NVIDIA's Rubin Ultra May Drop to 8-Layer HBM4E as Memory Shortage Squeezes AI Chipmakers

Why is NVIDIA revising Rubin Ultra's HBM configuration, and what are the alternatives?

NVIDIA (輝達) originally planned its next-generation AI chip, Rubin Ultra, to carry 12-layer HBM4E, according to Liberty Times Net (自由時報電子報). The outlet reported that NVIDIA has since reviewed downgraded alternatives — either 8-layer HBM4E or 12-layer HBM4 — "to ensure design flexibility."

Market research firm TrendForce (集邦科技) disclosed on August 6 that NVIDIA had not yet finalized Rubin Ultra's final specification and was discussing a reduction in HBM specs, per Liberty Times Net. Financial outlet Cnyes (鉅亨網) framed the same development under the headline "NVIDIA's Rubin Ultra reportedly to cut HBM capacity — how will chipmakers break through amid the memory shortage?", reflecting the same narrative reported by Liberty Times Net.

How lopsided is AI chip demand for HBM, and what does the industry expect for next year?

SpaceX CEO Elon Musk (馬斯克), speaking at a recent earnings call, said memory production is growing 20% year-over-year while demand is "exploding at over 200%," according to Liberty Times Net.

Samsung Electronics (三星電子) said at its own earnings call that it expects the supply-demand gap to be more severe next year than this year. SK Hynix (SK海力士) said it expects demand expansion to continue undiminished, citing steady capital investment from major technology companies.

What is the root capacity bottleneck behind the HBM shortage, and when might it ease?

BusinessKorea reported, as cited by Liberty Times Net, that NVIDIA's downgrade review stems primarily from DRAM capacity constraints — memory makers have limited production lines they can allocate to HBM — and that this situation is expected to persist into next year. BusinessKorea also noted that validation progress and mass-production yield for 12-layer HBM4E remain uncertain.

That persistence lines up with Samsung Electronics' own outlook: the company told its earnings call that it expects the supply-demand gap to widen further next year, reinforcing BusinessKorea's timeline that the DRAM-driven bottleneck will not resolve in the near term.

What dual dilemma will rising HBM prices create for chipmakers?

TrendForce said it expects HBM prices to rise further, and that AI chip manufacturers will face a "dual burden" — higher procurement costs alongside difficulty securing sufficient procurement volume — according to Liberty Times Net.

That price pressure sits directly on top of the demand-supply imbalance Musk described: with demand growing over 200% annually against 20% supply growth, TrendForce's forecast of further price increases is the market's direct response to that gap.

Facing the HBM shortage and cost pressure, what are chipmakers' options and trade-offs?

The evidence points to a narrow set of levers, none of them clean. NVIDIA's own move — reviewing 8-layer HBM4E or 12-layer HBM4 as alternatives to 12-layer HBM4E for Rubin Ultra, per Liberty Times Net — is one route to reduce dependence on the highest-layer, hardest-to-validate HBM configuration.

But BusinessKorea's reporting shows why that route is constrained rather than free: DRAM production lines allocable to HBM are limited, and 12-layer HBM4E's validation and yield remain unresolved, a condition expected to persist into next year. Meanwhile, TrendForce's forecast of further HBM price increases means that even chipmakers who secure supply still face the "dual burden" of rising cost and uncertain volume. A spec downgrade may ease the volume side of that equation, but nothing in the evidence indicates it resolves the pricing pressure TrendForce described.

What this means

Taken together, the reported figures form a chain: Musk's 20%-supply-versus-200%-demand gap (via Liberty Times Net) is the imbalance that Samsung Electronics expects to widen further next year, and that BusinessKorea traces to a specific cause — limited DRAM capacity allocable to HBM plus unresolved 12-layer HBM4E validation and yield. NVIDIA's review of 8-layer HBM4E or 12-layer HBM4 for Rubin Ultra, and TrendForce's forecast of further HBM price rises with a "dual burden" for chipmakers, both read as direct responses to that same bottleneck rather than separate developments — though none of the sources in this evidence set state whether NVIDIA's downgrade review will ultimately be adopted as Rubin Ultra's final specification.