DRAM Supply to Remain Tight in 2027, Prompting NVIDIA to Lower HBM Configurations for Rubin Ultra
TrendForce's latest research indicates that DRAM supply will remain constrained through 2027, with uncertainty surrounding memory suppliers' HBM4e validation timelines. These constraints are prompting AI chip vendors to reduce HBM configurations in their next-generation products.
Starting in Q3 2026, NVIDIA has begun evaluating alternative HBM configurations for its Rubin Ultra beyond the originally planned 12-Hi HBM4e, including 8-Hi HBM4e, 12-Hi HBM4, and 8-Hi HBM4 options. No final specification has been set. Other cloud service providers are similarly assessing lower HBM capacities for their next-generation in-house AI ASICs.
Recent memory shortages have already prompted several specification adjustments in AI chips. During H1 2026, cloud service providers and server OEMs reduced RDIMM capacities for server configurations. More recently, NVIDIA halved the SOCAMM capacity for its next-generation Vera Rubin Superchip modules, citing expectations that LPDDR5X supply constraints will persist through 2027.
NVIDIA held 12-Hi HBM4e as the baseline design for Rubin Ultra through H1 2026, but began evaluating lower-specification alternatives in early Q3 2026. Two supply-side constraints are driving this shift: the anticipated 2027 DRAM shortage will limit wafer capacity available for HBM production, and uncertainties remain regarding the validation schedule and yield ramp for 12-Hi HBM4e.
For Rubin Ultra, NVIDIA's priority is increasing I/O speed; GPU shipment volume expansion is secondary. If NVIDIA reduces HBM specifications, the reduction will come through fewer DRAM stack layers per GPU.
Whether HBM4e reaches validation and mass production on schedule will determine if Rubin Ultra can achieve 14–16 Gbps I/O speed—a gain from Rubin's 8–11.7 Gbps—or whether it will be limited to 11–12 Gbps through HBM4 optimization.
Within a given generation, DRAM stack-layer count determines the trade-off between HBM capacity per GPU and total GPU shipments.
The final configuration will also hinge on wafer allocations from memory suppliers. Despite projected 50–60% year-over-year growth in HBM bit shipments in 2027, supply is expected to fall short of demand growth.
HBM suppliers are expected to maintain pricing power through 2027, with significant price increases already widely anticipated. AI chip vendors will thus face limited supply and higher costs, strengthening the case for lower-capacity HBM configurations.