Compute

Anthropic taps Samsung
for its first
custom AI chip,
targeting 2nm

The Claude maker has opened early-stage talks with Samsung's foundry to manufacture a bespoke accelerator on the SF2 node, one month after Samsung joined Anthropic's $65B Series H — and weeks after a rival Samsung project with OpenAI collapsed.

Anthropic has opened early-stage talks with Samsung Electronics to manufacture its first custom AI chip on the SF2 2-nanometer process, The Information reported on July 2, with Bloomberg confirming the discussions the same afternoon. The specs, target workload, and server integration remain undefined; Korea Herald notes the project hasn’t reached detailed design, testing, or manufacturing.

The timing is the story. Samsung joined Anthropic’s $65 billion Series H in May 2026 as a strategic infrastructure partner alongside SK Hynix and Micron, and when Anthropic announced the round it flagged that its memory-maker partners would help supply logic chips as well. That line, largely overlooked at the time, now reads as a foundry cue. Samsung’s foundry arm is trying to close the gap with TSMC and return to profit after years of losses. A marquee Claude customer would be exactly the anchor tenant it needs.

It arrives, too, in the shadow of a failure. Samsung had been developing an ARM-based inference NPU for OpenAI, and those talks stalled in early June over what Korean media called “strategic differences.” Sam Altman cancelled a planned Seoul visit. Weeks later, on June 24, OpenAI unveiled “Jalapeño,” a Broadcom-designed inference accelerator that TechCrunch reports offers better performance-per-watt than competing chips. Anthropic, meanwhile, hired Clive Chan in early June, the second engineer to join OpenAI’s dedicated custom chip team, who’d spent roughly two and a half years on Jalapeño. The talent flowed one way while the foundry deal reshuffled the other.

Some reports peg SF2P yields around 70 percent in early 2026, which is either a genuine turnaround or a number Samsung’s foundry very much wants circulating. The Anthropic project would test it.

Anthropic isn’t abandoning its multi-vendor posture. AWS Trainium, Google TPUs, and Nvidia GPUs remain central to how it scales compute, and The Information reports it’s also in conversations with Microsoft and U.K. startup Fractile. Reuters had noted in April that Anthropic was “toying with the idea” of its own silicon in response to chip shortages. The Samsung talks turn that toying into a supplier shortlist.

The structural read is straightforward. The Information estimates Nvidia now holds 74 percent of the AI chip market, higher than before the custom-silicon race began, because training demand has outrun the alternatives faster than anyone could tape out a replacement. Every frontier lab is now expected to have a chip program, less because the economics obviously work than because not having one has become illegible to investors. Anthropic is joining the ritual on a schedule that suggests it agrees.

Sources