Anthropic in early talks with Samsung on custom AI accelerator chip; Clive Chan hired from OpenAI

Anthropic is in preliminary discussions with Samsung Electronics to manufacture a custom AI chip, The Information reported July 2, 2026. The talks remain at an exploratory stage with no finalized design, performance targets, or intended workload (training vs. inference) yet determined. Anthropic has hired Clive Chan, an engineer who spent two and a half years on OpenAI's dedicated chip team building its Broadcom-designed Jalapeño inference accelerator, signaling the move from exploration to active development. Chan's hire in early June preceded the public disclosure, suggesting Anthropic is advancing hardware roadmap work in parallel with its software initiatives.

Samsung offers multiple potential advantages to Anthropic: 2-nanometer foundry process (SF2P), advanced packaging facilities optimized for AI/data center chips (gate-all-around transistors, power-efficiency focused), and in-house HBM memory production. Samsung participated in Anthropic's May 2026 Series H funding round as a "strategic infrastructure partner" alongside SK Hynix and Micron. The move mirrors OpenAI's June 24 unveiling of Jalapeño, a custom inference processor built with Broadcom that reportedly delivers ~50% cost savings versus standard GPU inference in early testing.

Anthropic's chip effort reflects a broader industry pivot toward vertical integration and supply-chain diversification away from near-total Nvidia dependence. The company reported annualized revenue run-rate surpassing $30 billion in early 2026, making custom silicon economics increasingly attractive. Success would give Anthropic tighter hardware-software integration, lower inference costs, and differentiation against competitors. Samsung's foundry business, historically underperforming TSMC, would gain a marquee frontier AI customer to strengthen its margin position. Neither company has disclosed timeline, design partner selection (OpenAI worked with Broadcom), or deployment targets.