Anthropic confirms it is building an in-house team to design its own AI chip

Anthropic confirmed on Wednesday that it is building a team to design its own chip for AI workloads. The Claude maker said it plans to co-design hardware and models so its technology runs faster and more efficiently, according to the company. Business Insider was first to report the move; Anthropic later confirmed it directly to TechCrunch.

The job listing describes the group as a "custom silicon team" and is looking for engineers with experience in chip design. That is about where the public details end. There is no chip name, no roadmap, and no target date. What Anthropic has confirmed is intent, not a product.

Samsung is the rumored manufacturer

Last month, The Information reported that Anthropic was scouting Samsung as a potential partner to build the chip. Anthropic has not commented on that report, and there is no official confirmation of a foundry relationship. With the company not disclosing who would manufacture the chip, that detail remains a report, not a fact.

The bet comes from a compute crunch

Anthropic is already one of the biggest consumers of third-party AI compute. The company has signed deals with AWS, Google, Nvidia and AMD to access AI computing hardware, per TechCrunch. The in-house silicon push arrives as demand for Claude rises and AI companies grab every infrastructure deal they can. TechCrunch's read is blunt: to scale to that level of demand, relying on other companies clearly is not enough.

Anthropic is not the first AI company to go down this road. OpenAI unveiled its Broadcom-built Jalapeño chip in June, designed specifically for inference workloads. Google DeepMind has long run its models on Alphabet's TPU chips, and Meta has been developing its own MTIA accelerators.

For teams building on these models, the near-term effect of Anthropic's move is small: there is no chip, no specs, and no announced timeline. What is practical to take away is the strategic signal: another frontier lab has decided to own part of its compute stack instead of renting all of it.