Anthropic has reportedly hired the engineer who built Google's TPU program from scratch. Bloomberg's Dina Bass reported on August 21, 2026 that Anthropic hired Amir Salek — founder and former head of Google's custom silicon effort — for its compute team, as the company lays the groundwork for its own AI chips. Salek most recently spent over four years at Cerberus Capital Management.
This is not a standalone personnel item. It is the third chapter this summer in Anthropic's compute-diversification story, after the AMD Instinct MI450 deal and the Google-backed $15B Texas data center. Here's who Salek is, what's actually confirmed versus reported, and why a nine-year TPU veteran joining Anthropic's compute team matters more than a typical hiring headline.
TL;DR — What People Are Asking
| Question | Answer |
|---|---|
| Who got hired? | Amir Salek, founder and former head of Google's TPU program (2013–2022) |
| Confirmed by whom? | Bloomberg's Dina Bass, Aug 21, 2026 — no Anthropic press release yet |
| What was he doing before this? | Senior Managing Director at Cerberus Capital Management since March 2022 |
| What did he do at Google? | Founded Custom Silicon for Google Technical Infrastructure/Cloud; built TPUv1–v4 and Edge TPU |
| What did he do before Google? | ~8 years at Nvidia as Senior Director of Engineering; founded Nvidia's SoC Design org (GPU, Tegra) |
| What team is he joining? | Anthropic's compute team, working toward custom AI chips |
| Does this replace Nvidia/AMD/Google deals? | No — it layers on top of an already multi-sourced supply chain |
| Does this change Claude pricing today? | No — custom silicon timelines run 18+ months minimum |
Who is Amir Salek?
Amir Salek's résumé is one of the most direct paths in the industry from "built the chip" to "leads the next chip." According to his Cerberus bio and corroborating reporting:
- Nvidia (~8 years) — Senior Director of Engineering. Founded and led Nvidia's System-on-a-Chip (SoC) Design organization, contributing to its GPU and Tegra product lines.
- Google (2013–2022) — Recruited to establish custom silicon development for Google's data centers. As founder and head of Custom Silicon for Google Technical Infrastructure and Google Cloud, he directed the design and deployment of TPUv1 through TPUv4 and Edge TPU — the first seven generations of Google's Tensor Processing Units. Financial Times reporting has described him as "the brains behind Google's TPU chip."
- Cerberus Capital Management (March 2022–2026) — Senior Managing Director and Partner of Tracker Ventures, investing in semiconductors, AI/ML, edge computing, and deep-tech infrastructure.
- Anthropic (reported, August 2026) — Joining the compute team as Anthropic works toward custom AI chips.
That's a rare combination for a chip hire: he isn't a researcher who published a paper on accelerator architecture — he ran a production silicon program that shipped seven chip generations inside one of the largest AI infrastructure operations on Earth. If Anthropic wants to go from "hiring a chip design team" (which TechCrunch reported on August 5, 2026) to "shipping a chip," Salek is the kind of leader who has already done that exact job once, for the closest thing in the industry to a comparable customer profile.
What's actually confirmed vs. reported
Be precise about sourcing here, the same way explainx.ai has flagged on Anthropic's reported Decart acquisition talks and the Google-backed Texas data center reporting:
| Confirmed | Reported, not yet confirmed by Anthropic |
|---|---|
| Salek's Nvidia and Google TPU career history (his own Cerberus bio, FT coverage, Wikipedia) | That Salek has formally joined Anthropic, exact title, and start date |
| Anthropic is publicly hiring for an in-house chip design team (TechCrunch, Aug 5, 2026) | Anthropic's official framing of Salek's specific mandate |
| Anthropic's public line that "relying on others clearly isn't enough" despite AWS, Google, Nvidia, and AMD partnerships | Whether this hire changes any near-term supply commitment |
As of publication, there is no Anthropic press release naming Salek — the story is Bloomberg's reporting, corroborated by wire pickups like Seeking Alpha. Bloomberg reporters covering enterprise tech and chip beats have a strong track record on personnel scoops like this one, so treat it as credible, single-sourced reporting rather than official confirmation — the same posture this coverage has taken on other recent Anthropic infrastructure stories before they were confirmed.
Chapter three of Anthropic's 2026 compute story
Read on its own, "AI lab hires a veteran chip engineer" is a small story. Read against explainx.ai's coverage of Anthropic's compute buildout this summer, it's the connective tissue:
- Chapter one — July 22, 2026: AMD invests up to $5B in Anthropic for 2 gigawatts of MI450 Helios capacity, with Claude used internally to help close AMD's ROCm software gap against Nvidia's CUDA.
- Chapter two — July 30, 2026: Google backstops a reported $15B Nexus data-center financing package for an Anthropic campus in Hubbard, Texas, filled with TPUs co-designed by Google and Broadcom.
- Interlude — August 5, 2026: TechCrunch reports Anthropic is hiring an in-house AI chip design team, quoting the company's own framing that existing AWS, Google, Nvidia, and AMD partnerships aren't enough on their own.
- Interlude — August 13, 2026: Bloomberg reports Anthropic in talks to acquire Decart for ~$6B, an Israeli startup specializing in squeezing more inference performance out of existing chip fleets — a faster, nearer-term lever than designing silicon from scratch.
- Chapter three — August 21, 2026: Anthropic reportedly hires Amir Salek, giving the custom-chip team announced on August 5 a leader who has already built and shipped this exact category of product.
