DRAM Supply to Remain Tight Through 2027; NVIDIA Downgrades Rubin Ultra HBM Configuration

According to the latest memory industry research by TrendForce, due to the continued tight supply-demand imbalance of DRAM through 2027 and uncertainties in original manufacturers' HBM4e validation timelines, NVIDIA (NVDA-US) has revised its HBM configuration plan for Rubin Ultra since the third quarter of this year. Instead of sticking to HBM4e 12hi, the company is now concurrently evaluating multiple design options, including HBM4e 8hi, HBM4 12hi, and HBM4 8hi. A final decision has not yet been made.

Beyond NVIDIA, some cloud service providers (CSPs) are also considering reducing HBM capacity in their next-generation in-house ASIC designs.

TrendForce notes that recent memory supply shortages have already triggered several downgrades in DRAM specifications for AI chips. Since the first half of 2026, CSPs and server OEMs have successively reduced RDIMM capacities. Recently, NVIDIA decided to halve the SOCAMM capacity on its next-generation Vera Rubin Superchip module, citing ongoing LPDDR5X shortages expected to persist through 2027.

For Rubin Ultra, NVIDIA had initially adopted HBM4e 12hi as the baseline design from 2025 through the first half of 2026. However, starting in early Q3, the company opened evaluations for lower-tier specifications and continues to deliberate. This move primarily addresses two supply-side bottlenecks.

First, overall DRAM supply will remain tight through 2027, limiting wafer production capacity that manufacturers can allocate to HBM. Second, there remains uncertainty regarding the validation timeline and yield improvement progress for HBM4e 12hi.

TrendForce indicates that NVIDIA’s primary goal for the Rubin Ultra generation is enhancing I/O speed, while the secondary objective is increasing GPU shipment volume. If a final decision involves downgrading HBM specifications, adjustments to the number of DRAM stack layers are expected to be the main approach.

In summary, whether HBM4e can complete validation and ramp to volume production on schedule will determine if Rubin Ultra’s I/O speed can upgrade from Rubin’s previous 8–11.7 Gbps range to 14–16 Gbps—or only reach 11–12 Gbps through optimized HBM4 designs. Meanwhile, within any given generation, the number of stack layers determines the trade-off between per-GPU HBM capacity and total GPU units available for shipment.

TrendForce believes that final specification decisions will also depend on original manufacturers’ wafer allocation. On the supply-demand front, HBM shipment bit volume is projected to grow 50–60% year-on-year in 2027, still insufficient to meet demand growth. Under persistent supply shortages, TrendForce expects HBM pricing power to remain with suppliers in 2027, with significant price hikes becoming an industry consensus. AI chipmakers will face dual pressures of constrained HBM supply and rising procurement costs, strengthening their motivation to adopt lower-capacity HBM solutions.