NVIDIA is reportedly testing multiple lower-memory configurations of its forthcoming Rubin Ultra accelerator to cope with high-bandwidth memory supply constraints. Designs under test include versions with as little as 192 GB or 256 GB (down from the full 1 TB HBM4E initially announced) and now use HBM4 instead of HBM4E, which has customizable logic dies and greater manufacturing complexity.
The memory downgrade reflects a broader constraint: SK Hynix, Samsung, and Micron have sold through their HBM capacity through 2027, and SK Hynix's CEO warned 2027 will be the 'worst year' for memory shortage, with supply crunch extending through 2030. NVIDIA has also trimmed its Rubin Ultra from a quad-die design (canceled in June due to manufacturing complexity) to a likely dual-GPU variant, making lower memory per unit more palatable. Per-GPU memory capacity in the base Rubin is currently 288 GB.
For enterprises planning Hopper-to-Rubin migration, this signals memory will be the gating factor for 2027–2028 deployments, not compute chiplet availability. Customers prioritizing long-term Rubin relationships over per-GPU VRAM will face tighter per-system memory budgets. Watch for further NVIDIA SKU announcements as test results mature.