Memory pricing and demand are wobbling
A recent infrastructure round-up and industry commentary is signaling that the overheated memory narrative may be cooling, after reports that large letters of intent for DRAM wafer allocations were non‑binding and markets reacted — including a notable share-price drop at major DRAM names. The discussion also flagged a major dynamic: efficiency tricks (one report referenced a technique that could cut model memory needs by roughly 6×) can either reduce component pressure or simply let teams run far bigger models, while real-world bottlenecks like power capacity and industrial supplies keep datacenter builds constrained. (x.com) (youtube.com)
For months, the memory boom looked like one of the cleanest stories in tech. AI companies needed vast amounts of DRAM and flash. Memory makers shifted capacity toward high-margin server parts. Prices jumped. Analysts started talking as if the industry had escaped its old boom-and-bust cycle for good. Then the story hit a softer patch, not because demand vanished, but because some of the loudest demand signals turned out to be less solid than they looked. The first crack showed up in the market itself. TrendForce said on April 1 that DRAM spot momentum for DDR4 and DDR5 had turned weak, with buyers resisting elevated prices and only making small purchases at steep discounts. NAND spot trading looked sluggish too. That matters because spot markets are where strain shows up first. Tight supply can keep contract prices high for a while, but weak spot trading is often the first sign that buyers are blinking. (trendforce.com) That was awkward timing, because just days earlier the industry was still telling a very different story. TrendForce had forecast conventional DRAM contract prices rising 58% to 63% quarter over quarter in the second quarter, driven by suppliers steering more capacity toward HBM and server products. CNBC reported that Micron, Samsung, and SK Hynix were all describing a market where customers could get only part of what they wanted and where multi-year supply agreements were becoming normal. The shortage story was real. It was also getting priced as if nothing could interrupt it. (trendforce.com) Then investors got a reminder that “demand” is not the same thing as “binding demand.” Reporting in late March described large letters of intent tied to DRAM wafer allocations as non-binding expressions of interest, not enforceable purchase commitments. A non-binding letter can still move a supply chain if manufacturers treat it as a serious signal. But it is not the same as an order book. If the market had been treating those allocations as locked in, the repricing was inevitable. (thedeepdive.ca) The share-price reaction made that clear. CNBC reported on March 26 that memory names sold off after Google unveiled TurboQuant, a compression method it said could cut the memory needed to run large language models by six times. SK Hynix fell 6% and Samsung nearly 5% in South Korea, while Micron had already dropped in the U.S. The move looked dramatic, but the logic was simple: if AI can do more work with less memory, the industry’s most crowded trade suddenly looks less certain. (cnbc.com) The catch is that efficiency does not neatly reduce hardware demand. TurboQuant targets the key-value cache used during inference, which is one of the biggest memory bottlenecks in serving large models. Remove a bottleneck and companies do not always buy less hardware. They often use the same budget to run larger models, serve more users, or lower latency. Even the analysts quoted in the sell-off said better efficiency could make AI systems more capable, which can pull memory demand right back up. (cnbc.com) That is why the real constraint may be elsewhere. Data center growth is still pinned to power, construction, and industrial capacity more than to any single component. OpenAI, Oracle, and their partners have kept describing Stargate as a multi-gigawatt build-out, with Abilene already operating and total planned capacity across sites reaching nearly 7 gigawatts. But one planned expansion at Abilene was scrapped after financing talks dragged and OpenAI’s needs changed, and the site was later picked up for Microsoft infrastructure instead. The lesson is not that AI demand disappeared. It is that this market is being shaped by real estate, grid access, and financing just as much as by chips. (openai.com) That leaves memory in an uncomfortable middle. Suppliers are still tight on paper. Contract prices are still elevated. But the clean narrative has broken. Some demand signals were softer than advertised. Some efficiency gains are arriving faster than bulls expected. And the biggest AI projects still have to pass through the slowest parts of the physical world, where even a flagship campus in Abilene took a second phase of six buildings to reach 1.2 gigawatts by the end of 2026. (trendforce.com)
