Anthropic is in talks with Samsung about manufacturing a custom AI chip, The Information reported this week — with the Korea Herald adding that the discussions center on Samsung's 2-nanometer foundry process and advanced packaging. The plans are early: Anthropic is reportedly still working out what the processor should do and how powerful it should be, and is speaking with multiple chip design firms.
Asked about it, Anthropic didn't confirm a partnership — the company told TechCrunch that a diversified hardware stack built on Google, Amazon, and Nvidia chips "will continue to be pivotal" to its compute strategy, and said it had nothing further to add. That's not a denial. It's what you say while you're still negotiating.
The pattern, not the deal
Zoom out and the news is bigger than one conversation. Reuters reported back in April that Anthropic was weighing building its own chips in response to shortages. Last week, OpenAI announced a custom inference processor — dubbed Jalapeño — built with Broadcom. Google has run its models on in-house TPUs for a decade; Amazon has Trainium. Every serious player in AI is concluding the same thing: when compute is your biggest cost and your hardest constraint, renting it at someone else's margin stops making sense.
Why Samsung, and why it moved markets
For Samsung, a frontier-lab logo on its 2nm line would be the marquee foundry win it has chased for years in its pursuit of TSMC. Investors did the math instantly: the report was cited among the drivers of Friday's violent rebound in Korean chip stocks, with Samsung up 8.2% — and it lands squarely inside Korea's $880 billion national AI buildout, which needs exactly this kind of anchor demand.
Our take: Talks are not tape-outs — custom silicon takes years, and "exploring options" is also a negotiating tactic with Nvidia, whose roughly $4.7 trillion valuation assumes the renting continues. But the direction is now unambiguous: model companies are becoming compute companies. The winners of the next leg won't just have the best models — they'll have the cheapest tokens. Watch who signs, not who talks.
What to watch next
- Commitment signals: a design-firm selection or a Samsung capacity reservation would turn this from exploration into strategy.
- Samsung's 2nm roster: foundry economics are won on anchor customers. One frontier lab tends to attract another.
- Nvidia's response: every custom-chip headline is pricing leverage for its biggest buyers. Watch the next round of supply deals.