NVIDIA's Huang Sees No 2029 Memory Glut, Cites HBM and LPDDR Demand

NVIDIA CEO Jensen Huang said on October 22 that he does not expect a memory glut, or oversupply, in 2029 and that demand for high-bandwidth memory (HBM) and LPDDR will keep growing with AI data centers, according to Korean outlets JKN and SBS News.

Huang spoke on the sidelines of NVIDIA's GTC conference in Berlin, in an interview with CNBC, SBS reported. Asked about a possible glut in 2029, he answered "I don't think so," and said that for memory, more is always better. He added that HBM will stay in use for years and that NVIDIA plans to keep buying large volumes of LPDDR, the low-power memory used in phones.

How Huang Answered Korea's 2029 Oversupply Forecasts

The remarks read as a rebuttal to South Korean forecasts that new supply in 2029 would saturate the market and end the current memory upcycle, SBS said. JKN reported that memory stocks had recently fallen on fears of oversupply and cooling AI demand, and that such an outcome could hit pricing and profits at SK hynix, Samsung Electronics and Micron.

Huang argued that the memory market should be read by product group. In his view, HBM and high-end server memory can follow different supply and demand conditions from memory for consumer devices, JKN said. HBM sells at higher prices and adds more value than standard DRAM, and its share is rising with demand for AI chips, the outlet added.

How HBM and LPDDR Differ in AI Servers

High-bandwidth memory (HBM) and LPDDR both feed AI chips, in different ways. HBM stacks DRAM dies beside the processor and widens the link: JEDEC's HBM4 standard uses a 2,048-bit interface for up to 2 TB/s per stack. LPDDR is the low-power memory made for phones, which Huang said NVIDIA now uses in data-center products at twice the energy efficiency of existing ones, a vendor claim.

NVIDIA is not alone in using it. Intel's GPU D270 uses LPDDR5X instead of HBM, with 160 GB on Intel's own card, and Qualcomm's AI200 carries 768 GB of LPDDR per card. Memory makers therefore sell into two demand streams, which fits Huang's view that the market should be read by product group.

What It Means for HBM Supply and Capex

For data-center planners, the point is who is buying. SBS calls NVIDIA the world's largest memory-purchasing company. NVIDIA buys HBM, which sits beside the GPU in the package, and has said it will keep using large volumes of LPDDR, which JKN says it now places in data-center products. Huang claimed NVIDIA was the first to do this; JKN did not report independent figures.

A long demand runway would support the capacity additions already under way at the Korean makers. Voltgrid has covered Samsung's restart of Fab 5 at Pyeongtaek to chase SK hynix in HBM, and NVIDIA's reported move to put all three HBM suppliers on TSMC 3nm base dies. Voltgrid also reported Huang's remarks on neocloud demand and scarce capital.

Asked whether NVIDIA would use memory from China's ChangXin Memory Technologies (CXMT), Huang said "of course, we have an open mind," according to SBS. JKN noted that CXMT has unveiled a new design structure that raised investor concern about its reach in general-purpose DRAM, while it still faces yield and advanced-packaging hurdles in making high-end HBM at volume.

Huang also said US restrictions have sped up China's push for chip self-sufficiency, and that US technology companies can compete with any rival.

Huang gave no supply, price or volume figures, and no purchase or partnership with CXMT was announced. The 2029 glut scenario is a forecast from South Korean experts, not a measured surplus, and Huang's view is an opinion from the largest memory buyer, as SBS describes NVIDIA, rather than a supply-side projection. Watch for HBM and DRAM contract-price commentary from SK hynix, Samsung and Micron in their coming results, and for any NVIDIA supplier decision on CXMT.