Nvidia reportedly tests cheaper, lower-memory Rubin Ultra configs

Nvidia is reportedly testing at least three Rubin Ultra configurations with as little as 192GB of memory, well below the 1TB of HBM4E (fourth-generation high-bandwidth memory) originally announced. The change reportedly comes as HBM supply tightens industry-wide.

Why it matters: If confirmed, this is a direct sign that the AI hardware buildout is now running into memory supply constraints, not just GPU compute limits. Lower-memory variants could mean higher effective costs or reduced model-serving capacity for buyers expecting the original 1TB spec, adding to the compute-versus-memory bottleneck concerns that have shadowed this GPU generation.