clauding.de
DE
• 3 min read AI-generated

Anthropic Is Building Its Own Chip Team — and Wants to Co-Design Hardware With Its Models

Copy article as Markdown

Anthropic is hiring engineers for a 'custom silicon team' and has confirmed it: the Claude maker wants to build its own AI chip, co-designed with its next models. Samsung is reportedly in talks as the manufacturing partner.

Anthropic Hardware Infrastructure
Featured image for "Anthropic Is Building Its Own Chip Team — and Wants to Co-Design Hardware With Its Models"

The rumors had been floating around for a while. Now it is official: Anthropic is assembling a team to design its own AI chips. A spokesperson confirmed to Business Insider on August 5 that the Claude maker plans to co-design its hardware and models going forward.

Why build your own chip?

The logic is simple. Off-the-shelf AI accelerators rarely match the model running on them — sometimes a bit too much memory, sometimes too little. Design the chip and the software together and you squeeze out exactly those inefficiencies. And for Anthropic there is real money on the line: inference — running the models in production, day in and day out — is usually the single biggest infrastructure cost. Custom silicon can be far cheaper than off-the-shelf graphics cards.

So far Anthropic has struck deals with AWS, Google, Nvidia and AMD to get compute. But with demand for Claude climbing, that clearly is not enough anymore. Hence the move into its own hardware.

The interesting part: Claude tests the chips

In a job posting, Anthropic is looking for chip designers and — maybe even more telling — verification experts. Verification is the art of making sure a chip design actually works the way it is supposed to. And here is where it gets recursive: according to the listing, Anthropic plans to automate parts of that verification with AI. The idea is to build simulations in which Claude learns to test new chip designs. So Claude would help build the hardware that Claude later runs on.

Anthropic is not alone here. In June, OpenAI unveiled its Broadcom-built inference chip, “Jalapeño” — designed in just nine months because so many steps were automated with AI. Google has leaned on its own TPUs for years, and Meta builds its MTIA accelerators. Like Jalapeño, Anthropic’s chip will likely be optimized for inference.

Reports point to Samsung as a possible manufacturing partner. Samsung’s new zHBM technology stacks memory directly on top of the logic cores instead of beside them — shorter distances, lower power draw. For an AI chip, that is a genuine edge.

My take

This is more than a hiring note. Building your own chips is vertical integration at scale — and it loosens your dependence on Nvidia and the cloud providers. The real lever is cost per token: if Anthropic can bring inference costs down, things that are expensive today get cheaper.

And the image of Claude helping design the next generation of chips captures pretty well where we are right now. Nothing is shipping yet — there is a team, a job posting and a plan. But the direction is clear.

Sources: