Anthropic has opened talks with Samsung about a custom AI chip, The Information reported Thursday, though the company has not decided what the chip would power, how it would fit into a server rack, or how much compute it would carry. The stakes are simple: every frontier lab now wants leverage against Nvidia’s pricing and allocation queue, and Anthropic just signaled it wants a fifth supplier. What the disclosure does not tell us is whether this becomes a shipping product or a due-diligence exercise that quietly fades.
TechCrunch asked Anthropic directly about the report. The company said its compute strategy would keep relying on chips sourced from Nvidia, Google, and Amazon, and that it had nothing else to share regarding Samsung specifically. Confirming a relationship while declining to describe its scope is itself a signal. It suggests conversations are real but nowhere near a signed agreement.
This is not Anthropic’s first hint at proprietary silicon. Reuters reported in April that the company was weighing an in-house chip program to reduce its exposure to compute shortages. Thursday’s report moves that idea from internal deliberation to an external conversation with a named manufacturing partner, though the company’s own hedging suggests the shift is preliminary rather than settled.
The timing invites a direct comparison. One week earlier, OpenAI and Broadcom announced a finished custom chip, an inference processor named Jalapeño that OpenAI claims outperforms rival silicon on power efficiency. Anthropic’s Samsung talks surfaced days later and are, by its own account, unresolved on basic design questions. That gap, a completed chip announcement from OpenAI against an open-ended conversation from Anthropic, is the clearest evidence of how far apart the two labs are on this front.
Custom silicon appeals to labs burning through billions in training and inference costs for three reasons: it can be tuned to a company’s specific workloads rather than general-purpose GPU design, it caps exposure to Nvidia’s order backlog, and at scale it can undercut the margin Nvidia builds into every chip it sells. Amazon and Google have already proven the model works, each running its own TPU line inside its cloud business for years.
Samsung’s position complicates the independence narrative. The company already fabricates chips that Nvidia depends on and, in turn, licenses Nvidia’s software for its own manufacturing process. The two are jointly building a chip plant in South Korea, and Samsung is separately said to be in early discussions with Google about another manufacturing partnership. A Samsung-built chip for Anthropic would sit inside a supply web that already touches two of Anthropic’s current chip vendors, which raises a fair question: how much diversification does a new supplier actually buy when that supplier is this tightly wound into the existing one?
None of this changes what Anthropic runs on today. Nvidia, Google, and Amazon hardware remains, per the company’s own statement, central to its stack, and custom chip programs typically take 18 to 24 months to move from early design talks to production silicon. Teams building on Claude should watch for a concrete production timeline or a named chip in a future model release, not this week’s headline, before treating Anthropic’s compute story as having changed.
Reported by TechCrunch on July 2, 2026.
