Anthropic Hires Google TPU Founder to Build Its Own AI Chips

Anthropic announced on August 22, 2026 that it has hired Amir Salek, one of the co-founders of Google's custom chip program, to join its compute team. Salek ran Google's tensor processing unit (TPU) business until 2022, delivering the first seven generations of the chip. The hire signals that Anthropic is laying the groundwork for in-house semiconductor development, moving beyond its current dependence on chips sourced from Nvidia, Google, and Amazon.

What happened

Detail	Fact
New hire	Amir Salek
Role	Compute team, reporting to James Bradbury
Previous employer	Google (TPU program co-founder, until 2022)
TPU generations delivered	7
Other past employers	Nvidia; Cerberus Capital Management (senior MD)
Current chip suppliers	Nvidia, Google, Amazon

Anthropic confirmed it has brought Amir Salek onto its compute team, where he will report to James Bradbury. Salek is one of the people who built Google’s tensor processing unit program from the ground up. A TPU is a chip designed specifically to run machine-learning workloads, as opposed to the general-purpose graphics processors (GPUs) that most AI labs, including Anthropic, currently buy from Nvidia.

After leaving Google in 2022, Salek served as a senior managing director at Cerberus Capital Management, a private equity firm co-founded by a former US Deputy Secretary of Defense. He also spent time at Nvidia before his Google tenure.

Why is Anthropic building its own chips?

Right now, Anthropic sources compute from three external suppliers: Nvidia, Google, and Amazon. That dependence creates real risk. Supply shortages at any one of those vendors can slow model training and inference capacity. Custom silicon, built around Anthropic’s own workloads, could reduce that exposure and potentially lower cost per computation over time.

Anthropic has already started posting job listings related to an internal silicon effort, according to the source report. Salek’s hire is the most senior public signal yet that the program is moving from planning to execution.

Every major AI lab has been making similar moves. Designing your own chips means you can optimize memory bandwidth, precision formats, and interconnect speeds for the specific operations your models run most often, rather than accepting whatever a general-purpose chip offers. For a company spending heavily on GPU cloud costs, the long-term economics can be compelling.

Our take

This is a credible hire for a credible ambition. Salek did not just work on TPUs; he co-founded the program and shipped seven generations of production silicon. That is the kind of operational depth most chip startups spend a decade trying to recruit.

The harder question is timeline and cost. Custom silicon programs routinely take three to five years from first tape-out to production at scale, and they consume enormous engineering resources before a single chip ships. Anthropic is competing with Google, which has a decade-long head start on TPU design, and with Amazon, which has been shipping its own Trainium chips since 2021.

Still, the direction is clear. If you are an enterprise customer deciding between AI vendors for the next few years, Anthropic’s infrastructure independence matters. A lab that controls its own compute stack is less likely to pass supplier cost increases on to customers or face capacity crunches that delay your workloads. If you are exploring AI integration for your business, the long-term supply stability of your chosen model provider is worth watching alongside benchmark scores.

For more coverage of how the major AI labs are positioning their hardware and model strategies, see our AI news coverage.

Source: Bing News · Anthropic