Anthropic Turns to Samsung's 2nm Foundry for Its First Custom AI Chip, Reports Say

Anthropic is building its own chip. Reports circulating July 14 say Samsung's foundry division has agreed to manufacture the Claude maker's first custom AI accelerator on its cutting-edge 2-nanometer process, using Samsung's advanced chip-packaging technology. Neither company has officially confirmed the arrangement, and earlier reporting described the discussions as preliminary — but the direction of travel is unmistakable: Anthropic is joining the race to design the silicon its models run on rather than renting all of its compute from Nvidia.

The story broke on July 2, when The Information reported that Anthropic had approached Samsung as a manufacturing partner for a pending processor. At that stage the project was early: people familiar with the effort said Anthropic was still determining what the chip should do, how powerful it needed to be, and how it would slot into a server, and that the company had not yet begun detailed design, testing, or manufacturing. Choosing Samsung's 2nm node — a process that trails TSMC in the foundry market — would also hand Samsung a marquee AI customer at a moment when it is fighting to close that gap.

Anthropic has been quietly assembling the talent to pull this off. In June it hired Clive Chan, the second hardware engineer to join OpenAI's custom-chip program, who worked on that effort from its early stages. The financial firepower is already in place: Anthropic closed a $65 billion Series H on May 28, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, with Samsung, SK Hynix, and Micron joining as strategic infrastructure partners responsible for the supply of memory, storage, and logic chips. The Samsung chip talks are a natural extension of that alliance rather than a bolt from the blue.

The move puts Anthropic in increasingly crowded company. OpenAI unveiled its first custom inference chip, Jalapeño, with Broadcom in late June; Google has shipped its own TPUs for years; and Amazon builds Trainium and Inferentia silicon that Anthropic already uses alongside Google's TPUs and Nvidia GPUs. What each of these labs is chasing is the same prize — lower cost per token, tighter control over supply, and an architecture tuned to their own models — at a time when GPU allocation is the single biggest constraint on scaling and inference bills dominate the income statement.

For Anthropic specifically, the timing matters. The company is running at a roughly $47 billion annualized revenue pace and is widely reported to be preparing an S-1 for a public listing later this year, which means every structural lever on gross margin — including owning its accelerator roadmap — becomes part of the pitch to investors. A first-generation, unconfirmed chip years from volume won't change Anthropic's compute mix in 2026. But the signal is loud: the era in which frontier labs simply bought whatever Nvidia would sell them is ending, and custom silicon is becoming table stakes for anyone who wants to run models at global scale.