Anthropic confirms an in-house chip team for Claude, with Nvidia, AMD, and Google still in the mix

Anthropic has confirmed, for the first time publicly, that it is building an in-house silicon team to design custom chips for Claude. The build-out is paired with a multi-chip strategy that keeps Google, Nvidia, and AMD central to how the company scales.

The confirmation came through a spokesperson statement to Business Insider, and it is the first time the company has put a name to the in-house silicon effort that earlier reporting had only sketched. Anthropic is not pitching a break-up with Nvidia. It is pitching a world where the lab co-designs chips with the existing supplier base rather than replacing it.

The trail leading here is short. In April 2026, Reuters reported that Anthropic was weighing custom-chip development. CNBC's write-up the next day carried the same sourcing. The Information separately reported, second-hand in the Business Insider piece, that Anthropic had held talks with Samsung Electronics as a potential manufacturing partner, a detail still unconfirmed at the foundry level and best treated as exploratory, not signed.

What is signed is the hiring. Anthropic's careers page lists a role on the "custom silicon team" with a salary range of $320,000 to $485,000, a band for one specific engineer position, not a team-wide figure. The job description calls for "direct personal contribution" to chip finalization and shipping, which puts the role on the tape-out side of the work rather than at the research-and-prototype stage.

The economic logic is the part the wire cannot easily explain. A frontier lab running Claude at customer scale pays for inference every time a user prompt hits a model. Even a modest per-token efficiency gain compounds into a meaningful cost delta when the model is serving hundreds of millions of requests. Co-designed silicon lets the lab tune the chip to the model's actual arithmetic patterns: which kernels it leans on, which memory access patterns it triggers, which precision it can run at without quality loss. Off-the-shelf GPUs are generalists. A custom chip can be a specialist, and the specialist wins on the metric the lab cares about.

A specialist that replaces a generalist is a different bet. Anthropic's multi-chip strategy has the company saying, on the record, that the custom-silicon work is additive, not substitutive. The lab gets the co-design benefit on the workloads where it pays off, and the existing supplier base keeps absorbing the training and inference workloads where commodity silicon is already good enough.

That makes Anthropic the most deliberate entrant in a category OpenAI has already defined. OpenAI unveiled in June its Broadcom-built purpose-built inference chip, internally dubbed Jalapeño, and positioned it as a layer on top of, not a replacement for, the existing compute base. Anthropic's announcement puts the two frontier labs on parallel tracks: co-designed custom silicon for inference, multi-vendor supply for everything else.

Several pieces of the Anthropic picture are still genuinely unconfirmed. The chip architecture is not public. The foundry partner is not named. The tape-out timeline is not on the record. The Samsung talks, as referenced through Business Insider's recap of The Information, are talks, not a deal. Treating any of those as confirmed would be reading intent as silicon.

The next signal to watch is straightforward: a foundry announcement, a tape-out, or a public migration of a Claude inference workload onto Anthropic-designed silicon.