Dell's CEO, Michael Dell, has discussed his estimates of the AI memory supercycle at a recent event, claiming that the explosive demand will persist for several years.
If Hyperscalers Do Not Spend Money on Memory, There's a Fear Within Them of Getting Behind the Competition
We have been tracking the memory supply chain for quite some time now, noting that the supply-demand gap has widened in the past few quarters; however, questions remain about how long we will see such conditions in the memory industry. Following the recent TurboQuant fiasco and the wider selloff within memory companies, there was a perception out there that somehow, the industry will now see a decline in demand; however, according to Dell's CEO, he expects the supercycle to persist moving up to 2028, and more importantly, the need for DRAM will increase signifcantly higher than where it is today.
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
There is a rather interesting metric discussed here: the memory demand for a particular AI accelerator and how it is growing steadily. When you look at the transition from Ampere to Vera Rubin, you'll realize that the need for memory has expanded significantly, and it isn't just limited to HBM advancements. NVIDIA and other manufacturers have introduced memory technologies like SOCAMM to support specific parts of AI workloads, which means that DRAM requirements have climbed exponentially when you narrow them down to a per-accelerator scope.
Dell estimates that memory demand could grow up to 625 times by 2028, claiming that a single accelerator would see a 25x increase in memory capacity and that accelerator deployment will grow by 25x as well. Of course, these estimates aren't backed by factual data for now, but this is what we are here for. Here's a comparison of raw memory capacity, scaling from Hopper all the way to Vera Rubin:
Raw capacity increase probably isn't the only metric for judging demand growth, but it is a solid indicator of the 'macro' perspective on how the memory industry is growing. The requirements for memory are increasing with each chip generation, and they will be much higher as we move towards the world of inference. A huge portion of DRAM orders is driven by hyperscalers and their CXL memory pools, which ultimately means that demand prospects are at a level right now that it would probably be a wrong judgment to say we are seeing a decline.
Another important factor to consider is that suppliers are now entering into agreements with hyperscalers that span up to 5 years, and so far, the reception has been tremendous for such arrangements. This is also an indication that buyers are willing to spend 'anything' to secure memory supply, and for suppliers, that demand will remain persistent. It wouldn't be wrong to say we should expect shortages to last for a few years, at least through H2 2027, since that's when new capacity comes online.
About the author: Muhammad Zuhair is a hardware and technology reporter for Wccftech, specializing in the semiconductor industry and the complex interplay between technology, manufacturing, and geopolitics. His coverage focuses on the corporate strategies and technological roadmaps of industry giants like TSMC, NVIDIA, Samsung, and Intel. Zuhair's expertise lies in deconstructing complex topics such as fabrication nodes (e.g., 2nm process), the economic impact of policies like the CHIPS Act, and the strategic development of AI infrastructure from NVIDIA, AMD and Intel.
Follow Wccftech on Google to get more of our news coverage in your feeds.
