Infrastructure

Anthropic Poaches Google's TPU Architect to Build Its Own Chips

Anthropic hired Amir Salek, the ex-head of Google's TPU program, to build in-house AI chips - a hire EU and UK enterprises with multi-year compute deals should watch closely.

InfrastructureBy Servola Tech Desk2026-08-224 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Anthropic Poaches Google's TPU Architect to Build Its Own Chips

Anthropic Hires the Architect of Google's TPU Program

Anthropic has hired Amir Salek, the founder and former head of Google's Tensor Processing Unit program, to join its compute team, Bloomberg reported on August 21, 2026, citing people familiar with the matter. Salek led seven generations of TPU chip development at Google, the custom silicon that has powered Google's own AI training and inference workloads since 2015 and now underpins a growing share of Google Cloud's external AI business. Business Standard and Seeking Alpha corroborated the Bloomberg report the same day, adding that Salek most recently worked at Cerberus Capital, a private equity firm, before Anthropic recruited him.

At Anthropic, Salek reports to James Bradbury on the company's compute team, according to the reporting. Bradbury has been a central figure in Anthropic's infrastructure strategy, and placing Salek under his reporting line signals the hire is meant to feed directly into Anthropic's hardware roadmap rather than sit in a research or advisory capacity. Neither Anthropic nor Google has issued its own statement on the move beyond what Bloomberg's sources described.

Why Seven TPU Generations Is the Detail That Matters

Salek does not bring generic chip-industry experience to Anthropic - he built the exact program that freed Google's AI infrastructure from full dependence on Nvidia GPUs. Leading seven generations of TPU development means Salek carried Google's custom silicon from an internal experiment into the chip family now training Gemini models and renting out to external Google Cloud customers, a decade-long arc that very few engineers anywhere have run start to finish. That is a materially different hire than recruiting a chip-industry generalist or a fabrication-process specialist.

For Anthropic, the value is not just Salek's engineering judgment but his firsthand knowledge of what it actually took Google to get from a first-generation TPU to a production fleet serving billions of inference requests: the design trade-offs, the fabrication partnerships, and the software stack changes needed to make custom silicon usable at scale. That is precisely the playbook Anthropic needs if its own in-house chip effort is going to move faster than a from-scratch attempt would.

This Confirms a Shift From Renting Compute to Owning It

Anthropic recruiting the person who built Google's chip program marks a concrete step in a shift from renting AI compute to owning the silicon roadmap outright. Servola reported on August 7, 2026, that Anthropic had confirmed it was assembling an in-house chip team; that story described an ambition in general terms, with no named hire attached. The Salek hire is the first specific, dated data point that turns that confirmed ambition into an active program with a named leader and a reporting line.

Anthropic's compute today runs on a mix of Google TPUs, Amazon Web Services Trainium chips, and Nvidia GPUs, secured through multi-year commercial deals with each of those providers. None of the current reporting sets a timeline for Anthropic reducing its reliance on any of those three, and Salek's hire alone does not change existing contracts. What it does change is the credibility of Anthropic's long-term intention to design its own chips rather than remain a permanent customer of Google, Amazon, or Nvidia silicon.

The Procurement Question This Raises for EU and UK Buyers

Enterprises signing multi-year Anthropic API or infrastructure commitments now have a concrete reason to ask what happens to pricing and availability once Anthropic controls more of its own chip supply. Most EU and UK organizations buying Anthropic capacity today do so through Google Cloud, AWS, or Microsoft Azure marketplace agreements layered with the same multi-year terms that make switching providers costly - exactly the terms procurement teams should be re-reading now, not when a chip transition is already underway. A lab that designs its own silicon has a different cost structure and a different incentive to keep favorable terms with today's cloud partners than one that permanently rents capacity from them.

None of this means current Anthropic pricing or availability will change in the next year; nothing in the Bloomberg, Business Standard, or Seeking Alpha reporting sets that timeline. It does mean that renewal terms, exit clauses, and multi-sourcing strategy deserve a fresh look before Anthropic's silicon roadmap - not just its model roadmap - starts showing up in contract negotiations two to three years from now.

Frequently asked questions

Who is Amir Salek?

Amir Salek is the founder and former head of Google's Tensor Processing Unit program, where he led seven generations of custom AI chip development before most recently working at Cerberus Capital and then joining Anthropic's compute team.

Is this the same story as Anthropic's August 7, 2026 in-house chip announcement?

No. That earlier report confirmed Anthropic was building an in-house chip team in general terms. The Salek hire, reported by Bloomberg on August 21, 2026, is the first specific, named hire that puts a person and a reporting line behind that plan.

Does this mean Anthropic will stop using Google TPUs, AWS, or Nvidia chips?

Not based on current reporting. Anthropic's existing multi-year compute deals with Google, Amazon, and Nvidia remain in place, and no source has set a timeline for phasing any of them out. The hire signals long-term intent, not an immediate supply change.

What should EU and UK enterprises with Anthropic contracts do now?

Review renewal terms, exit clauses, and multi-sourcing strategy now, since pricing and availability could shift once Anthropic's own silicon roadmap, rather than just its model roadmap, starts factoring into contract negotiations two to three years out.

Anthropic just hired the engineer who proved a frontier AI lab can own its chip roadmap end to end, not rent someone else's. Enterprises whose multi-year Anthropic contracts assume today's vendor arrangement holds through 2028 are the ones with the most reason to reread those contracts now.

AnthropicAI ChipsTPUCompute InfrastructureVendor Risk

GIC and Macquarie Become Anthropic's Landlords

Anthropic just became a tenant, not an owner. Macquarie Asset Management and Singapore's GIC will own and fund Theseus Infrastructure - what that means for Claude's EU and UK enterprise customers.

3 min read

Artificial Intelligence

DeepSeek Adds Vision to V4-Flash, Nears Opus 4.8

DeepSeek released DeepSeek-V4-Flash-Vision-Exp on August 21, 2026, an experimental multimodal model scoring 64.3 on Chartography and nearing Opus 4.8 on visual-agent tasks, raising new vendor-risk questions for European buyers.

4 min read

Decisions

The Third Pivot Was the Warning Magic Leap Missed

Magic Leap filed a WARN notice for 193 layoffs effective 1 October 2026, its fourth strategic pivot after $4 billion raised. The lesson is the pattern, not the pivot.