NVIDIA is under pressure from supply shortages: Rubin Ultra chips may consider reducing HBM configuration, and the DRAM shortage is expected to persist until 2027.

Author: TrendForce

Translated by: TechFlow

Editor's Note: Persistent memory supply shortages are changing AI chip design logic. Starting from Q3 2026, NVIDIA is considering lowering the HBM configuration for its next-generation Rubin Ultra chips, expanding from the original plan of 12-layer HBM4e to include options such as 8-layer HBM4e, 12-layer HBM4, and 8-layer HBM4. This is not a technical decision but a compromise forced by supply chain constraints—DRAM shortages are expected to continue until 2027, the mass production timeline for HBM4e is still uncertain, and AI chipmakers must balance performance with shipment volume.

According to TrendForce's latest memory industry research, DRAM supply shortages will persist until 2027, and memory suppliers' HBM4e verification timelines remain uncertain. These constraints are prompting AI chip manufacturers to cut HBM configurations in their next-generation products.

Starting from Q3 2026, NVIDIA has broadened its evaluation of Rubin Ultra chip HBM configurations, expanding from the initial 12-layer HBM4e design to include alternative options such as 8-layer HBM4e, 12-layer HBM4, and 8-layer HBM4. The final specifications have yet to be determined. In addition to NVIDIA, several cloud service providers are also assessing reduced HBM capacities for their next-generation self-developed AI ASIC chips.

TrendForce notes that recent memory shortages have resulted in multiple adjustments to AI chip memory specifications. For example, cloud service providers and server OEMs reduced RDIMM capacity in server configurations during the first half of 2026. Recently, after determining that LPDDR5X supply limitations may persist until 2027, NVIDIA decided to halve the SOCAMM capacity of the next-generation Vera Rubin superchip module.

From 2025 through the first half of 2026, NVIDIA maintained 12-layer HBM4e as the baseline design for Rubin Ultra. But starting from early Q3 2026, the company began evaluating lower-spec alternatives. This change is primarily driven by two supply-side limiting factors: first, the anticipated overall DRAM shortage in 2027 will restrict the wafer capacity that memory suppliers can allocate to HBM production. Second, the verification timeline and production ramp-up for 12-layer HBM4e remain uncertain.

TrendForce points out that NVIDIA's top priority for this generation of Rubin Ultra products is increasing I/O speed, while expanding GPU shipment volume is of secondary importance. If NVIDIA ultimately decides to reduce HBM specifications, it is expected to do so by reducing the number of DRAM stack layers.

Ultimately, whether HBM4e can pass verification and enter mass production as scheduled will determine whether Rubin Ultra's I/O speed can increase from the previous Rubin generation's 8-11.7 Gbps to 14-16 Gbps, or only reach 11-12 Gbps through HBM4 design optimizations.

At the same time, within the same product generation, the number of DRAM stack layers determines the trade-off between per-GPU HBM capacity and the number of GPUs available for shipment.

TrendForce believes the final configuration will also depend on memory suppliers' wafer allocation decisions. From a supply and demand perspective, HBM bit shipments are expected to increase by 50-60% year-on-year in 2027, but this will still not be enough to keep up with demand growth.

TrendForce expects that under these supply-constrained conditions, HBM suppliers will retain pricing power throughout 2027, and the industry widely anticipates a significant increase in HBM prices. Thus, AI chipmakers will face the dual challenge of limited HBM supply and rising procurement costs, increasing the incentive to adopt lower-capacity HBM configurations.