Anthropic Taps Samsung for Custom AI Chip: 2nm Process, Nvidia Exit Ramp

Even Anthropic has had enough — GPUs are too expensive, time to build your own.

Three Key Takeaways

According to The Information on July 2, Anthropic is in preliminary talks with Samsung Electronics to manufacture a custom AI accelerator using Samsung's 2-nanometer (SF2P) process. The project is at a very early stage: the chip's intended workload (training vs. inference), architecture specs, and performance targets are all undecided, and Anthropic could walk away at any point. But choosing Samsung over TSMC is itself a signal — Samsung offers advanced packaging capabilities and in-house HBM memory production, enabling a one-stop chip-plus-packaging-plus-memory solution.

Anthropic has hired Clive Chan, an engineer from OpenAI's chip team. Chan previously worked on OpenAI's custom inference chip, Jalapeño, built in partnership with Broadcom. Multiple outlets interpret this hire as evidence that Anthropic's chip ambitions have moved past the "casual conversation" stage.

This is the latest wave in AI labs' collective push to reduce Nvidia dependence. OpenAI has taped out Jalapeño with Broadcom, Google has TPUs, Amazon has Trainium, and Meta has its in-house MTIA accelerator. Nvidia still commands roughly 80% of the AI training chip market, but its top customers are chipping away at its inference business with custom silicon.

WangDou's Take

Anthropic's 2025 revenue was around $850 million — it burns cash far faster than it earns it — and now it wants to chew on chip design too. But the math makes sense: an Nvidia H100 costs $30,000 a pop, and training Claude requires tens of thousands of them. If a custom inference chip cuts per-API-call costs by even 30%, the annual savings pay for an entire chip team. As for choosing Samsung over TSMC — this probably isn't a technology preference. TSMC's capacity is locked up by Apple and Nvidia; Samsung's 2nm line is hungry for marquee customers. It's a trade where both sides get what they need: Anthropic wants fab capacity, Samsung wants orders and an AI flagship client. Whether the chip ever reaches mass production is another question — but in Silicon Valley, poaching the right engineer is itself a strategic move.

Source: TechCrunch · Bloomberg · The Information