Claude’s maker is hiring engineers to design its own chips, the clearest sign yet that renting compute has stopped being enough.
On August 5, 2026, Anthropic confirmed it is assembling an in-house team to design custom silicon for Claude, describing the effort internally as a “custom silicon team” that will co-design hardware and models together. The job postings, offering between $320,000 and $485,000 for candidates who have already “shipped silicon,” point to a small, senior group rather than a large-scale chip organization. Officially, the goal is efficiency: faster inference, lower power draw, tighter integration between architecture and silicon. Unofficially, arithmetic explains at least as much: Anthropic’s annualized revenue has grown from roughly nine billion dollars at the end of 2025 to more than thirty billion today, with over a thousand enterprise clients spending above one million dollars a year each.
That is close to the logic that pushed Samsung and SK Telecom to stop depending on someone else’s stack and build their own AI native network intelligence, a shift we examined when the RAN stopped being a fixed cost and started becoming a layer worth owning. Anthropic’s move starts from a different position, though. Samsung was already a hardware manufacturer, with memory, foundry and packaging capacity to redeploy; Google and Amazon built TPUs and Trainium after years running verticalized cloud infrastructure they already controlled end to end. Anthropic controls none of that today: for the first time, a model company is asking to also become a chip designer.
The closer precedent is Meta, whose Training and Inference Accelerator (MTIA) chips have run internally since 2020, never sold or rented to anyone, purely to cut the cost of serving its own ranking, recommendation and, more recently, generative workloads. Anthropic’s ambitions look similar in scope: silicon tuned to Claude’s own inference patterns rather than a general purpose accelerator sold to third parties. OpenAI took the more capital intensive route in June 2026, unveiling Jalapeño, an inference chip built with Broadcom; that makes Anthropic the fifth major lab to commit publicly to custom silicon, after Google, Amazon, Meta and OpenAI.
What makes the timing notable is how exposed Anthropic already is to outside supply. The company holds roughly 3.5 gigawatts of Google TPU capacity coming online from 2027, on top of 1 gigawatt already deployed in 2026, alongside up to 2 gigawatts of AMD Instinct MI450 systems tied to a cross investment of up to five billion dollars, plus separate arrangements with AWS and Nvidia. Four external suppliers, each with its own architecture, delivery schedule and pricing power, already carry a coordination cost before a single proprietary chip ships: reducing that dependency is a direct lever on cost per token, not merely a technical preference.
None of this changes anything for buyers over the next twelve to twenty four months. No foundry has been confirmed; talks with Samsung over a two nanometer process remain exploratory, and design to production cycles in this category typically run eighteen to twenty four months even with a well resourced team. Anthropic says its four hardware partnerships stay central regardless of outcome, so for IT leaders sizing AI infrastructure decisions today, the custom silicon team is just a signal worth tracking, not a variable to plan around yet.
Was this move predictable? Given where Google, Amazon, Meta and OpenAI already stood, largely yes: past a certain scale of revenue and compute consumption, designing one’s own silicon stops being optional and becomes the same conclusion every peer has already reached. What remains open is execution: whether Anthropic, starting later and further from manufacturing than any of them, can compress a multi-year hardware cycle into something that matters before the market moves again.
