Analysis
Samsung Electronics is reportedly in talks to manufacture custom 2-nanometer AI chips for Meta as well as Anthropic, according to SamMobile's July 4 report building on The Information's original July 2 scoop about the Anthropic discussions. If both materialize, Samsung would be manufacturing frontier AI silicon for two of the largest AI compute buyers in the world simultaneously.
The Anthropic conversation centers on a custom accelerator using Samsung's advanced packaging to place compute closer to memory, cutting data-transfer bottlenecks; Anthropic has hired Clive Chan, an early member of OpenAI's custom-chip team, as part of the buildout. Meta's interest follows the same logic driving Google's TPU, Amazon's Trainium and OpenAI's in-house 'Jalapeño' chip: reducing dependence on Nvidia GPUs for a growing share of inference and training workloads, especially as Meta guides to $115 billion-$135 billion in 2026 capital spending on chips, land and power.
For Samsung, landing two frontier AI labs as custom-silicon customers on its leading-edge 2nm node would be a meaningful competitive answer to TSMC, which has manufactured the bulk of the industry's most advanced AI accelerators to date, including for Nvidia and Apple. Samsung, alongside SK Hynix and Micron, also participated in Anthropic's $65 billion May fundraising round, giving it a financial stake in Anthropic's success beyond a pure manufacturing contract.
Both conversations remain exploratory -- no design has been finalized and no production timeline has been disclosed for either customer -- but the pattern of frontier AI labs and hyperscalers pursuing custom silicon in parallel, rather than sequentially, suggests foundry capacity at the leading edge is about to get considerably more contested.
What to watch: whether Samsung converts either conversation into a signed manufacturing agreement, and whether TSMC responds with pricing or capacity commitments aimed at keeping Meta and Anthropic on its own roadmap.
