Anthropic Negotiates Proprietary Chip with Samsung and Challenges Nvidia's Oligopoly
The Information and Bloomberg revealed on July 2 that Anthropic wants to use Samsung's 2-nanometer process to design its first custom silicon. The era of chipless labs is ending.
The Request Made to Samsung
The Information revealed this Thursday that Anthropic is in talks with Samsung Electronics to manufacture its first proprietary artificial intelligence chip, a fact later confirmed by Bloomberg and Korean outlets such as Korea Herald. The target is the 2-nanometer process of the South Korean foundry, the same node where Samsung is trying to gain ground after losing much of the 3nm race to TSMC.
The American company stated to TechCrunch that "a diverse stack of hardware, including chips from Google, Amazon, and Nvidia, will remain central to our compute strategy." The statement acknowledges the project without confirming it. According to The Information, Anthropic is still deciding what the chip should do, how powerful it will be, and how it will fit into the rack. It may not even make it off the drawing board.
A More Concrete Signal: Hiring
The natural skepticism regarding chips that do not yet have a specification sheet encounters a factual data point: in June, Anthropic hired Clive Chan, an engineer who was part of the custom silicon team at OpenAI. Such hiring is costly and rarely signals mere curiosity. Samsung had already entered Anthropic's capital in Series H of $65 billion, closed on May 28 with a valuation of $965 billion, alongside SK Hynix and Micron, the only three manufacturers of HBM memory on the planet.
The reading for CIOs and CFOs of major compute contractors is twofold. On the software side, the three frontier labs are now pursuing custom silicon: OpenAI unveiled Jalapeño in June, an inference chip co-designed with Broadcom, and Google has been using TPU for a decade. On the hardware side, Nvidia, with a 74% market share in AI chips according to estimates from Wall Street analysts, is beginning to see its largest customers designing alternatives in parallel. This is not abandonment; it is negotiating pressure and a secondary supplier roadmap.
South Korea and Japan in the Memory Funnel
The dependence on HBM became explicit when Samsung, SK Hynix, and Micron appeared simultaneously on Anthropic's cap table for the first time in industry history. The three manufacturers have already sold their entire capacity for 2026, and the imbalance between supply and demand for HBM, according to Korean estimates cited by Korea Herald, is between 20% and 50% until 2028. Investing in Anthropic is like purchasing a preview of what clients will demand in two cycles.
For Samsung, the eventual foundry contract serves a secondary function. The company negotiated with OpenAI for an ARM-based inference chip until early June, when conversations cooled due to, according to the Korean press, "strategic divergences." An agreement with Anthropic would reestablish the South Korean foundry division in the frontier silicon game on equal footing with TSMC, which currently dominates the high-end. In Tokyo, the reading is of interest to Japanese hyperscalers like SoftBank and KDDI, who buy HBM from Micron and SK Hynix and already face smaller allocations in this two-year period of constraint. Each chip that Anthropic takes from TSMC is capacity that remains for HBM clients who are not frontier labs.
What Weakens the Thesis
The opposing view has a name. Analysts from Bernstein and Morgan Stanley have been reminding for quarters that custom silicon projects typically take 24 to 36 months from the first tape-out to production at scale, and the unit cost of a frontier chip only breaks even at volumes that currently only workloads like Google Search can support. Anthropic may get there if the Claude Sonnet 5 family and the Mythos line maintain their daily token curve, but there are no guarantees. It is plausible that the conversation with Samsung will follow the path of the OpenAI-Samsung deal: prospecting that does not turn into a contract.
The point that Anthropic's reading does not change: the hyperscaler buying GPU in 2026 is no longer comfortable only purchasing GPU. Amazon launched Trainium 3 in May and is reserving half of its internal training compute for its own silicon, Google has scaled TPU v7 and charges a latency premium on Vertex AI, and Meta is developing MTIA for recommendation inference. Until this week, Anthropic was the only one of the major independent labs without a known proprietary project. It is no longer.