Nvidia evaluates lower-memory configuration for Rubin Ultra chip amid HBM shortage
Nvidia $NVDA is evaluating a lower-memory version of its upcoming Rubin Ultra AI chip, including an option with 192GB of HBM4, a 33% cut from the 288GB in the current Vera Rubin GPU. The change is driven by a DRAM and HBM supply shortage expected to persist into 2027.
Nvidia $NVDA is evaluating reduced-memory configurations for its upcoming Rubin Ultra AI chip, according to industry reports, including an option that would cut on-package memory to 192GB of HBM4, down 33% from the 288GB carried by the current Vera Rubin GPU. Nvidia had maintained a higher-memory baseline design through the first half of 2026 before beginning to evaluate lower-spec alternatives starting in the third quarter. The change is driven by a broader DRAM and high-bandwidth-memory supply shortage that industry analysts expect to persist into 2027, limiting how much wafer capacity memory suppliers can allocate to HBM production. That a memory downgrade is being considered for the industry's most anticipated AI chip underscores how deep the current shortage runs, even for the company that dominates AI accelerator sales, with implications for the compute budgets cloud providers plan around Rubin Ultra deployments.