Memory Demand Warning
Michael Dell warned that AI memory demand could surge as much as 625‑fold by 2028, suggesting memory and DRAM could become binding constraints alongside accelerators. The projection underlines how AI’s economics extend beyond GPUs to broader bill‑of‑materials pressures. (punemirror.com)
Artificial intelligence systems are starting to run into a different limit: memory chips, not just the processors that do the math. Michael Dell said this week that total memory demand in artificial intelligence infrastructure could rise about 625-fold by 2028. (timesnownews.com) Dell made the forecast at a Bank of America event, where he said the jump comes from two multipliers at once: roughly 25 times more memory per accelerator and roughly 25 times more accelerators deployed. Multiplying those two figures produces the 625-times estimate. (msn.com) Memory is the working table for an artificial intelligence model: it holds the model weights, the incoming data, and the intermediate results while the accelerator computes. When that table is too small, companies have to split jobs across more chips and more servers, which raises cost, power use, and network traffic. (nvidia.com) That is why the memory attached to each accelerator has been climbing. Nvidia’s H100 shipped with 80 gigabytes of high-bandwidth memory, while the H200 moved to 141 gigabytes, giving buyers a larger pool of very fast memory on each chip. (nvidia.com, nvidia.com) The supply side is tighter than a normal commodity cycle because this is not standard laptop memory. The fastest stacks, called high-bandwidth memory, are built in advanced packages next to accelerators, and adding new output takes new fabrication, packaging, and testing capacity that can take years to expand. (nvidia.com, forbes.com) Buyers are already acting as if memory is scarce. Micron’s high-bandwidth memory supply for calendar 2025 was fully booked last year, and industry reporting in 2026 has described high-bandwidth memory output as effectively sold out well into the next cycle. (semimedia.cc, forbes.com) Dell has a direct view into that squeeze because it sells the servers that bundle accelerators, networking, storage, and memory into artificial intelligence systems. In its annual report for the fiscal year ended January 31, 2025, Dell said backlog levels for its artificial-intelligence-optimized servers remained elevated as it exited the year. (sec.gov) The warning also shifts the conversation away from graphics processing units alone. A server build now depends on a wider bill of materials — high-bandwidth memory, conventional dynamic random-access memory, advanced packaging, power delivery, cooling, and networking — and a shortage in any one of those parts can slow deployments. (sec.gov, nvidia.com) Not everyone accepts the 625-times figure at face value. Some coverage of Dell’s remarks has noted skepticism about whether deployments and per-chip memory can both compound that quickly through 2028, especially if software becomes more efficient or if buyers slow spending. (notebookcheck.net) Dell’s core point is narrower than the headline number: the cost of artificial intelligence infrastructure is being pushed by memory as well as compute. If that pattern holds through 2028, the next bottleneck in the artificial intelligence buildout may sit beside the accelerator, not on it. (punemirror.com, msn.com)
