Anthropic wants control of the hardware underneath its models. The company has confirmed that it is assembling a “custom silicon team” tasked with designing chips for Claude — a path OpenAI, Google and Meta have already taken.
Confirmation arrived awkwardly, by way of a careers page. Business Insider caught a posting yesterday seeking a senior engineer who has shipped semiconductor designs. Two roles are currently open: a silicon engineer and a technical program manager, silicon. Anthropic’s spokesperson subsequently confirmed the plans to Business Insider and TechCrunch alike.
One detail deserves attention: according to the spokesperson, the company intends to maintain a “multi-chip approach.” In other words, third-party hardware will sit beside its own designs as it continues to scale. Nobody is being cut loose here.
The rumor was already out there
This didn’t appear out of thin air. The Information had already reported that Anthropic was weighing Samsung as a manufacturing partner for hardware, and the recruitment drive fits that account.
OpenAI beat everyone to the announcement. Jalapeño, its custom chip, was developed alongside Broadcom and targets large language model inference inside data centers. Google has been running its models on homegrown hardware for some time. Meta has both designed and deployed chips of its own. Mistral is said to be weighing the same step.
Nvidia is the reason, mostly
A large share of the industry sits on Nvidia hardware, and that degree of concentration amounts to a strategic weak point. Compute infrastructure is being fought over hard at the moment, with demand continuing to exceed what capacity exists. Leaning on a single supplier in a market like that is an uneasy place to be.
Performance is the other motivation. Building chips around specific models — and models around specific chips — ought to beat what off-the-shelf hardware delivers. Should OpenAI cash in on that vertical integration, Anthropic and every other frontier model provider will be chasing the same advantage.
The company says its teams will co-design fresh hardware and models in parallel. Co-designing hardware with partners is nothing new for Anthropic. The change is bringing the silicon expertise inside the building.
There’s an edge-computing angle too
Developers and other users are increasingly running cheaper, smaller or open-weight models on hardware they own and on edge devices. Custom silicon may be Anthropic’s bet on giving its frontier models a quality those setups can’t replicate.
The team, though, is still hypothetical. Anthropic has yet to hire the people who would carry out the work, and until those chairs are occupied, neither the company nor anyone using Claude gets anything out of it.