- Lead. Anthropic is in early-stage discussions with Samsung Electronics to develop a custom AI chip, TechCrunch reported July 2, with no terms, specifications or timeline yet disclosed.
- Fact. The move follows OpenAI’s June 2026 announcement of “Jalapeño” — a custom inference processor developed with Broadcom that OpenAI says delivers superior performance per watt versus comparable Nvidia hardware.
- Stake. Every major AI lab is now pursuing or actively developing proprietary silicon, a structural shift that could gradually erode Nvidia’s dominant position as the default compute supplier for frontier AI training and inference.
Anthropic confirmed it is in discussions with Samsung, while offering minimal detail. The company’s statement — that “a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal to its compute strategy” — was careful not to characterise the Samsung talks as a departure from existing partnerships, positioning any custom chip as an addition rather than a replacement. Anthropic declined further comment on specifics of the Samsung discussions, according to TechCrunch. No financial terms, intended use case, or chip specifications have been determined at this stage.
Samsung’s position in the AI chip race
Samsung’s involvement carries strategic weight. The South Korean conglomerate already manufactures chips for Nvidia — whose H100 and B200 series remain the workhorse of frontier AI training — and has an established chip co-development relationship with Google. Separately, Samsung and Nvidia are jointly building an AI chip fabrication facility in South Korea. Samsung’s foundry capabilities, combined with its existing relationships across the AI supply chain, make it a credible partner for a lab looking to develop proprietary inference hardware without building its own fab.
The Anthropic discussions were presaged by April reporting from Reuters that the company was exploring custom chip production to address recurring compute shortages. The Samsung talks represent the most concrete step yet disclosed toward that goal.
The race from Nvidia
The Anthropic announcement fits a pattern accelerating across the AI industry. OpenAI and Broadcom unveiled Jalapeño in June 2026 — the first custom AI inference chip disclosed by a frontier lab — with OpenAI describing efficiency gains versus off-the-shelf Nvidia alternatives. Google has long operated its own Tensor Processing Units. Amazon’s Trainium and Inferentia chip families serve AWS customers and internal Bedrock workloads. Meta’s MTIA inference chip is in production deployment.
Custom silicon offers two primary advantages: efficiency, measured in performance per watt or per dollar, and independence from Nvidia’s supply allocations and pricing power. For Anthropic, which runs Claude across AWS infrastructure, a proprietary inference chip could reduce the marginal cost of each API call as inference volume scales — a significant operational advantage if the company’s revenue growth continues at its current trajectory.
What remains unknown
The talks are early enough that the fundamental parameters — whether the chip would target training, inference, or both; whether Samsung would manufacture exclusively or alongside other foundries; and what investment Anthropic would commit — have not been determined. The absence of a signed deal also means the discussions could widen to include TSMC, which manufactures most of Nvidia’s chips and has deeper advanced-node capacity than Samsung’s current process nodes. For now, the Anthropic-Samsung talks signal intent and direction, not a committed timeline.