Anthropic Hires the Founder of Google's TPU as It Lays the Groundwork for Its Own Chips

Anthropic has hired Amir Salek — the electrical engineer who was recruited by Google in 2013 to build custom silicon for its data centers and who went on to found and lead the Tensor Processing Unit (TPU) program — to join its compute team as the AI lab lays the groundwork for developing chips of its own. The news, first reported by Bloomberg on August 21, 2026 and quickly picked up across tech media, is the strongest signal to date that Anthropic intends to follow Google, Amazon, and Microsoft down the path of designing its own AI accelerators.

Who Anthropic just hired

Amir Salek is not a household name outside the semiconductor world, but inside it he is one of the most consequential chip architects of the past decade. Google recruited him in 2013 — in his own words, “with the mission (and a dream) to build custom silicon development capabilities for Google’s datacenters.” The program he founded matured into the TPU, the application-specific accelerator that now powers the bulk of Google’s machine-learning workloads and has become Nvidia’s most credible rival in AI inference at scale.

Salek ran Google’s TPU business until 2022, contributing to multiple generations of the chip, including the TPUv4 family. He co-authored “Ten Lessons From Three Generations Shaped Google’s TPUv4i,” a widely cited retrospective on accelerator design presented at the Chips & Compilers for MLSys workshop. After leaving Google, he served as a Senior Managing Director at Cerberus Capital Management, where he advised on semiconductor and technology investments, and he joined the board of AI-chip startup Untether AI. He was also earlier in his career a Senior Director of Engineering at Nvidia — meaning Anthropic’s newest hire has now worked on all three sides of the AI compute equation: chip vendor, chip customer, and chip investor.

At Anthropic, Salek joins the compute team — the group responsible for the lab’s sprawling training and inference infrastructure. Bloomberg frames the hire as “part of a push into hardware,” and Techmeme’s summary is more direct: Anthropic hired Salek “as part of a push to develop its own chips.”

Why this matters: Anthropic’s compute math

To understand why a frontier lab would want its own chip architect, look at Anthropic’s existing commitments. In April 2026, Anthropic announced an expanded partnership with Google and Broadcom for approximately 3.5 gigawatts of next-generation TPU-based compute capacity, coming online starting in 2027 — described at the time as its biggest compute deal yet. That sits alongside its long-running arrangements with Amazon Web Services (Anthropic’s trains on Trainium chips and its models ship as a first-class offering inside AWS Bedrock) and its substantial Nvidia GPU footprint.

That is a rare position: Anthropic is simultaneously one of the largest customers of Google’s TPUs, Amazon’s Trainium, and Nvidia’s GPUs. It gives the lab supply resilience in a market where compute is the binding constraint on progress — but it also means Anthropic’s economics are largely set by other companies’ pricing, roadmaps, and allocation decisions. Every frontier competitor is facing the same squeeze. OpenAI has been widely reported to be working on its own silicon with Broadcom. Google designs its own. Amazon designs its own. Microsoft has Maia. Meta has MTIA. Anthropic was the last major frontier lab without a public custom-silicon effort of its own.

The hire closes that gap — at least in intent. Someone who founded the industry’s most successful custom-AI-chip program does not join a compute team to manage vendor contracts. He joins to build.

What custom silicon would actually mean for Anthropic

Designing an AI accelerator is a multi-year, multi-billion-dollar undertaking, and the realistic near-term payoff is not a shipping chip but leverage. A credible internal silicon program changes Anthropic’s negotiating posture with every supplier: pricing on TPU capacity, allocation priority at Nvidia, and the terms of the AWS relationship all improve when suppliers know the customer has an alternative — even one that is years from landing in a data center.

There is also a technical argument that general-purpose GPUs are an imperfect fit for the inference workloads that now dominate frontier-lab costs. Training runs are massive and rare; inference is continuous and latency-sensitive. Google’s TPU program demonstrated that a workload-specific architecture, co-designed with the software stack that runs on it, can deliver better performance-per-dollar at scale. Salek literally wrote the lessons-learned document on how to do this.

The risk is equally clear: chip design is unforgiving, capital-intensive, and slow. Google took roughly three years from recruiting Salek to a first-generation TPU deployed internally, and a decade more to reach today’s seventh generation. Anthropic, even with Salek on board, would need design partners (Broadcom is the obvious candidate, given the existing relationship), foundry capacity at TSMC, and the patience to iterate through at least one full generation before seeing competitive silicon.

The broader picture: everyone is building chips now

The Salek hire lands in a week when AI infrastructure is dominating the news cycle. Nvidia’s new playbook — licensing model factories and hiring teams rather than acquiring companies outright, as seen in its reported $6 billion licensing deal with coding-model startup poolside — shows how aggressively the compute layer of the industry is being contested. Meanwhile analysts are tallying the industry’s multi-trillion-dollar infrastructure commitments, and even the Swiss National Bank has warned about AI-driven inflation effects from the building boom.

Against that backdrop, a frontier lab hiring away the founder of the TPU program is more than personnel news. It is a statement that Anthropic believes the next phase of AI competition will be decided as much in silicon as in model architecture — and that being purely a customer of everyone else’s chips is not a durable position.

For now, Anthropic remains deeply dependent on its partners: Claude still trains and serves on Google TPUs, Amazon Trainium, and Nvidia GPUs, and the 3.5-gigawatt compute pipeline is locked in through the end of the decade. But the groundwork has formally started. The company that made its name on constitutional AI and interpretability research has just begun the long, expensive journey toward owning the metal its models run on.

Nothing has been announced about timing, chip generations, or design partners, and Anthropic has not publicly detailed the scope of Salek’s mandate. But as Bloomberg’s headline put it, the AI lab is tapping a Google chip veteran “as part of a push into hardware” — and in the compute wars of 2026, that push is the story.