The conversation is happening in Samsung’s foundry division, and it is the kind of meeting Nvidia used to win by default. Anthropic, the artificial intelligence company behind the Claude models, is in talks with Samsung Electronics about manufacturing custom AI chips, according to people familiar with the discussions, in a move designed to reduce its dependence on Nvidia’s processors and cut the cost of running its models.
The discussions have focused on Samsung’s advanced manufacturing processes, including its 3-nanometer node and the generations that follow it, the people said. A deal would make Samsung one of the few foundries supplying custom chips to a leading AI lab, a market that has been dominated by Taiwan Semiconductor Manufacturing Co. and, until recently, by Nvidia’s standard designs.
The talks come a week after OpenAI announced its first custom chip, a processor developed with Broadcom and given the code name Jalapeño. The two announcements, taken together, mark the end of an era in which AI labs bought their computing power off Nvidia’s shelf, and the beginning of a period in which the biggest buyers of chips are designing their own.
The economics explain the shift. Nvidia’s GPUs are expensive, scarce and built for the broadest possible market, and AI companies have concluded that chips tailored to their specific workloads can run inference at a fraction of the cost. Inference, the process of generating answers from a trained model, is where Anthropic spends most of its computing budget, and every percentage point of efficiency improvement translates into real money at the scale it operates.
Anthropic has said publicly that it works with chipmakers on co-designed hardware, and the Samsung talks would deepen that strategy from collaboration to full custom silicon. A chip built on Samsung’s 3-nanometer process would be designed around the mathematical operations that Claude performs most often, trading flexibility for speed and efficiency in the way that specialized processors have done in every computing era.
Samsung needs the business. The company’s foundry unit has struggled to match TSMC’s share of advanced manufacturing, and it has won few of the high-profile AI chip contracts that have flowed to its Taiwanese rival. A partnership with one of the world’s most valuable AI companies would be a statement win, a sign that Samsung can serve the customers defining the next generation of computing rather than the last one.
Cost is the pressure behind the talks. Anthropic’s spending on computing has climbed with the scale of its models, and executives have said inference efficiency is one of the few places where the company can improve its margins without changing what it sells. Custom silicon is the most direct path to that efficiency, which is why the company has been willing to take on the complexity of designing its own chips rather than accepting the bill from Nvidia.
The technical hurdles are substantial. Designing a custom chip takes years and hundreds of engineers, and Samsung’s process technology, while advanced, has a mixed record of delivering the yields that high-volume customers expect. People familiar with the talks said the two sides are still early, and that no final design or production schedule has been set.
The strategic stakes extend beyond the two companies. Nvidia has built its dominance on the combination of hardware and software, and its CUDA platform has made its chips the default choice for AI developers. Custom chips from Anthropic, OpenAI and others attack that position from the hardware side, giving the AI labs alternatives they can tune to their own needs, though they still face the software moat Nvidia has spent a decade building.
The industry has seen this pattern before. Google built its own tensor processors to cut the cost of running its AI workloads, Amazon developed its Trainium chips for its cloud customers, and Microsoft designed the Maia accelerators for its data centers. Each project began as a cost-saving measure and grew into a strategic asset; Anthropic’s talks with Samsung suggest the AI labs are following the same path, only faster, because the cost pressure is sharper.
Anthropic has options beyond Samsung. The company has relationships with other chipmakers and cloud providers that could serve as manufacturing partners, and a final decision has not been made, the people said. But the Samsung talks are the furthest along, and both companies have reason to want a deal announced sooner rather than later, one to secure capacity, the other to secure credibility.
Foundries are watching closely. TSMC’s order books are full, and it has told customers it can allocate only so much capacity to any single one, a constraint that has pushed AI companies to look elsewhere. Samsung’s willingness to take on custom work, and to invest in the process improvements that work requires, could redraw the map of advanced chip manufacturing over the next few years.
For Anthropic, the prize is control: control over its costs, its supply chain and the pace at which its models improve. For Samsung, the prize is relevance in the fastest-growing corner of the chip business. The talks are early, the outcome uncertain, and either side could walk away. But the direction of travel is clear, and it runs away from the standard GPU.
