Summary: A DIGITIMES report relayed by TweakTown and IGN says five-year purchasing agreements have committed Samsung, SK Hynix, and Micron DRAM and HBM capacity through 2027, largely for AI data centers. None of the three manufacturers has confirmed the report, although a year of component-price increases supports its direction. The reported squeeze already appears in SSD, console, and PC costs, so hardware budgets should treat 2026 memory inflation as a planning assumption rather than expect normal prices.
One more threshold has fallen in the memory crunch. Word of a DIGITIMES report went first to TweakTown, and IGN followed this week: the DRAM and HBM capacity of Samsung, SK Hynix, and Micron is committed through 2027 under purchasing agreements lasting five years. The report’s own phrasing is that “no further supply planned” exists for 2027, and the 2028 picture is no cheerier. Whoever is absorbing that capacity, it is largely not the PC and phone makers. It is AI firms.
A qualifier belongs at the top: confirmation has come from none of the three manufacturers. Twelve months of component pricing, though, has supplied the supporting evidence.
What the Report Actually Claims
The structure of the contracts is where the weight sits: long-term agreements spanning five years, signed against memory that does not exist yet. Chip buying is too narrow a description. What happens instead is advance reservation of fab output, mirroring how long-term GPU compute contracts work. Timing separates the two cases — a GPU bought today ships inside the year, while 2027 allocation is being committed before any wafer holds a scheduling slot. Paying up to erase uncertainty is the new posture, not pricing it in. Over on NAND, demand is also heading up, but that market carries more suppliers and its output is not yet fully eaten. IGN’s takeaway on where this goes: absent fresh capacity, retail prices should keep rising, and relief is nowhere on the horizon.
Who Is Doing the Buying
AI data centers. HBM sits in every AI accelerator card; high-capacity DDR5 fills every AI server; and the quantities ordered are sized for data centers rather than desktops. With training and inference clusters growing on a quarterly rhythm, the best output naturally flows toward the longest contracts and the richest bidders. HBM is where the squeeze bites hardest, since its advanced packaging lines are slow to bring up — so each new generation of AI hardware turns into another multi-year HBM pledge. An AI company treats securing 2027 HBM exactly as it treats securing GPUs or electricity: silicon, memory, land, power — every ingredient of compute expansion now trades like futures, where the choice is to buy early or to queue.
Consumers Are Footing the Bill Already
The numbers are concrete, not abstract. IGN’s arithmetic: an SN7100 1TB PCIe 4 SSD from Western Digital was about $110 in January and lists at $189 now, a 52% climb. This month the Xbox Series X went up in price. Memory costs pushed Valve’s Steam Machine to roughly $750 at launch, above Valve’s original target. What the reviewer wanted was modest — a decent RAM kit for under $500 that she could buy again. Once the dull, invisible line item of every build, memory is now the likeliest candidate to blow the budget. Console gamers and PC builders are, hike by hike, subsidizing AI infrastructure.
CXMT and the OEM Fallback in China
Procurement maps are being redrawn by the shortage as well. Nikkei Asia has reported that HP, Asus, and Acer now take memory chips from CXMT (ChangXin Memory Technologies), the Chinese maker, for products sold in China. Guru3D pushes the point further: CXMT’s DRAM capacity is itself reported as sold out through 2027, while the major OEMs compete for allocation. With the big three’s standard-memory output tied up by AI contracts, a supply source once filed under second-choice is turning up in mainstream lines from first-tier brands — geopolitical and supply-chain signal in one.
What Developers and Buyers Should Be Doing
Three suggestions. First, write DRAM and NAND inflation into the hardware budget: whether the plan involves workstations, boxes for local inference, or holding out for cheaper cloud pricing, do not count on memory prices returning to normal at any point in 2026. Where a project leans on machines with large RAM — beefy CI runners, in-memory caches, local vector databases — the RAM cost deserves its own line item, revisited every quarter. Second, secure hardware procurement early: in GPU server tenders and enterprise refresh cycles alike, memory’s share of cost will keep growing. Third, watch the increases travel down the stack: phones, consoles, NAS devices, and cloud instance pricing will each reflect them sooner or later; only the timing is open.