Anthropic is in talks with Samsung to build a custom AI chip

Anthropic's compute strategy has always been plural — it trains and serves across Google's TPUs, Amazon's Trainium and Nvidia's GPUs rather than betting on a single vendor. A new report suggests it wants a fourth option it helped design. According to The Information, the company has held preliminary talks with Samsung Electronics about collaborating on a custom AI accelerator, potentially fabricated on Samsung's 2nm process. Crucially, the report frames this as early-stage: Anthropic has not settled on what the chip is for, how powerful it should be, or whether it proceeds at all.

The signal here is exploration, not commitment. The talks are preliminary, the specification is undefined, and no agreement has been announced. What is confirmed is the direction of travel: Anthropic told TechCrunch that a diversified hardware stack remains central to how it plans to buy and build compute, and a Samsung part would slot in as an additional leg rather than a wholesale switch away from its existing suppliers.

Read plainly, that positioning does two things. It confirms the talks are real, and it pre-empts the obvious misreading — that Anthropic is abandoning Nvidia or its cloud partners. A custom Samsung accelerator, if it happens, would be a hedge and a cost lever, not a rupture.

Anthropic's current silicon footprint spans three suppliers, each playing a role. A Samsung part would be a fourth, with its brief still to be defined:

The pattern is not unique to Anthropic. OpenAI has worked with Broadcom on its own accelerator effort, and Amazon has pushed Trainium precisely to lower the cost of frontier compute. The through-line is that the minimum unit of competition has moved from the model to the fleet: when serving millions of agentic requests, a few cents per thousand tokens compounds into the difference between a viable margin and a loss.

That is sharpened by the moment. The report lands during a broader efficiency crunch, as buyers shift spend away from raw scale toward cost-per-token and cheaper open-weight alternatives crowd the market. Owning a slice of your own silicon is one of the few durable levers on that economics — and on exposure to a compute supply chain under growing political strain.

The honest caveat: this is a report of early talks, not a signed programme. Custom silicon is a multi-year, capital-heavy commitment, and a 2nm accelerator would not touch Anthropic's economics for a while even if the deal closes tomorrow. But the strategic tell is clear — Anthropic wants to shape the metal its models run on, not just rent it.

It is worth situating this in Samsung's own strategic position, because the deal would cut both ways. Samsung's foundry business has trailed TSMC at the leading edge, and a marquee AI customer designing to its 2nm process would be exactly the credibility it needs to close that gap. So the reported talks are as much about Samsung buying its way into the AI-accelerator conversation as about Anthropic diversifying supply — a mutual hedge in which each party de-risks a dependence it would rather not carry alone. That mutual interest is part of why the report is plausible even though nothing is signed, and part of why it would be a mistake to read it as merely a procurement footnote.