Anthropic Hires Amir Salek, the Engineer Behind Google's TPU, as Its Custom Chip Effort Gets Real

Anthropic has reportedly brought Amir Salek, the engineer who founded Google's Tensor Processing Unit (TPU) program, onto its compute team. The move lands just weeks after word emerged that the company was staffing an in-house chip design group, meaning someone who has actually shipped seven generations of TPUs is now inside that effort. Anthropic's answer to the question of where its compute comes from is starting to look like silicon it designs itself.

※1 USD = 159 JPY (as of August 21, 2026)

The Person Who Shipped Seven Generations of TPUs

Salek's background is unusually direct for this kind of hire. He spent roughly 8 years at NVIDIA as a senior director of engineering, where he founded the System-on-a-Chip design organization behind the company's GPU and Tegra lines. In 2013 he moved to Google to establish a custom silicon group for its data centers.

Over 9 years at Google, that group delivered TPU generations 1 through 7 plus the Edge TPU. The foundation of the TPU fleet that now carries Google's training and inference workloads was laid during that stretch. From March 2022 he served as a senior managing director at the investment firm Cerberus Capital Management, covering semiconductors and edge computing.

In other words, Salek is not a researcher who writes papers about accelerator architecture. He is a practitioner who has run production silicon programs through multiple generations, at infrastructure scale comparable to what Anthropic faces today.

Why "Relying on Others Isn't Enough"

Anthropic currently sources chips from several suppliers, including NVIDIA, Google and Amazon. Even so, when its plan to build an in-house chip design team surfaced in early August, the company framed existing partnerships as insufficient on their own.

The reasoning is straightforward. As demand for Claude grows, three things press down harder: allocation queues, lead times, and pricing the company does not set. Adding suppliers does not change the fact that all three remain variables controlled by someone else. Designing its own silicon pulls those variables back onto Anthropic's side of the table.

The company has also reportedly placed an initial order of roughly 250 million USD (about 40 billion yen) with the UK chip startup Fractile, with an expressed intent to expand the contract later. Buy off the shelf, squeeze more out of what exists, and build your own: Anthropic is running all 3 tracks at once.

The Third Frontier Lab to Go Custom

Anthropic is not the first major AI company to head down this road. Google has been shipping TPUs in volume for more than 9 years, and OpenAI is developing its own inference-focused chip with Broadcom. This hire signals that Anthropic is joining that group in earnest.

What stands out is that Anthropic went after the founder of the TPU program specifically, rather than a generic semiconductor veteran. Standing up a production silicon organization from zero has a fairly well-known set of failure modes. Bringing in someone who has already worked through them once suggests an intent to compress the trial-and-error phase.

Nothing Changes This Week

There are good reasons not to get ahead of the story. Anthropic has issued no official announcement about the hire. This is reporting, and the exact title and start date remain unstated.

On the technical side, nothing is known about the architecture, whether the target is training or inference, which foundry would build it, or when it might arrive. Chip development does not end at design: fabrication, packaging and software bring-up all follow, and none of those stages compress the way software schedules do. A first-generation part reaching production racks takes years.

Custom silicon programs also get quietly shelved with some regularity. A high-profile hire is a strong statement of intent, not a guarantee of a shipping product, and it says nothing about near-term API pricing.

Summary

Amir Salek, who delivered Google's TPU from its first generation through its seventh, has reportedly joined Anthropic's compute team. The in-house chip design effort that surfaced in early August now has a leader who has completed that exact job before. There is no official confirmation, and no chip specifications or timeline exist yet, but Anthropic's shift toward not leaving its compute supply entirely in other companies' hands is now considerably clearer.