The quietest part of an AI server is becoming the loudest line item. The accelerator still gets the keynote applause, because of course it does, it is the vault door in the heist movie. But Omdia’s latest forecast says the getaway driver is memory, and the getaway driver is now negotiating hazard pay. If you buy, build, or budget AI infrastructure, the lesson is simple: the chip you can afford may matter less than the memory and packaging capacity you can actually get. ## The forecast under the heat spreader According to the July 30, 2026 FT.com Company Announcement carrying Omdia’s release, Omdia raised its 2026 semiconductor revenue forecast to 94.1% year-over-year. The announcement attributes that jump to exceptional growth in DRAM and NAND as AI demand continues to outpace global supply. Here is the buried spec that changes the whole board layout: Omdia now expects memory ICs to account for more than 50% of total semiconductor revenue in 2026. That is not a sidebar, that is the bill of materials walking into the boardroom and asking for a reserved parking space. The useful read is not merely that memory prices respond to demand. We all know supply curves bite, the same way a linear regulator gets warm when you ask it to impersonate a toaster. The important shift is that memory ICs are moving from supporting cast to revenue majority, according to the FT.com Company Announcement. For planners, that means AI capacity conversations need to start with memory availability, not end there after the accelerator spreadsheet is already laminated. ## The package is now part of the product The same FT.com Company Announcement says AI demand has exceeded the industry’s current ability to produce and package chips. It names bottlenecks across high bandwidth memory, advanced packaging, and node capacity, with those constraints expected to persist until at least 2027. Let’s talk about what they did not mention in the keynote: a processor without enough surrounding supply chain is a race car on jack stands. It may be glorious, expensive, and totally unable to leave the garage. That matters because AI infrastructure is not one heroic component. It is a chain of physical dependencies, DRAM, NAND, HBM, advanced packaging, and the node capacity needed to manufacture chips at scale. When one link tightens, procurement teams do not feel it as an abstract market note, they feel it as lead times, redesigns, and uncomfortable meetings about whether the rack plan still fits the calendar. This is where good engineering beats good marketing: the system ships only when the dull parts of the supply chain agree to cooperate. ## The April clue that became the July warning Evertiq reported that Omdia had already lifted its 2026 semiconductor revenue forecast to 62.7%, citing AI-driven demand and tighter memory supply conditions. That earlier report said DRAM was forecast to nearly double in value compared with 2025, while NAND was expected to rise to almost four times its 2025 level. The April 24, 2026 FT.com Company Announcement also said conventional memory IC supply constraints were being exacerbated by the industry’s focus on HBM production, which delivers lower volumes but commands significantly higher prices. In other words, the July figure did not fall from the sky, the warning light had been blinking on the dashboard. The April FT.com Company Announcement also tied the outlook to a major server refresh cycle in 2026 and exceptional hyperscaler capital expenditure. That is the demand side of the crime scene, lots of buyers trying to retire older hardware and feed more demanding AI workloads. Pair that with limited supply relief until well into 2027, as the April FT.com Company Announcement stated, and the market starts to look less like a simple GPU shortage and more like a multi-room bottleneck. Memory is not just being consumed, it is being structurally prioritized, repriced, and fought over. ## What builders should do with this signal For product teams, the Omdia forecast is a reminder to qualify memory assumptions as aggressively as compute assumptions. If your AI product roadmap depends on a specific memory configuration, treat that dependency like a voltage rail feeding the whole board, not like a replaceable sticker on the chassis. Ask suppliers about DRAM, NAND, HBM, advanced packaging, and node capacity exposure before locking a deployment schedule. The unpleasant surprise is never the spec sheet, it is the part of the spec sheet that becomes unavailable after the design review. For buyers, the practical move is to compare systems by availability, serviceability, and memory configuration, not just accelerator branding. For investors and operators, Omdia’s July forecast suggests the value in AI infrastructure may keep migrating toward the less glamorous parts of the stack. Watch whether the 2027 constraint language begins to soften in future Omdia updates, because that will tell us whether the supply chain is catching up or merely changing which gasket squeals first. Until then, the smartest AI hardware conversation may start with memory, the component that used to sit quietly in the corner and now appears to have the keys to the vault. ## Sources - Omdia: AI Demand Drives 94.1% Surge in Semiconductor Forecast for 2026, Company Announcement, FT.com

Sources