Jensen Huang Says a Memory Glut by 2029 Is Unlikely as the AI Buildout Keeps Swallowing HBM and LPDDR

NVIDIA CEO Jensen Huang has entered the fight over the memory market's direction, telling CNBC during the GTC Berlin conference on October 22 that a supply glut by 2029 is unlikely. "One of the really nice things about memory is you can always use more," Huang said, adding that HBM will "continue to see adoption for many years to come."

The comments are a direct response to a market scare that began with an October 15 column in a Korean business daily, in which experts argued that a wave of new fab capacity could push the memory industry into a "red ocean" by 2029 and that the supercycle may already be ending. The piece hit a nerve: storage stocks including SanDisk, SK Hynix and Micron sold off sharply in the following week, dragging the broader memory complex down with them.

Huang's rebuttal rests on demand rather than supply. NVIDIA's own products are among the largest memory consumers on earth — the company notes it was the first to bring LPDDR, originally a mobile memory standard, into the data center, claiming roughly double the energy efficiency of any alternative approach. As AI systems scale from chips to racks to entire factories, HBM, LPDDR and conventional server DRAM all get absorbed into a demand curve that Huang argues has no natural ceiling.

Third-party supply data roughly supports his timeline. TrendForce estimates the 2026 DRAM supply-demand gap at minus 1 to 2 percent, widening in 2027 as demand growth outpaces supply additions. Most new fabs that have been announced will not contribute meaningful output until 2028, because construction, equipment moves and production cycles all lag — and HBM consumes so many wafer starts per usable bit that headline capacity expansions translate into far less effective supply than buyers expect.

Prices tell the same story. Server DRAM contract prices rose about 64 percent in the second half of 2025, with TrendForce projecting cumulative growth of roughly 270 percent for 2026. HBM contract prices could climb another 70 to 140 percent in 2027. Combined DRAM and NAND spending is expected to jump from 47 percent of cloud providers' capital expenditure in 2026 to 68 percent in 2027 — memory is no longer a component line item but a defining constraint on AI infrastructure budgets.

Huang is, of course, an interested party: memory cost and availability directly shape how many accelerators his customers can deploy, and talk of a 2029 glut threatens to spook the same investors funding the buildout. But the counterargument to the glut thesis is structural rather than promotional. Even if every announced fab comes online on schedule, the industry's own research suggests demand growth through 2027 will outrun it — the real test arrives in 2028, when the new capacity finally lands and the AI buildout's appetite gets its first honest stress test.