DRAM Shortage Could Push Nvidia to Reduce Rubin Ultra HBM Configuration
TAIPEI – Tight DRAM supply expected to persist through 2027 is prompting Nvidia to evaluate lower HBM configurations for its next-generation Rubin Ultra AI platform, according to TrendForce.
Nvidia has expanded its evaluation of Rubin Ultra beyond the originally planned 12-Hi HBM4e configuration to include 8-Hi HBM4e, 12-Hi HBM4 and 8-Hi HBM4 alternatives. TrendForce said the final specification has not been determined.
The potential change is being driven by an expected industrywide DRAM shortage in 2027 that could constrain wafer capacity available for HBM production, along with uncertainty surrounding validation and production yields for 12-Hi HBM4e.
Other AI hardware developers are also adjusting memory configurations in response to supply conditions. TrendForce said several cloud service providers are considering lower HBM capacities for next-generation in-house AI ASICs, while Nvidia has decided to halve the SOCAMM capacity of its Vera Rubin Superchip modules amid expectations that LPDDR5X constraints will continue through 2027.
TrendForce said Nvidia's primary objective with Rubin Ultra is increasing I/O speed. Successful qualification of HBM4e could enable speeds of 14–16Gbps, compared with 8–11.7Gbps for Rubin. An optimized HBM4 configuration could instead reach 11–12Gbps.
HBM bit shipments are forecast to grow 50–60% year-over-year in 2027, but TrendForce expects that increase to remain insufficient to meet demand. The firm consequently expects HBM suppliers to retain pricing power throughout the year, creating both availability and cost pressures for AI chip vendors.
