Supply Tight! NVIDIA Considers Lowering HBM Memory Configuration for Rubin Ultra

According to a report by _The Information_ on the 6th, NVIDIA (NVDA-US) is considering lowering the HBM memory configuration for its next-generation AI GPU, Rubin Ultra. This is a measure to ensure production plans due to a shortage of high-end HBM chips. Insiders reveal that NVIDIA has tested at least three different versions of the Rubin Ultra GPU in recent weeks, some of which use lower HBM memory capacity than originally planned. After the report, NVIDIA's stock price fell 0.1%, and SK Hynix's stock price fell 4.97%.

HBM has become the 'hardest bottleneck' in the AI supply chain. According to TrendForce's research, the DRAM supply shortage will continue until 2027, and HBM shipment volume is expected to increase by 50% to 60% annually, but this will still be insufficient to meet demand growth. The reasons for the HBM supply shortage are twofold: limited overall DRAM production capacity and uncertainty in the verification and yield improvement of HBM4e 12hi.

NVIDIA has been using HBM4e 12hi as the baseline design for Rubin Ultra from 2025 to the first half of 2026, but began evaluating alternative lower-specification plans in the third quarter of 2026. Reducing HBM capacity will affect chip performance, but NVIDIA can partially compensate for this through other technological means, such as improving I/O speed.

This adjustment is a strategic choice for NVIDIA to maintain product competitiveness while ensuring production. However, customers may need to deploy more GPUs to handle large-scale AI models, potentially increasing data center construction costs for cloud providers. HBM suppliers are expected to continue to hold pricing power.