The news hit the hardware world like a forgotten alarm finally going off. Anthropic, the Paris-born AI darling that has spent years buying compute like a tourist in a duty-free shop, has quietly signed Amir Salek. The man who shipped seven generations of Google’s TPUs is now the senior director of silicon at Anthropic. On paper, it’s a hire. In practice, it’s the first loud, unmistakable signal that the company has stopped renting the future and started trying to own the factory.
Salek doesn’t join to audit GPUs. His mandate is definition: chip architecture, compiler stacks, data-center integration, and the brutal choreography of getting silicon out of a fab and into a server rack. The implication is direct. Anthropic is no longer willing to be a pure model company. It wants to be an infrastructure company that happens to have a world-class model.
For a year, the market has been reading this tea leaves. OpenAI’s Jalapeno project — a custom inference accelerator built with Broadcom — has been running since 2025. Google has TPUs. AWS has Trainium and Inferentia. NVIDIA is a nine-hundred-pound gorilla in a gorilla suit. Anthropic’s position was dangerously pure: a brilliant model company with no hardware identity. This hire ends that narrative.
The architecture angle matters more than the headlines. The common reading is that Anthropic wants to replace NVIDIA. The technical reading is different. Salek’s TPU experience isn’t about building a general-purpose GPU rival. It’s about designing an ASIC that is so tightly coupled to Claude’s architecture that it becomes a competitive moat. The target scenarios are the ones that actually hurt the wallet: inference, long context, MoE, and KV cache behavior. Anthropic is not trying to win the H100 benchmark. It’s trying to make a single token so cheap that enterprise adoption becomes a no-brainer.
That is the true breaking point. The market has always valued Anthropic on the strength of Claude’s output, but the cost of that output is the silent killer. They buy from NVIDIA, Google Cloud, and AWS, which means their cost structure is dictated by third parties. Salek changes the physics of that. A custom accelerator that trims inference cost by 30 to 40% doesn’t just improve margins. It rewrites the API pricing table. That is the kind of edge that makes a competitor unable to catch up.
But let’s not get high on our own supply. The hard truth about chip projects is they are capital-intense, brutal, and slow. The best-case scenario for Anthropic is that the first chip is a custom inference accelerator, manufactured via a TSMC-like partner, deployed in a hybrid cloud model. They will not build a data-center network from scratch. They will likely use their existing AWS and Google Cloud relationships as the physical layer.
The contrarian angle is that this is also a defense move, not just an offense one. The risk of a three-player model is that NVIDIA raises prices, or a hyperscaler prioritizes its own internal models over yours. By owning the silicon roadmap, Anthropic is hedging against both. It’s the same reason why independent artists start to print their own merch. Control over the means of production is the only stable power.
A Note on the security angle: A closed silicon stack is a closed audit path. It gives Anthropic more control over data flow, but it also makes third-party security assessment harder. In the bear market, survival matters more than gains. And in the AI market, survival matters more than feature speed. The firm’s “safety-first” brand may clash with a closed hardware stack if they don’t provide visibility to auditors.
What we’re really watching is a shift in the meta-game. It’s not just about the model’s intelligence anymore. The contest is now about the entire system: the chip, the compiler, the network, the data center, and the software stack. OpenAI has the Jalapoc lead. Google has TPU. AWS has its own chips. Anthropic now has the right general to lead its own battalion.
The strategic intent is clear, but the execution is a massive capital-intensive bet. The next signal isn’t a press release; it’s a job posting. Look for expansion in the silicon team: architecture, backend, compiler, and interconnect roles. If that happens in the next six months, the race is on.
Will Anthropic’s chip ever get a name like “Jalapoc”? Maybe. But in a market where speed is the only truth, the very fact that they are now building is the news. The dance has begun, and this time, the tune is being written in silicon. I won’t regret the dance. The only question is if the AI’s capex will keep up with the narrative.
Volatility isn’t a bug. It’s the byproduct of a market realizing that the AI competition is moving from the model layer to the physical layer. That’s a shift worth watching.