Anthropic and Samsung Eye Custom AI Chip, Still Figuring Out What It Does

Anthropic hasn't even decided what its custom chip with Samsung will do, how fast it will run, or where it plugs into a server — yet the company is publicly exploring the partnership anyway, and that tells you everything about the state of AI hardware.

The Undefined Chip That Challenges Nvidia's Grip

The Information reported Thursday that Anthropic is in discussions with Samsung to collaborate on a custom AI chip. According to the report, Anthropic hasn't settled on the chip's use case, its power profile, or even how it fits into a server rack. When TechCrunch asked, Anthropic offered boilerplate about diversifying its hardware stack with chips from Google, Amazon, and Nvidia — and said nothing else. This vagueness isn't incompetence; it's a deliberate admission that the company needs its own silicon, but doesn't yet know exactly what shape it should take.

Chip shortages have been squeezing AI companies for months. Back in April, Reuters reported Anthropic was toying with the idea of building its own chips. Now they're serious enough to call Samsung, which already manufactures Nvidia's AI chips and is building a massive AI chip factory in South Korea with Nvidia. Samsung also supplies Google's chip-making efforts. For Anthropic, Samsung is a logical partner: deep fabrication expertise, existing AI pipeline, and a seat at Nvidia's table.

Why Samsung Wants In (And Why OpenAI's Jalapeño Matters)

Samsung has its own reasons to work with Anthropic. The chip giant wants to reduce reliance on Nvidia's software stack while expanding its custom-chip client list. A deal with Anthropic diversifies Samsung's AI portfolio beyond the Nvidia factory deal and the Google discussions.

But the elephant in the room is OpenAI. Last week, OpenAI announced its own custom inference processor, "Jalapeño," built with Broadcom. OpenAI claims Jalapeño delivers better performance-per-watt than competitor chips. Anthropic's Samsung talks look like a direct response — not just to match OpenAI's hardware ambitions, but to signal to investors and partners that Anthropic isn't falling behind in the vertical integration game. Nearly every major AI player now has custom silicon: Google's TPUs, Amazon's Trainium and Inferentia, and now OpenAI's Jalapeño.

What This Means for Anthropic's Hardware Stack

Anthropic's current compute strategy relies on a multi-cloud approach: Google Cloud TPUs, AWS Trainium, and Nvidia H100s/B200s. A custom Samsung chip won't replace those anytime soon. The company explicitly told TechCrunch that diversified hardware remains pivotal. This new chip is a hedge — a long-term bet that owning your silicon gives you leverage in negotiations and supply chain, even if the first generation is narrowly scoped or underwhelming.

For engineers, the key question is what the chip will actually accelerate: training, inference, or both? The lack of a decision suggests Anthropic is still running workload surveys and cost projections. Given the lead time for chip design (18–24 months), we won't see silicon until late 2027 or 2028. By then, Nvidia's next-generation architectures will be on the market, and OpenAI's Jalapeño will have real-world benchmarks.

If Anthropic can pull off a custom chip — even one that starts with vague specs — it buys itself leverage against Nvidia's pricing and availability, and that's a bet worth watching through 2027.

Source: Anthropic is discussing a new custom chip with Samsung
Domain: techcrunch.com