed Jalapeno, co-developed with Broadcom, with deployment planned for later this year, according to Bloomberg. One lab is partnering with a foundry; the other is hiring a veteran to start from scratch. Either way, the frontier labs are moving from renting compute to designing it.

The next public data point will be a working chip in a data center.