Memory becomes a deal driver
Memory demand from AI is exploding and is now a first‑order commercial variable rather than a late-stage line item, with Dell’s CEO estimating total AI memory demand could be 625x 2022 levels by 2028. That scale makes bill‑of‑materials exposure and supply timing central to how buyers evaluate infrastructure deals (pcgamer.com).
Dell’s chief executive Michael Dell says the artificial intelligence buildout is turning memory from a quiet parts-list item into a constraint big enough to shape whole server deals, with total artificial intelligence memory demand by 2028 potentially reaching 625 times 2022 levels. He tied that jump to two multipliers at once: memory per accelerator rising from 80 gigabytes in Nvidia’s 2022 H100 era to about 2 terabytes by 2028, and accelerator deployments rising about 25-fold. (pcgamer.com, msn.com) That changes what buyers are shopping for. In a normal server purchase, processors get the attention and memory is a line item; in an artificial intelligence cluster, the memory can decide whether the system is fast enough, available on time, and affordable enough to deploy at scale. (pcgamer.com, cnbc.com) The reason is simple: artificial intelligence chips spend their lives moving giant piles of model data around, and memory is the workbench sitting right next to the chip. If that workbench is too small or too slow, a very expensive accelerator waits around doing nothing. (ieee.org, wevolver.com) The premium version of that workbench is called high-bandwidth memory, which means memory built in vertical stacks and placed close to the processor so data can move across a very wide path. That design uses less space and delivers far more throughput than conventional server memory, which is why Nvidia and other accelerator makers depend on it for modern training systems. (wikipedia.org, wevolver.com) Once that kind of memory became essential, supply stopped behaving like a commodity market and started behaving like airline seats before a holiday weekend. SK hynix said in late 2025 that its 2026 supply of dynamic random-access memory, flash memory, and high-bandwidth memory was already sold out, showing how far ahead customers are now booking capacity. (techspot.com, bloomberg.com) Micron said in March 2026 that its high-bandwidth memory capacity was sold out through 2026 as well, and its quarterly results showed just how much pricing power tight supply can create. The company reported fiscal second-quarter 2026 revenue of $23.86 billion and gross margin of 74.9%, both lifted by artificial intelligence demand and constrained memory supply. (cnbc.com, storagenewsletter.com) That is why Dell is talking about memory in the same breath as revenue. Dell said on February 26, 2026 that it closed more than $64 billion in artificial intelligence-optimized server orders in fiscal 2026, shipped more than $25 billion, and entered fiscal 2027 with a record backlog of $43 billion. (dell.com, finance.yahoo.com) A backlog that large is not just a sales story. It is also a parts story, because every delayed memory shipment can push back a rack, a cluster, or a whole data center handoff, and every memory price jump can rewrite the bill of materials before the customer signs. (cnbc.com, dell.com) That is why memory is becoming a deal driver instead of a procurement footnote. Buyers now have to ask not just which accelerator they want, but which memory stack comes with it, when it is available, and whether the supplier locked in enough capacity months ago to actually deliver the system. (pcgamer.com, techspot.com, cnbc.com) If Michael Dell’s 625-times estimate is even close, the next few years of artificial intelligence infrastructure will be shaped as much by who can secure memory as by who can design chips. In that market, the winning server quote is not the one with the best brochure; it is the one with memory attached and a delivery date that holds. (msn.com, dell.com)
