Anthropic has opened preliminary discussions with Samsung Electronics to manufacture a custom AI inference chip on Samsung's 2nm SF2 process, according to reporting first published by The Information on July 2, 2026. The talks position the $965 billion Claude maker as the latest frontier lab to break its near-total dependence on Nvidia GPUs, and would hand Samsung Foundry the marquee anchor customer it needs to close the gap with TSMC at the leading edge.
Early Hires, Not Yet a Design
Anthropic has not fixed the accelerator's specifications, power envelope or server-rack topology, and no prototypes exist. What has moved is people: Anthropic hired Clive Chan - the second engineer to join OpenAI's custom-silicon team - in early June, straight off the Broadcom-designed "Jalapeño" inference ASIC OpenAI unveiled on June 24. Chan brings direct blueprints for how to design an AI accelerator from the software layer down.
Why Samsung, Why Now
Samsung participated in Anthropic's $65 billion Series H in May 2026 as a strategic infrastructure partner - the only investor in that round that also operates a foundry. Anthropic's talks reportedly cover both Samsung's 2nm Gate-All-Around SF2 process and its advanced 2.5D/3D packaging, letting the chip sit as close to memory as possible. Samsung also has open 2nm capacity previously earmarked for an OpenAI inference NPU whose talks stalled in early June, freeing engineering muscle to redirect at Anthropic.
The Inference Cost Prize
The conversations target inference, not training. OpenAI has said Jalapeño delivered roughly 50% cost savings versus commodity AI GPUs in early benchmarks. Anthropic, serving Claude to millions of enterprise and consumer users a day, faces the same cost structure and stands to unlock similar economics with a chip tuned to Claude's specific transformer workload. Nvidia still holds an estimated 74% of the AI chip market, but rivals are compounding faster than Nvidia's growth can supply.
The Yield Question Still Hangs
Samsung's SF2 yields sat at 50-60% through much of 2025, well below the 70-80% economic threshold TSMC's N2 has cleared with anchor customers Apple and Nvidia. The SF2P variant is reportedly closer to 70% in early 2026, but volume stability is unproven. Anthropic told The Information that AWS Trainium, Google TPUs and Nvidia GPUs will remain central to its compute stack - the Samsung chip, if it ships, would be a three-to-five-year additive layer, complementing rather than replacing existing partnerships like Microsoft's Frontier Company push.
Reporting based on coverage from The Information, TechTimes and TrendForce.