Jensen Huang says no 2029 memory glut. The HBM numbers say 2027

Jensen Huang spent Thursday telling investors the memory cycle has years left in it. The strongest evidence that he is partly right is also the strongest evidence that the bill arrives well before 2029: chipmakers are already redesigning accelerators to use less high-bandwidth memory.

What he actually said

Asked about the memory selloff on the sidelines of Nvidia's GTC Berlin conference, Huang gave CNBC's Arjun Kharpal the sentence that will get quoted: "One of the things that's really great about memory is that you can always use more of it," he told CNBC. He said HBM will be used for years to come, and claimed Nvidia as the first company to put LPDDR — the memory that goes in phones — into data centres, which he said delivers performance "twice as energy efficient as anything in the world today." His conclusion: "we're going to continue to use abundant amounts."

He was also asked about CXMT, the Chinese memory maker whose new cell design helped trigger the selloff, and said Nvidia is "open-minded" about buying from it: "We work with Chinese companies; just about the whole world works with Chinese companies in the supply chain somewhere." On export controls he was blunter than any Washington policy paper — the restrictions "definitely accelerated" China's push into semiconductor manufacturing.

The selloff he was answering started with an opinion column. On October 15 a major South Korean publication urged the government to brace for the AI boom slowing, citing two experts who expect "a flood of new supply" to turn the memory market "into a red ocean" in 2029 — "the supercycle could be over, and the industry could enter a downturn," they wrote. CXMT then supplied the technical half of the fear: server DRAM built on a 4F² cell that cuts cell area by about a third against the 6F² design the industry runs on, which we worked through when CXMT forecast $19.5B of nine-month profit and its shares still fell 6.5%. Denser cells mean more bits per wafer, and more bits per wafer is what a supply flood is made of.

The squeeze is a 2027 problem

Set the 2029 argument against what the price data says about the next four quarters and the two do not describe the same market. TrendForce raised its 2027 HBM price outlook to a 121% year-on-year increase in blended average selling price, and expects HBM supply to stay constrained through 2027 as HBM4 takes a larger share of the mix and HBM4e ramps in the second half. For the current quarter it projected conventional DRAM contract prices up 10–15% quarter over quarter, with server DRAM still undersupplied and PC DRAM supply possibly falling in 2027 as fabs shift wafers to servers.

The spending behind those numbers is the part the 2029 column glosses over. TrendForce's August estimate has cloud capex up 98% year over year in 2026 and another 50% in 2027, with memory rising from 47% of that spend this year to 68% next year, and server DRAM contract prices up roughly 270% across 2026 after a 64% gain in the second half of 2025. Its forecast for 2027 combined server DRAM and HBM bit supply is a 27% increase — real capacity growth, but smaller than the capex number it is meant to serve.

The tell is 8-Hi

The detail worth more than any forecast sits inside TrendForce's HBM note. GPU and ASIC vendors are now evaluating 8-Hi HBM stacks instead of 12-Hi ones — eight DRAM dies per stack rather than twelve — not because demand fell but because supply is short and system cost is climbing. Fewer DRAM dies per stack lowers the bill of materials and lets the same HBM bit output serve more accelerators. (If stack layers are new ground, What is HBM (high-bandwidth memory)? covers the architecture in plain English.)

That is what a shortage looks like from the inside: customers redesigning silicon around the price of a component. And the workaround does not even deliver a cheaper gigabit. Each stack needs a base die, and that cost does not shrink with the layer count, so an 8-Hi stack spreads it over less capacity — TrendForce expects 8-Hi to carry a 10–20% per-gigabit premium over 12-Hi in 2027. The saving is per GPU, not per bit. If you want a single number that argues against a 2029 glut, it is that one: the industry is paying more per gigabyte to get more gigabytes.

Who wins, who loses

Samsung is the clearest winner and knows it. Preliminary third-quarter guidance puts operating profit at 107.4 trillion won, up 782.5% year on year, on sales of 195 trillion won, and its executives have said HBM will consume close to 30% of DRAM makers' wafer capacity in 2027 — the mechanism that keeps commodity DRAM tight while HBM volumes grow. Margins like that fund the fabs that the 2029 column is worried about.

CXMT is the wildcard, and the honest read is that its threat is to the commodity end, not the HBM end. Counterpoint put it at 10% of global DRAM revenue in the third quarter, double a year earlier, and its 4F² cell is a cost-per-bit claim rather than a bandwidth one; US export controls still keep the leading HBM out of its reach. A genuine flood would come from Chinese commodity DRAM, and it would hit DDR5 and LPDDR margins before it touched HBM pricing.

The losers are downstream and already visible: PC makers carrying higher component costs into notebooks, phone makers whose margins absorb mobile DRAM allocation, and every server buyer negotiating against 270% cumulative increases. That squeeze is exactly the kind of thing that shows up in hardware earnings two quarters later — the same repricing we watched run through TSMC's record quarter and the repriced order book last week.

What the skeptics get right

Huang is the world's largest HBM buyer answering a question about whether HBM will be scarce. A glut would suit Nvidia's cost of goods; what he is actually defending is the demand curve, not the price. Treat "no glut" as a statement about bits consumed, not dollars per bit, and it survives contact with the evidence.

The bear case is not wrong either. Memory cycles end the way the column describes: high prices invite capex, capex lands in a lump, and the lump is what breaks the price. The capacity funded in 2026 and 2027 arrives in 2028 and 2029. What the column gets wrong is the tense. Memory equities turn twelve to eighteen months before prices do, which is why Samsung and SK Hynix fell 17.8% and 24.3% in a single week on a newspaper opinion piece, per Seoul Economic Daily's tally — the market was not pricing next quarter, it was pricing the turn after it. The 2029 glut and the 2027 squeeze can both be true. Only one of them is a fact today.

What to watch

Fourth-quarter contract settlements in January, to see whether the 10–15% holds. The 2027 HBM4 annual price negotiations, where suppliers are pushing for increases in the several-hundred-percent range. Whether CXMT's 4F² server DRAM actually ships this year as announced. Samsung's October 29 results, which break memory out from the handset and appliance divisions that are being squeezed by the same prices. And whether 8-Hi shows up in shipping configurations rather than evaluation — if chipmakers cut HBM content per device at scale, HBM bit demand takes a real hit, and that is the glut arriving early, years before anyone's 2029 model.

If the memory bill lands in your hardware budget before it lands in the spot price, how are you planning for it? Write to us via our contact page.