Anthropic in Talks With Samsung to Manufacture Custom AI Chip
The Takeaway
- Anthropic explores custom AI chip development with Samsung.
- AI developer seeks greater control over costly computing infrastructure.
- Samsung eyes high-profile win for its advanced foundry business.
Anthropic has begun early-stage work on its own AI chip and held talks with Samsung Electronics as a potential manufacturing partner, following the lead of rival OpenAI in trying to get more control over the costly computing systems behind its models, according to three people with direct knowledge of the project.
If the Claude maker goes ahead with the chip, it would be a relative newcomer in developing its own AI server chips, compared to others. Google and Amazon Web Services have both been successful at developing their own chips over many years, while Meta Platforms and Microsoft have also developed their own chips. OpenAI tapped Broadcom to design its own chip in 2024 and last month unveiled the first product of that partnership, Jalapeño, an inference chip built to run large-language models more efficiently.
Anthropic is still at the stage of figuring out what it wants the processor to do, how powerful it should be and how it would fit into a server or clusters of servers, according to the three people. It has held discussions with multiple chip design firms but it has yet to move into detailed design, testing and manufacturing work.
Such processors are hard to design because engineers must balance speed, power consumption, memory, networking and cooling all at once, and manufacturing reliably at high volume is even harder.
The project underscores a broader shift where AI developers such as Anthropic are looking to control the infrastructure behind their models, including chips, cloud contracts, power and data centers. The biggest AI models run on vast clusters of processors, and at that scale even small efficiency gains can lower costs and free up scarce computing capacity. An in-house AI chip would add leverage as AI developers compete for processors, data-center space and electricity.
It’s possible Anthropic may not proceed with the chip, although the company has begun hiring for chip engineers. Earlier this month, it hired Clive Chan, an early member of OpenAI’s custom chip team.
In response to questions from The Information, Anthropic said Amazon Web Services’s Trainium chip, Google tensor processing units and Nvidia graphic processors will remain central to how the company scales its compute strategy. The company declined to share more about its roadmap. Samsung declined to comment.
Samsung is no stranger to Anthropic. The company, one of the world’s leading memory suppliers, invested in Anthropic’s $65 billion fundraising in May alongside the two other main memory suppliers, SK Hynix and Micron Technology. The investments, happening at a time when the memory chip supply crunch has led to Apple and other consumer electronics makers to hike prices, can tie Anthropic closely to some of the chip suppliers needed for its growth.
South Korea recently announced a decade-long investment plan led by Samsung Group and SK Group, parent companies of Samsung Electronics and SK Hynix, with the two investing a combined $518 billion to build four memory-chip plants in the country.
A deal with Anthropic, if finalized, would mark a high-profile win for Samsung’s foundry business. A leading manufacturer in memory chips, the South Korean company has also been trying to expand its AI chip business and close the gap with Taiwan Semiconductor Manufacturing Co., whose advanced production lines remain the industry’s gold standard for cutting-edge AI processors. As demand for AI chips has overwhelmed the Taiwanese firm, Samsung has found an opening to pitch its 2-nanometer technology to more customers. Google is another customer considering using Samsung for part of a future tensor processing unit, The Information reported.
Anthropic is considering using Samsung’s 2-nanometer manufacturing process and advanced packaging facilities, two of three people said. The 2-nm label, an industry shorthand for one of the most advanced chipmaking technologies rather than a physical measurement, is designed to make processors denser and more power-efficient. The advanced packaging technique places the main processor closer to high-speed memory so data can move quickly inside graphics processing units.
Anthropic has long differentiated itself from other AI developers by renting a variety of server chips to develop and run its technology so it isn’t dependent on Nvidia’s hardware as OpenAI and xAI are. It uses AI server chips from Amazon and Google as well as Nvidia. Anthropic is also discussing using chips from Microsoft and from a young U.K.-based startup, Fractile.
Despite the frenzy of financing and design activity in the inference chip market, Nvidia has actually increased its market share in recent years to 74%, according to estimates from The Information. Nvidia CEO Jensen Huang insists the company’s chips can handle inference more effectively than any alternative.