Anthropic weighing Samsung partnership to develop custom AI chips: U.S. media

Anthropic, the firm behind the AI service Claude, is reviewing a partnership with Samsung Electronics to develop its own custom AI chips, according to the U.S. outlet The Information on Thursday.

The two companies are in talks to manufacture custom AI chips, as the U.S.-based AI firm has reportedly entered the early stages of developing its own chips and has identified Samsung Electronics as a potential manufacturing partner.

The company is reportedly keen to leverage Samsung Foundry’s 2-nanometer process and advanced chip packaging capabilities.

Anthropic also hired Clive Chan, who led OpenAI's custom AI chip program, last month to bolster its semiconductor efforts. The company has yet to finalize the chip's specifications or performance targets.

Anthropic signaled its hardware ambitions in May when it announced that Samsung Electronics, SK hynix and Micron had joined its Series H funding round as investors. The AI company labeled three chipmakers as its "strategic infrastructure partners whose technologies play a critical role in the world's supply of memory, storage and logic chips.”

“As demand for Claude continues to grow, these relationships will help us scale our compute reliably at the pace our customers need,” Anthropic said.

Among the three, Samsung Electronics is the only company with a foundry business capable of manufacturing logic, or system, semiconductors.

Samsung Electronics is therefore the most likely manufacturer if Anthropic moves forward with mass-producing its custom AI chips, industry observers expect.

The competitive battleground among AI developers is increasingly shifting from AI models to semiconductors.

Google trained its Gemini AI models using its in-house Tensor Processing Units, or TPUs, last year. Amazon Web Services launched its homegrown AI chip, Trainium, in December of the same year.

This year, Microsoft commercialized its Maia AI chip in January, and Meta unveiled four AI chip prototypes in March. OpenAI, the operator of ChatGPT, showcased a prototype of its custom AI chip, Jalapeño, developed with Broadcom, on June 24.

Big Tech companies are racing to develop their own chips as Nvidia continues to dominate the GPU market.

Using scarce and costly GPUs to run AI models can sharply increase operating expenses. Custom chips optimized for a company's own AI models could help lower those costs.

The biggest challenge, however, is the cost of development.

"Developing proprietary AI chips is strategically the right move for AI companies, but chip design alone costs at least 500 billion won ($327 million),” an official at a Korean chipmaker said. “There is no guarantee of when the development will be completed.”

BY OH HYEON-WOO [[email protected]]