Each chapter targets a different point on the build-vs-buy spectrum: AMD and Google deals lock in someone else's silicon at scale; Decart (if it closes) buys software that makes existing chips run more efficiently; Salek's hire is the build your own from scratch option, which is slower, more expensive, and only pays off years out — but removes vendor dependency entirely once it does.
Why "reduce dependence" keeps showing up
Every one of Anthropic's 2026 compute moves points at the same underlying pressure: Claude demand has outpaced the capacity Anthropic can reliably secure from any single supplier. TechCrunch's reporting quoted Anthropic's own framing directly — despite deals with AWS, Google, Nvidia, and AMD, "relying on others clearly isn't enough." That's a company saying, in public, that multi-sourcing GPUs and TPUs from partners still leaves it exposed to allocation queues, lead times, and pricing it doesn't control.
Custom silicon is the standard answer to that exposure once a company is large enough to justify the cost. Anthropic is now the third frontier lab to visibly pursue it:
| Lab | Custom chip path | Status as of Aug 2026 |
|---|---|---|
| TPU program, founded by Salek in 2013 | Nine years in production, TPUv1–v8 | |
| OpenAI | Jalapeño, co-designed with Broadcom | Unveiled mid-2026, inference-focused |
| Anthropic | In-house chip design team (announced Aug 5) + Salek hire (reported Aug 21) | Team-building stage — no chip announced |
Hiring the founder of the TPU program specifically — rather than a generic chip-industry veteran — reads as Anthropic wanting to compress the "how do we even start" phase. Salek has already made every mistake a from-scratch custom-silicon program makes once, at Google, on infrastructure with similar scale demands to what Anthropic now runs for Claude.
What this means for Claude users and API pricing
Hedge this appropriately: nothing changes this week. Custom AI accelerators take a long time to go from "hire the lead" to "chips in racks serving production traffic" — typically 18-24 months at minimum for a first-generation design, even with an experienced team and unlimited budget, because chip fabrication, packaging, and software bring-up don't compress the way software timelines do.
What it signals for the medium term, if the effort succeeds:
- More capacity headroom. Fewer Claude availability crunches during demand spikes, the same motivation cited for the AMD and Google deals.
- Better unit economics. Owning silicon designed specifically for Claude's inference patterns can lower cost-per-token more than renting general-purpose GPUs, which is the same efficiency logic behind the reported Decart talks.
- No expectation of an imminent price cut. Anthropic just made Sonnet 5 pricing permanent at $2/$10 per million tokens days before this hire — custom silicon is a multi-year infrastructure bet, not a pricing lever available today.
- Continued multi-sourcing, not a pivot away from it. The AMD and Google/TPU deals remain the near-term capacity plan; a Salek-led chip effort is additive, arriving years later if it ships at all.
Honest limitations
- Not officially confirmed by Anthropic. This rests on Bloomberg's reporting and wire pickups as of August 21–22, 2026 — no press release, no confirmed title or start date.
- No chip specifics exist yet. Nothing reported describes an architecture, target workload (training vs. inference), fab partner, or timeline for Anthropic's custom silicon.
- Salek's exact mandate is unstated. "Compute team" is the only role description in the reporting so far — whether he leads chip design specifically, or a broader hardware strategy function, isn't yet clear.
- Custom silicon programs fail or stall often. Even well-funded efforts can be quietly folded years later; a hire is a strong signal of intent, not a guarantee of a shipped chip.
Bottom line
Anthropic reportedly hired the person who built Google's TPU program from its first generation to its fourth — a rare, specific, and credible signal that Anthropic's custom-silicon ambitions (first reported August 5) are being staffed with someone who has already done exactly this job. It's not a pivot away from the AMD and Google partnerships covered here all summer; it's the third leg of the same stool: rent capacity at scale (AMD, Google), buy efficiency software (Decart, if it closes), and now, build the chip itself, with a leader who has nine years of evidence he can.
Related on explainx.ai
- AMD Invests Up to $5B in Anthropic for 2GW of MI450 Helios — chapter one of Anthropic's 2026 compute buildout.
- Google Backstops $15B Anthropic Texas Data Center (Nexus) — chapter two, and Anthropic's existing Google TPU relationship.
- Anthropic Reportedly in Talks to Buy Decart for $6 Billion — the inference-efficiency acquisition talks that first flagged Anthropic's custom-chip hiring push.
- OpenAI's Jalapeño: First Custom AI Chip, Co-Developed With Broadcom — the rival lab's custom-silicon path, for comparison.
- Google Cloud Next 2026: TPU 8t / TPU 8i — the TPU program Salek founded, as it stands today.
- Anthropic Makes Claude Sonnet 5 Pricing Permanent at $2/$10 — current API pricing this hire does not change.
- Anthropic IPO Path 2026: S-1, Banker Meetings — the capital-markets context behind Anthropic's compute spending this year.
- Dario Amodei Worries New Anthropic Hires Are Chasing Pay, Not Mission — more on Anthropic's 2026 talent strategy.
Sources: Dina Bass on X — Anthropic hires Google TPU founder Amir Salek (Aug 21, 2026) · Seeking Alpha — Anthropic hires former Google chip developer · TechCrunch — Anthropic is hiring an AI chip design team (Aug 5, 2026) · Amir Salek — Cerberus Capital Management bio · Amir Salek — Wikipedia
This article reflects reporting as of August 22, 2026. Anthropic has not issued an official press release confirming Amir Salek's hire, title, or start date as of publication — this post will be updated if Anthropic confirms, denies, or adds detail.