$XIAOMI-W(01810)$  SK Hynix to Cut HBM4 Volumes This Year... Expanding HBM3E and Other Products Instead

SK Hynix is reportedly pursuing a plan to reduce its 6th-generation High Bandwidth Memory (HBM) shipments to NVIDIA by approximately 20–30% compared to its original plan. This is understood to be a consequence of NVIDIA's next-generation AI accelerator "Vera Rubin" facing difficulties in scaling mass production.

However, the reduced HBM4 volumes from SK Hynix are expected to be substituted by demand for the previous-generation HBM3E and server DRAM. Given that margin profiles differ across products, the industry consensus is that the impact on this year's business results remains to be seen.

According to ZDNet Korea's reporting on April 14, SK Hynix plans to reallocate a portion of its NVIDIA-bound HBM4 shipment allocation for this year to HBM3E and server DRAM volumes.

HBM4 is the latest HBM entering full-scale commercialization this year. It will be first adopted in "Vera Rubin," the AI semiconductor that global big tech company NVIDIA is targeting for launch in the second half of this year. Accordingly, major memory companies including Samsung Electronics, SK Hynix, and Micron are all going all-out to supply HBM4 to NVIDIA.

However, the industry expects that Rubin series shipment volumes this year will fall short of the original plan.

This is based on analysis that the optimization of various components constituting the Rubin platform has not yet been fully completed. Notably, NVIDIA had requested HBM4 data processing speeds significantly above the industry standard, at the 11 Gbps level.

In a recently published report, market research firm TrendForce stated: "Beyond the time required for HBM4 validation, several challenges remain, including the transition to network interconnects, increased power consumption, and optimization of more advanced liquid cooling solutions." It added: "As a result, the Rubin series' share of NVIDIA's high-performance GPU shipments is expected to decline from the original 29% to 22%."

Conversely, the shipment share of the "Blackwell" series—NVIDIA's current primary production focus—is projected to increase significantly from 61% to 71%. Blackwell uses HBM3E.

The domestic memory industry is also reportedly adjusting its HBM business strategies accordingly. SK Hynix in particular is seeing the largest degree of business variability, as the dominant player in the HBM market with the largest shipment share for both NVIDIA-bound HBM4 and HBM3E.

A source familiar with the matter explained: "As the volume of Rubin series shipments this year is being reduced, adjustments to SK Hynix's HBM4 shipment plan have become unavoidable. However, since that volume is being redirected to HBM3E and server LPDDR (low-power DRAM), total memory demand is not decreasing."

SK Hynix had originally planned to ship approximately 6 billion Gb (gigabits) of HBM4 to NVIDIA this year. The level currently under discussion is 20–30% below that figure. As some of the reduced volume is being converted to Blackwell series use, HBM3E volumes are also expected to exceed the original forecast of 8 billion Gb.

Another source said: "SK Hynix is internally discussing plans to convert a portion of HBM4 volumes to HBM3E and server LPDDR. In practice, procurement orders for materials and components needed for HBM4 mass production have also been increasing more slowly than initially expected."

From jukan