Nvidia testing lower memory Rubin Ultra configurations (192 GB, HBM4) as HBM supply shortage pressures SKU mix.
Nvidia is reportedly testing variations of its upcoming Rubin Ultra accelerator with less memory than originally planned, due to concerns about sourcing sufficient HBM. According to The Information, some configurations under evaluation include just 192 GB of memory and use HBM4 instead of the originally announced HBM4E. The report confirms earlier commentary from SemiAnalysis about a potential Rubin Ultra memory downgrade.
Rubin Ultra was unveiled earlier this year at GTC, where Nvidia displayed a compute tray housing four compute chiplets alongside 1 TB of HBM4E memory. The accelerator forms part of Nvidia's Kyber NVL144 design, scheduled to roll out in 2027, though SemiAnalysis reported the rack has been delayed to 2028. Nvidia told Tom's Hardware that "Our roadmap is intact," though the company declined to clarify whether the delay was confirmed.
According to The Information, Nvidia is testing Rubin Ultra versions with 192 GB or 256 GB of memory, as well as configurations using fewer than the 16 announced memory stacks. Most significantly, Nvidia is reportedly testing with HBM4 rather than HBM4E as originally announced. While HBM4E shares traditional generation-over-generation improvements, it uniquely offers a customizable base logic die—a feature that Micron developed with TSMC to allow customers to tailor the logic die to their specific needs. This complexity has reportedly strained supply chains, leaving memory manufacturers unable to keep pace with Rubin Ultra's rollout. At least three lower-memory designs have been tested, though full details remain unclear.
The number of memory dies carries significance. In June, reports surfaced that Nvidia cancelled its quad-die Rubin Ultra design due to manufacturing complexities and would instead pursue a dual-GPU configuration. With a dual-die Rubin Ultra, reduced memory capacity would make more sense, though the tested capacities of 192 GB and 256 GB remain lower than expected—each base Rubin GPU currently ships with 288 GB of HBM4.
Nvidia is clearly attempting to navigate a world where memory agreements have been locked in years in advance. The company holds multiple such agreements. In June, Nvidia announced a partnership with SK Hynix to develop next-generation memory technology, including HBM, LPDDR5X, and DDR5. In July, Nvidia expanded that relationship with a $500 billion strategic commitment that includes a long-term memory supply agreement.
Memory shortages are affecting nearly every design currently on the market, with enterprise systems using HBM particularly vulnerable. Last week, Digitimes reported that Samsung, SK Hynix, and Micron have sold through their HBM capacity through 2027. SK Hynix CEO Kwak Noh-jung recently stated that 2027 will be the "worst year" for the memory shortage, with supply constraints extending through 2030.
Despite Nvidia's consideration of lower-memory configurations, at least one Nvidia customer told The Information that per-GPU memory is not their primary concern, valuing the long-term relationship with Nvidia over specific hardware specifications. In response to the report, Nvidia stated: "With NVIDIA Rubin Ultra, we are optimizing across compute, networking and memory to deliver the best performance and efficiency for our customers' AI deployments. These optimizations let us build more GPUs and deploy more AI systems."