SK Hynix has reportedly secured customized high-bandwidth memory (HBM) orders from Nvidia, Microsoft, and Broadcom, pulling ahead of domestic rival Samsung Electronics in the race to supply AI-optimized memory solutions, according to reports by The Korea Economic Daily and Hankyung.
The South Korean chipmaker has received certification from all three tech giants to supply bespoke HBM solutions tailored to specific workload and system requirements. Industry observers now widely expect customized HBM to gradually displace off-the-shelf versions, becoming the dominant format in a global market projected to reach US$130 billion by 2033.
Tailored HBM targets AI-era needs
SK Hynix has already begun designing customer-specific HBM chips in collaboration with Nvidia, Microsoft, and Broadcom—marking roughly 10 months since the company received its first customized HBM inquiries from members of the US "Magnificent Seven" tech group, which includes Apple, Microsoft, Google, Amazon, Nvidia, Meta, and Tesla.
The first commercial shipments of customized HBM are expected in the second half of 2025. SK Hynix has previously stated that from the HBM4E generation onward, it will fully pivot to a custom production model. Current mainstream adoption centers around fifth-generation HBM3E, with the industry preparing to shift toward sixth-generation HBM4 in the near term.
Customized HBM is designed to meet increasingly complex performance and power-efficiency requirements as hyperscalers shift away from generic memory toward specialized modules optimized for AI workloads. To support this transition, SK Hynix has partnered with TSMC to manufacture the logic die, which is often described as the "brain" of an HBM stack. While SK Hynix previously produced the logic die in-house for HBM3E and earlier models, the move to HBM4 requires more advanced process nodes, making outsourcing to TSMC a more practical solution.
Capacity ramp-up and competitive response
SK Hynix has transformed its M15X fabrication plant in Cheongju, which was previously dedicated to NAND flash production, into a facility focused on manufacturing advanced DRAM and HBM. The company intends to invest KRW20 trillion (US$14.5 billion) in the site to address increasing market demand.
Meanwhile, Samsung Electronics is engaged in active negotiations with several clients for customized HBM supply agreements. After recently supplying HBM3E chips to AMD, the second-largest AI accelerator provider globally, industry analysts anticipate that Samsung will soon secure orders for HBM4 and other custom versions. Reports indicate ongoing discussions with both Broadcom and AMD regarding next-generation HBM supplies.
Article edited by Jerry Chen
