HBM and DRAM manufacturers have reportedly sold out their entire production capacity through 2027, with all remaining slots committed to hyperscalers.
The memory industry has reportedly booked out its available production capacity for all of 2027, with both DRAM and high-bandwidth memory already assigned to customers. If accurate, it means new customers today would find almost nothing left to purchase for a year that hasn't begun.
This doesn't mean store shelves will be empty in 2027. In semiconductor terms, "sold out" means suppliers have assigned their currently planned output to customers through contracts, volume commitments, prepayments, and reserved production slots. Chips will still be manufactured; there is simply very little uncommitted capacity for late arrivals.
A new report suggests that all three memory manufacturers—Samsung, SK Hynix, and Micron—have collectively sold through their 2027 memory manufacturing capacity to AI companies. The driver is straightforward: AI accelerators require enormous quantities of HBM (high-bandwidth memory), which stacks DRAM dies vertically to feed data to chips at the necessary speed. The servers around those accelerators then require conventional DRAM as well. A single data center order can consume more memory than an entire console generation.
Hyperscalers and AI firms are also ideal customers from a supplier perspective. They plan years ahead, sign multi-year deals, pay upfront, and accept premium pricing—far more attractive than laptop brands ordering quarter by quarter.
Advanced memory comes from a very short list of suppliers: Samsung, SK Hynix, and Micron. HBM narrows this further, requiring die stacking, through-silicon vias, advanced packaging, and platform-specific qualification. Packaging and testing can bottleneck supply even when DRAM wafers are available.
New fabs require years to build. They demand billions in capital, clean rooms, lithography equipment with lengthy lead times, substantial power and water infrastructure, trained workforces, and a long production ramp before yields become profitable. Memory also carries a brutal boom-and-bust history—Qimonda went insolvent in 2009 after a demand crash—making suppliers cautious about overbuilding.
Instead of expanding capacity, suppliers can shift their production mix toward higher-margin HBM. This benefits AI customers but can starve conventional DRAM, the memory that ends up in gaming PCs, laptops, and consoles.
The knock-on effects for gaming hardware are plausible but not guaranteed. DRAM contract prices have already climbed sharply, and hardware makers have warned about component costs. Higher-memory PC configurations could become pricier, discounts could thin, and standard RAM and VRAM capacity growth could slow. Console makers typically lock in supply years in advance, so existing deals may shield them, but future renegotiations and next-generation hardware remain exposed.
Retail prices depend on far more than factory bookings: existing inventory, contract cycles, retailer markups, tariffs, logistics, and whether device makers quietly reduce memory specs.
Worth noting: NAND flash—the storage in SSDs—is a separate category. A DRAM and HBM capacity report says nothing directly about storage supply.
The key question remains whether suppliers can add meaningful output before 2027 arrives, or whether AI customers begin trimming commitments and release capacity back into the market.