Anthropic in talks with Samsung over a custom AI chip

TL;DR:

  • Anthropic is reportedly in contact with Samsung to explore developing a custom AI chip, according to The Information.
  • The company has not decided what the chip would do or how powerful it would be, and says a diversified hardware stack remains central to its strategy.
  • It follows rival OpenAI’s custom inference chip, “Jalapeño”, built with Broadcom.

Anthropic appears to be getting serious about making its own silicon. The company is in early discussions with Samsung about a bespoke AI chip, TechCrunch reports, building on earlier signals that it was weighing in-house hardware to cushion against chip shortages. Anthropic told TechCrunch that a diversified stack spanning Google, Amazon and Nvidia chips would stay “pivotal” to its compute strategy, and declined to elaborate on Samsung.

The scramble for silicon independence

The move fits a clear industry pattern: frontier labs want custom hardware both to tune performance for specific tasks and to loosen their dependence on Nvidia, still the dominant supplier. The timing is notable — it follows last week’s launch by rival OpenAI of a custom Broadcom-built inference processor, dubbed “Jalapeño”, which OpenAI says delivers better performance-per-watt than competing chips. Amazon and Google already offer their own custom TPUs through their clouds.

Samsung is an unusual but logical partner: it manufactures chips Nvidia relies on, uses Nvidia software to make them, and is building an AI chip factory in South Korea. For UK readers, the story sits upstream of everything else in AI. Compute scarcity and cost are what ultimately gate model access and pricing — the same forces behind Britain’s push for sovereign AI compute capacity. When the labs vertically integrate into silicon, the supply chain that UK firms buy their AI through shifts with them.

Looking forward

Nothing is settled — Anthropic has not fixed the chip’s purpose, power or place in the server rack. But the direction of travel is unmistakable: the biggest labs increasingly treat bespoke hardware as strategic, not optional. If these efforts succeed, they could ease the compute bottleneck that has kept advanced AI expensive. If they stall, they underline just how hard it is to design around Nvidia — a lesson with consequences for every business, British ones included, that runs on someone else’s models.

Share this article