Dell CEO Michael Dell has released a forecast that the demand for memory for artificial intelligence infrastructure will increase 625-fold by 2028.

He assumes that the memory volume per accelerator will increase 25-fold, and the number of accelerators themselves will also increase 25-fold. The entrepreneur noted that ramp-up takes a long time, while the need for AI equipment does not diminish.

The generalization of NVIDIA GPUs confirms this trend, with memory consumption increasing significantly from Ampere to Vera Rubin. This is not just about the evolution of HBM, but also about implementing new standards like SOCAMM. Comparison of the net capacity from Hopper to Vera Rubin with HBM4 and SOCAMM shows a gap of 22.8 times, and the total value per accelerator is approaching 1824 GB. As the industry shifts toward the demands of the industry, memory requirements will continue to tighten.