Anthropic Hires Google's TPU Founder to Build Its Own Chips

Anthropic has recruited Amir Salek, the executive who built Google’s custom-chip program and ran its Tensor Processing Unit business, to join its compute team. Salek will report to James Bradbury, who leads compute at Anthropic, according to RuntimeWire, which cited a Bloomberg report published August 21 in which Anthropic confirmed the hire. Dina Bass first surfaced the move in a post on X.

The hire matters because compute has become the primary cost line for frontier AI labs, and Anthropic still depends on three external chip suppliers even as it starts building an internal option. Bringing in the person who built Google’s alternative to Nvidia is a statement about where Anthropic wants its cost structure to go, long before any chip of its own reaches a data center.

Salek’s record is specific to this problem. At Google, he stayed on the TPU program long enough to carry it through seven separate chip generations, only exiting in 2022, per Bloomberg’s account. Before that stretch, he had already stood up and led the team inside Nvidia responsible for system-on-a-chip design. An engineering doctorate from USC, blending computer science with hardware, sits underneath both jobs but is the least relevant line on the resume for what Anthropic needs now.

After Google, Salek spent four years on the investing side. Cerberus Capital Management brought him in during March 2022 at senior managing director level, and he also took a partner role at Tracker Ventures, a Cerberus platform backing semiconductors, AI, edge computing and aerospace.

Anthropic said earlier this month that it was staffing a custom-silicon team with engineers experienced in taking chip designs into production. Bringing Salek on board converts that recruiting effort from an intention into a functioning group with someone who has already run one. At Google, he was recruited to stand up a chip capability from nothing and then scale it across the company’s data centers and cloud business. Anthropic is asking him to do something similar for workloads built around Claude.

One hire does not produce a shipping accelerator. Competitive AI chips require architecture, verification, software, networking, packaging and manufacturing, work that took Google years to move from a founding decision to a production TPU. Salek’s value is that he ran that cycle seven times, not once: the relevant experience is repeatability, not a single design win.

Anthropic already sits closer to silicon than a typical software buyer. In 2024, the company disclosed that its own staff had begun writing hardware-level code for Amazon’s Trainium accelerators, on top of contributions to the AWS Neuron software stack and joint work with Amazon’s Annapurna Labs on future chip generations, by Anthropic’s own account.

That proximity has not slowed Anthropic’s external chip orders. In October 2025, Anthropic said it planned to deploy as many as one million Google TPUs, with more than a gigawatt of capacity expected online in 2026. In April, it added a separate agreement with Google and Broadcom covering multiple gigawatts of next-generation TPU capacity beginning in 2027. Anthropic trains and runs Claude across Nvidia GPUs, Google TPUs and Amazon Trainium chips today, and Amazon remains its primary cloud and training partner, per Anthropic’s own materials.

The overlap is the point, not a contradiction. A lab ordering silicon by the gigawatt from three merchant vendors, then hiring the person who built the fourth option, is hedging its supply, not abandoning its suppliers. A credible internal design gives Anthropic leverage in future capacity and pricing conversations with Nvidia, Google and Amazon, whether or not a Claude accelerator ever ships. Google’s own TPU program needed several years to go from an internal tool to a chip business Google now sells to outside customers. If Anthropic’s effort follows a similar arc, an Anthropic-designed chip is unlikely to touch a production training run before the end of the decade, and near-term progress will show up in headcount and infrastructure hires rather than in Claude’s training stack.

For competitors and cloud partners, the immediate signal is not a chip roadmap. It is a negotiating posture: Anthropic is positioning itself to argue for better GPU and TPU terms over the next several procurement cycles, regardless of when, or whether, its own silicon reaches production.

RuntimeWire (Ryan Merket) reported the hire on August 24, 2026, citing an earlier Bloomberg report published August 21.