Samsung poised to make part of Google's next AI chip as TSMC nears capacity

Samsung Electronics' foundry business is expected to produce part of Google's next-generation AI chip, following its work on Nvidia's self-driving chip. With the AI boom pushing the world's top foundry, TSMC, to the limits of its capacity, expectations are growing that Samsung will benefit.

The Information, a U.S. technology news outlet, reported Thursday that Google is in talks with Samsung to produce a key component of its 10th-generation Tensor Processing Unit (TPU), codenamed Icefish. Icefish is Google's newest AI chip, designed with the Taiwanese chipmaker MediaTek and targeted for mass production in 2028. Under the leading plan, TSMC would produce the main processor that handles computation on its 1.4-nanometer process, while Samsung would make the input-output component that links the TPU to high-bandwidth memory (HBM) on its 2-nanometer process.

Industry watchers also see a strong chance of a turnkey arrangement covering memory, foundry and packaging, under which Samsung's memory division would supply the HBM, its foundry division would make the key component and Samsung would assemble them.

Samsung and other global foundries have recently benefited from a production bottleneck at TSMC. Surging demand for AI chips from big tech firms such as Google, Amazon and Meta has left TSMC's advanced 3-nanometer process so busy that it can barely take new orders. Google recently concluded that TSMC could not handle its entire TPU volume because of capacity limits and asked Intel Foundry to produce 3 million TPUs by 2028.

Samsung has also landed a chance to produce Nvidia's self-driving chip. Jun Young-hyun, vice chairman and head of Samsung Electronics' chip division, met Nvidia CEO Jensen Huang during Huang's visit to Korea on Monday.

"We are cooperating to produce Nvidia's self-driving chip and Groq's AI chip, and we are continuing to discuss next-generation production," Jun said afterward.

It was the first time Samsung had publicly confirmed it would produce Nvidia's self-driving chip. Groq, a fabless chip-design firm that Nvidia acquired last December, recently commissioned Samsung's foundry to produce inference chips.

Samsung will also produce Tesla's self-driving chips, the AI5 and AI6, starting this year, using its 2-nanometer process at its plant in Taylor, Texas, which is set to begin operating in the second half. Samsung signed a $16.5 billion next-generation AI chip foundry contract with Tesla last year.

The Information pointed out that large tech firms were diversifying their supply chains to Samsung Electronics, the world's No. 2 foundry. 

With foundry orders mounting, some expect Samsung's foundry to climb out of the red.

"The yields on Samsung foundry's 4-nanometer and 8-nanometer processes have already stabilized, and the yield on its advanced 2-nanometer process is reported to have recently risen into the 60 percent range," said Kim Yong-seok, a chair professor at Gachon University's semiconductor college.

"A foundation for securing profitability appears to be in place."

The professor said if Samsung can raise its 2-nanometer yield it can swing into profit as early as the third quarter. 

"The key for Samsung is to raise its 2-nanometer yield," Kim said. "With TSMC's 2-nanometer process saturated, there is a strong chance of the company swinging to a profit as early as the third quarter if Samsung lands additional contracts with big tech."

Samsung itself is more cautious about the timing. "Even if the business turns profitable next year, the accounts will stay in the red once special performance bonuses and the like are factored in," Han Jin-man, president of Samsung's foundry business, said at an internal employee briefing Friday. "A clear profit will be achievable in 2028."

BY KIM KYUNG-MI [[email protected]]