Anthropic Hires Builder of Google’s First Seven TPUs for In-House Silicon Push

Anthropic has hired Amir Salek, who founded Google’s custom-chip program and delivered its first seven TPU generations, for the compute team building an in-house silicon operation. The move adds a veteran of large-scale accelerator development as Anthropic continues to run Claude on hardware from multiple suppliers.

Salek’s background spans two of the companies central to AI hardware. Before Google, he founded and led Nvidia’s system-on-a-chip design organization. At Google, he established its custom-chip program and delivered the first seven generations of Tensor Processing Units, or TPUs, before leaving in 2022.

Anthropic had already confirmed in August that it was recruiting a custom-silicon team, including engineers experienced in taking semiconductor designs through production. Salek joins that effort to help build the company’s internal silicon operation.

Building alongside suppliers

The internal effort sits beside Anthropic’s existing hardware strategy, rather than replacing it. Anthropic currently trains and runs Claude across Nvidia GPUs, Google TPUs and Amazon Trainium chips. Amazon is its primary cloud and training partner, while Anthropic describes the mix as a way to put workloads on different hardware and reduce reliance on one platform.

Experience at the hardware layer

Anthropic has already worked closely with one specialized accelerator platform. Its engineers wrote low-level software kernels for Trainium, contributed to Amazon Web Services’ Neuron software stack, and worked with Amazon’s Annapurna Labs on future chip generations.

• Salek brings experience founding a custom-chip organization and delivering multiple TPU generations.
• Anthropic plans to develop internal silicon while continuing to reserve large amounts of Nvidia, Google and Amazon compute capacity.

The hire is an organizational addition to that mixed approach: Anthropic is assembling an in-house silicon operation while its disclosed commitments keep external accelerators central to its compute supply.