Dell CEO Predicts AI Memory Demand Surge to Unprecedented Levels by 2028, Forcing Buyers to Pay High Prices

Michael Dell, the CEO of Dell Technologies, recently addressed the anticipated AI memory supercycle during a significant event, asserting that the surging demand for memory will continue unabated for several years to come.

The Urgency of Memory Investment Among Hyperscalers

As market analysts track the memory supply chain, it’s evident that the gap between supply and demand has widened in recent months. Despite this trend, uncertainties linger regarding the future of the memory industry. Recent developments, such as the TurboQuant incident and subsequent selloff of memory stocks, have fueled assumptions of a potential downturn in demand. However, Dell remains optimistic. He forecasts that the supercycle will extend through 2028, with an increasing demand for DRAM that far surpasses current figures.

“As memory per accelerator and system scale expand simultaneously in AI infrastructure, a structure is forming where total memory demand increases by approximately 625 times. While it takes years to expand memory supply, current demand for AI infrastructure is not slowing down.” – Michael Dell

One of the more compelling insights presented by Dell touches upon the growing memory demands per AI accelerator. Analyzing the evolution from NVIDIA’s Ampere architecture to the latest Vera Rubin design reveals a significant uptick in memory requirements. This surge isn’t solely reliant on advancements in High Bandwidth Memory (HBM); companies like NVIDIA are also pioneering technologies such as SOCAMM, which cater to specific AI workloads, resulting in an exponential rise in DRAM needs when assessed on an accelerator-by-accelerator basis.

Dell projects that memory requirements could skyrocket by up to 625 times by 2028, with each individual accelerator expected to experience a 25-fold increase in memory capacity. Furthermore, the deployment rate of these accelerators is also forecasted to grow 25 times. While these predictions may not have definitive empirical backing at present, they provide a critical viewpoint into the industry’s trajectory. Below is a snapshot of raw memory capacity from the Hopper architecture to Vera Rubin:

Bar chart showing total capacity increase, with the Vera Rubin (HBM4 + SOCAMM) reaching 1824 GB, noted as a '22.8x Total Increase' from the H100 (Hopper) and H200 (Hopper) models. — Per Accelerator Raw Memory Distribution | Image Credits: Wccftech

While raw capacity growth serves as a valuable measure for assessing demand trends, it is only one facet of a larger picture. Memory requirements are anticipated to rise with each generation of chips, particularly as we transition deeper into AI inference applications. A substantial share of DRAM purchases is fueled by hyperscalers and their investments in CXL memory technologies, indicating that projecting a decline in demand would be a misjudgment at this juncture.

Another critical development is the emerging long-term contracts between memory suppliers and hyperscalers, with agreements extending as much as five years. Such contracts indicate a strong willingness among buyers to allocate significant resources to secure their memory supplies. This scenario suggests that demand will not wane anytime soon, with potential shortages expected to persist well into the latter half of 2027, when new memory capacities are anticipated to come online.

Source & Images