Anthropic is in talks with Samsung to co-develop a custom AI chip, according to a Thursday report from The Information. The discussions extend a thread first surfaced by Reuters in April, when Anthropic was reported to be weighing its own silicon as a hedge against chip shortages. Neither the chip's purpose, its slot in the server stack, nor its performance target has been decided.
Anthropic confirmed the strategic direction without confirming the Samsung tie-up. Asked about the report, the company told TechCrunch that its compute strategy will continue to rely on a mix of processors from Google, Amazon, and Nvidia, and declined to add anything specific about Samsung.
The move mirrors a broader push across the frontier labs to reduce dependence on Nvidia, which still dominates AI training silicon. Custom chips give a lab two things at once: hardware tuned for its specific model architectures and inference workloads, and a negotiating counterweight against a single dominant supplier.
Key facts
- 01Anthropic is in talks with Samsung about a custom AI chip, per a Thursday report from The Information.
- 02The chip's purpose, server integration, and performance profile remain undecided.
- 03OpenAI announced its own Broadcom-built inference chip, dubbed Jalapeño, last week.
- 04Samsung is already an Nvidia manufacturing partner and is building an AI chip factory in South Korea with Nvidia.
- 05Anthropic said Google, Amazon, and Nvidia chips will remain pivotal to its compute strategy.
The timing is notable. Last week, OpenAI unveiled its own custom inference chip, built with Broadcom and internally named Jalapeño. OpenAI claims the chip offers better performance-per-watt than competing silicon. Anthropic's Samsung talks landing days later reads less like coincidence and more like the same competitive clock ticking on both companies.
“a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal to its compute strategy”— Anthropic, company statement
Samsung is not a newcomer to this market. It is one of Nvidia's key manufacturing partners, producing chips that Nvidia sells into AI training clusters. Samsung in turn uses Nvidia software to run its fabs, and the two are jointly building an AI chip factory in South Korea. Samsung has also held partnership discussions with Google on chip-making. In short, Samsung already sits at the center of the AI hardware supply chain — it just hasn't yet had a frontier-lab customer of its own.
For Anthropic, the calculation is straightforward. Frontier model training and inference costs are dominated by compute, and every major competitor is either building its own accelerator (Google with TPUs, Amazon with Trainium and Inferentia, OpenAI now with Jalapeño) or has one on the roadmap. Continuing to buy the majority of compute from Nvidia — at Nvidia's margins — becomes an increasingly expensive position over time.
It's also a hedge against supply. The April Reuters report framed Anthropic's chip interest as a response to shortage risk, and that risk hasn't gone away. Nvidia's top-tier accelerators remain allocation-constrained, and the labs with the most compute win the model race. A Samsung-manufactured chip, even one narrowly targeted at inference, would give Anthropic a second sourcing option outside the Nvidia-TSMC axis that currently supplies most frontier silicon.
The caveats are real. Anthropic hasn't decided what the chip does, which suggests the project is early — closer to a scoping exercise than a taped-out design. Custom silicon programs routinely take three to five years from first talks to production deployment, and the AI landscape at that horizon is not something anyone is predicting with confidence. OpenAI's Jalapeño, by contrast, is already being described in performance-per-watt terms, which implies real hardware. Anthropic is behind on this specific curve.
There's also the question of what Samsung brings that Broadcom or TSMC-plus-a-design-partner wouldn't. Samsung's foundry has trailed TSMC on leading-edge process nodes, and Nvidia's most advanced chips are made at TSMC. If Anthropic wants frontier-node performance, Samsung is a harder path than the industry default. The choice may say more about capacity availability and commercial terms than about pure technical fit.
The Anthropic-Samsung talks fit a pattern the site has been tracking for months: frontier labs vertically integrating away from a single-vendor compute stack. OpenAI has Broadcom. Google has its TPUs. Amazon has Trainium. Anthropic, uniquely, has been the frontier lab most publicly committed to buying rather than building — and that position is now shifting. If the Samsung deal materializes, the story of the next two years is no longer whether the labs build their own chips, but which foundry partner each of them picks, and how quickly Nvidia's share of frontier training compute starts to compress.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
