Why Nvidia's Huang Is Betting Against a Memory Bust

The Signal Behind the Statement

Jensen Huang did not just answer a question at GTC Berlin on October 22. He walked into an active market narrative and tried to derail it.

The narrative was simple and brutal: after years of soaring memory prices driven by AI demand, the semiconductor supercycle would hit a wall around 2029. Supply would outpace demand. Prices would collapse. Samsung Electronics and SK Hynix — the two Korean giants that have ridden this cycle to record profits — would feel it first.

Huang’s response was blunt. “One of the biggest advantages of memory is you can always use more of it,” he told CNBC. When asked directly about the 2029 oversupply concern, he said demand would continue expanding sufficiently. He singled out HBM — the vertically stacked, high-bandwidth memory that sits inside every flagship AI accelerator — and said it would remain in use for years. He also highlighted LPDDR, the low-power memory type traditionally associated with smartphones, calling Nvidia the first company to deploy it at scale in data centers, achieving double the energy efficiency of any competing product.

The subtext: Nvidia is not pricing its next generation of chips around a memory crunch ending in three years. It is pricing them around memory demand growing for the foreseeable future. That is a directional bet with real consequences for how the market values the companies supplying that memory.

Who Is This Aimed At?

The timing matters. South Korean memory stocks had already dipped on the back of the oversupply thesis and comments from Elon Musk, who said in September that China would produce enormous volumes of RAM by 2028. The combined effect pressed on Samsung and SK Hynix shares and contributed to broader sector anxiety heading into fourth-quarter earnings season.

Huang’s remarks were not addressed to Korean manufacturers specifically. But they landed directly in their neighborhood. If Nvidia’s chief engineer — the single most influential buyer in the high-end memory market — says demand will absorb whatever supply comes online, that is a powerful counterweight to the bear case. It does not erase the structural risk of overcapacity. But it shifts the burden of proof.

The implication for investors is straightforward: the memory supercycle thesis is not settled. Huang is effectively placing Nvidia’s name behind the continuation of AI capital expenditure well beyond 2027, which is when most of the current buildout is expected to wind down.

The LPDDR Move Is the Quiet One

Most of the coverage will focus on HBM. But Huang’s comments about LPDDR in data centers deserve attention.

LPDDR has always been about efficiency, not raw bandwidth. It runs at lower voltages and is optimized for mobile devices where power matters more than throughput. Putting it in a data center is unusual — and it is exactly the kind of architectural shift that defines Nvidia’s strategy. Instead of asking memory suppliers to keep building more HBM at higher prices, Nvidia is designing systems that reduce the amount of memory power each operation consumes.

That is a demand-side argument. More efficient memory means more memory per watt, which means more memory per rack, which means more memory per data center. Huang framed it as a competitive advantage — “double the energy efficiency of any product in the world” — but it also functions as a demand amplifier. If LPDDR can do what Huang claims, the total addressable memory market expands beyond what HBM-only models predict.

This is not a near-term revenue driver for Samsung or SK Hynix in the same way HBM is. Both companies are heavily concentrated on HBM for AI. But if Nvidia’s data-center LPDDR strategy scales, it opens a secondary demand channel that could matter in the 2028–2030 window when the current HBM cycle arguably matures.

The China Factor Huang Does Not Want You to Ignore

The interview took place against a backdrop that makes Huang’s comments about memory demand even more consequential: China’s semiconductor ambitions are accelerating, and the United States is partly responsible.

Huang did not shy away from this. He said American export controls on advanced chips have arguably sped up China’s drive for self-sufficiency. “China has clearly increased its development pace because of all the regulations,” he said. He praised China’s manufacturing capabilities and infrastructure as unmatched globally. He also refused to rule out working with Chinese memory makers, saying Nvidia holds an “open attitude” and that virtually every supply chain in the world already touches Chinese companies somewhere.

ChangXin Memory Technologies (CXMT), China’s leading DRAM producer, recently announced a new design structure aimed at improving production efficiency. That announcement fed directly into the oversupply anxiety — if Chinese memory output ramps faster than expected, global supply could flood the market sooner than 2029.

Huang’s overall posture is calibrated carefully. He is acknowledging China’s rise without endorsing it. He is keeping the door open for cooperation without committing to anything. He is defending American competitiveness while admitting that isolationist policy has backfired in this specific domain.

For the market, the takeaway is that the memory oversupply risk is not purely a demand problem. It is a supply-side problem shaped by geopolitics. US policy pushed Chinese memory capacity forward. Nvidia’s demand thesis depends on that capacity either staying constrained or being absorbed by growing AI workloads. Huang is betting on the second outcome.

What Happens Next

The 2029 oversupply thesis will not disappear because Huang said it might be wrong. It will be tested by three things: Nvidia’s own chip shipment guidance, actual HBM demand from cloud providers, and the pace at which CXMT and other Chinese producers can scale production.

Samsung and SK Hynix are the closest proxies. Their fourth-quarter results and forward guidance will signal whether the Korean memory industry is preparing for a soft landing or a harder correction than expected. If Huang’s demand narrative holds, memory stocks that sold off on the oversupply fear could see a re-rating. If it does not, the correction may simply be delayed, not avoided.

Musk’s 2028 prediction and Huang’s counter-assurance represent two different timelines for the same risk. One says Chinese supply hits the market fast. The other says AI demand absorbs it before that becomes a problem. The truth will likely land somewhere in between — but the companies positioned on the right side of whichever scenario materializes first will capture disproportionate value.

Huang placed his bet. The market will decide whether he was right.