The AI's Memory Hunger Will Not Stop: Dell Forecasts 'Out of Scale' Numbers

The memory market continues to be one of the key elements in the development of infrastructures for artificial intelligence. According to Michael Dell, the current expansion phase does not represent a temporary peak but a true "supercycle" destined to continue for several more years, at least until 2028.

The statements come in a context where demand does not seem to show signs of slowing down. On the contrary, the competitive pressure among hyperscalers is pushing investments to ever-higher levels, especially to avoid delays in the development of AI platforms.

One of the most relevant aspects concerns the scale of the expected growth. Dell estimates an overall increase in memory demand of up to 625 times compared to current levels. This dynamic would arise from combined factors: first, an increase of up to 25 times in memory per single accelerator, and, second, a similar growth in the number of accelerators distributed in data centers.

The result is a structural multiplication of the overall demand for DRAM, driven by the continuous evolution of AI architectures. The transition between generations of GPUs and accelerators already highlights a significant increase in the necessary memory, not limited to HBM solutions alone but extended to other emerging technologies.

The evolution of AI platforms relates not only to capacity per single chip, but also to how memory is organized and shared. Technologies such as memory pools based on CXL are becoming central to hyperscaler strategies, enabling more flexible resource management.

This approach implies further expansion of DRAM demand, as memory is no longer confined to a single accelerator but becomes a resource distributed at the system level. Alongside this, dedicated solutions for specific AI workloads further contribute to expanding the overall requirements.

A significant signal also comes from the commercial side. Major cloud operators are entering into supply agreements with memory manufacturers with horizons of up to five years. These agreements clearly indicate the intent to secure production capacity in advance, even in the face of high costs.

According to Dell, there is a genuine "fear of falling behind" among hyperscalers: the availability of memory has become a critical competitive factor for the development and scalability of AI models. The challenge is that, in the face of rapidly growing demand, supply requires much longer times to adjust. The expansion of production capacity in the memory sector requires significant investments and multi-year cycles, and this creates an imbalance that is likely to persist.

Current estimates indicate that any improvements may not materialize until the second half of 2027. Until then, the market may continue to operate under pressure conditions, with the risk of structural shortages.