SK Hynix Inc. has secured Nvidia Corp., Microsoft Corp. and Broadcom Inc. as key clients of its customized high-bandwidth memory (HBM), positioning the South Korean chipmaker at the forefront of both tailor-made and general-purpose AI memory markets.
According to industry sources on Thursday, SK Hynix has recently begun designing HBM tailored to individual client specifications. It has selected its clients based on the delivery schedule for its largest customer Nvidia.
“Given SK Hynix’s production capacity and the launch timelines for AI services by big tech firms, it’s not feasible to accommodate all requests from the Magnificent Seven,” said a semiconductor industry official.
“But the company is likely to add a few more clients, depending on shifts in HBM market conditions,” he told The Korea Economic Daily.
The move comes about 10 months after the company made public it had received requests for customized HBM from members of the US “Magnificent Seven,” which includes Apple, Microsoft, Google, Amazon, Nvidia, Meta and Tesla.
Its first customized HBM — expected to be the seventh generation HBM4E — is likely to roll out in the second half of next year, said the sources.
Currently, fifth-generation HBM3E is the most advanced AI chip in mass production, with a transition to sixth-generation HBM4 expected in the near future.
SAMSUNG IN SUPPLY DISCUSSION FOR HBM4
Samsung is said to be in advanced talks with Broadcom and Advanced Micro Devices Inc. (AMD) to supply customized HBM4, a generation behind HBM4E, according to industry sources.
During an earnings call in April, the world’s largest memory chipmaker projected that it could begin supplying HBM4 to its clients as early as the first half of 2026.
Samsung is lagging behind its crosstown rival SK Hynix in the HBM race, having failed to secure supply deals for its advanced memory chips with Nvidia.
SK Hynix, backed by the orders for tailor-made AI chips from Nvidia, Microsoft and Broadcom, is also set to lead the customized HBM market, projected to expand to $130 billion by 2033 from $18.2 billion in 2024, according to data from TrendForce and Bloomberg Intelligence.
Customized HBM is designed to meet the specific requirements of individual clients. They are in growing demand as Big Tech shifts away from general-purpose memory to optimize the performance of their AI services.
Taiwan Semiconductor Manufacturing Co. (TSMC) will make logic dies for SK Hynix’s customized HBM. A logic die plays the role of the brain in HBM.
Until HBM3E, SK Hynix had manufactured logic dies in-house. But beginning with HBM4, it turned to TSMC as the advanced chip requires ultra-fine process technology.
MASS PRODUCTION OF HBM4 EXPECTED IN H2, 2025
Early this year, SK Hynix became the world’s first chipmaker that delivered a sample of HBM4 to Nvidia. Industry observers expect SK Hynix to begin mass production of HBM4 in the second half of this year.
The company controls half of the global HBM market, followed by Samsung and Micron Technology Inc. with market shares of 30% and 20%, respectively, according to Taiwan-based TrendForce.
To ramp up HBM output, SK Hynix has converted its M15X plant in Cheongju, North Chungcheong Province, originally designated for NAND flash production, into a facility for advanced DRAMs and HBM with plans to invest 20 trillion won ($14.5 billion).
By Chae-Yeon Kim
why29@hankyung.com
Yeonhee Kim edited this article.
