NVIDIA plans to lower the Rubin Ultra HBM configuration to address the high-end shortage, and has tested at least three lower-memory versions.

Nvidia is evaluating adjustments to the HBM configuration of its next-generation AI GPU Rubin Ultra, planning to launch a version with lower video memory than originally scheduled to ease production pressure caused by the shortage of high-end HBM supplies. Sources familiar with the matter revealed that Nvidia has tested at least three versions with different memory capacities over the past few weeks, some adopting lower memory specifications. This means that even Nvidia, which holds a dominant position in the GPU market, has had to make compromises between product specifications and supply capabilities. The Rubin Ultra is positioned above the upcoming Rubin series set for mass production, and was originally planned to be equipped with higher-capacity, higher-bandwidth HBM to boost performance for large model training and inference. If the low-memory configuration is ultimately adopted, customers will need to deploy more GPUs when running large AI workloads such as large language models, leading to higher system costs and increased cluster complexity. For cloud vendors including Microsoft, Meta, Amazon, and Google, which are continuously expanding their AI capital expenditures, data center construction costs may rise further. Nvidia’s move also reflects that despite the company holding the strongest bargaining power in the supply chain, HBM supply constraints remain a core bottleneck for AI computing power expansion. High-end HBM production capacities at SK Hynix and Micron have already been tight, and Nvidia’s proactive configuration reduction further confirms that the HBM supply-demand imbalance will be difficult to ease in the short term.