GPU bubbles get the headlines, but 2026's most universal AI cost increase never touched a GPU. It landed on the DIMMs. Memory is where the AI buildout stops being somebody else's procurement problem, because DRAM is fungible: the same fabs, the same process nodes, the same wafers that make HBM stacks for accelerators make the DDR5 in a desktop, a smartphone and a NAS. When one buyer class — AI — is willing to pay several times more per bit for the same capacity, the supply does not stretch. It moves. Everything this forum has written about HBM's cost premium was the server-side version of a story whose PC-side chapter is being written now.

The mechanism is stated in as many words by the industry's own pricing authority. TrendForce's 4Q25 DRAM release (September 24, 2025) attributes the quarter's price hikes to a single cause: "the three major DRAM suppliers continue to allocate advanced process capacity primarily to high-end server DRAM and HBM, crowding out capacity for PC, mobile, and consumer applications." The published result: conventional DRAM contract prices up 8–13% QoQ in 4Q25, 13–18% including HBM — and DDR4, the legacy product being run off the newest nodes fastest, having nearly doubled in 3Q25 alone. Note the shape of that sentence: nothing in it is about demand collapsing or booming on the consumer side. The consumer price rose because the consumer supply left.

2026: the compounding year

What makes the 2026 episode unusual is not the direction of the move — memory cycles have always existed — but the magnitude and persistence once AI demand met a supply base that had already been redirected. TrendForce's quarterly forecast series for server DDR5 contract prices through 2026 reads: +93–98% in Q1, +53–58% in Q2, +13–18% in Q3, +3–8% in Q4. Our arithmetic on that published series: compounding the midpoints gives roughly 3.3× within calendar 2026 alone, and TrendForce itself frames the full-year move as a projected ~270% cumulative increase in DRAM contract prices for 2026, on top of ~64% cumulative across 2H2025. The firm also expects 4Q26 conventional DRAM to rise another 10–15% (NAND 15–20%), and — the line item that should reprice every 2027 capacity model — memory rising toward 68% of CSP capital expenditure by 2027, with HBM and RDIMM alone taking 51% of DRAM supply bits in 2026.

The spot market, which reacts before contract negotiations do, moved even harder. TrendForce spot-price tracking reported by Chinese financial media in December 2025 put DDR5 2Gbx8 die spot prices up 307% and DDR4 1Gbx8 up 158% in roughly the four months from early September 2025. Spot runs hotter than contract and mean-reverts faster — but the direction of travel across both series was the same, and the contract series has kept climbing for five consecutive quarters since. A supply shock that lasts this long stops being a shock and becomes a repricing.

The corporate tell: Micron walks away from its own brand

If you want to know how a supplier values its consumer business versus its AI business, do not read the marketing — read the exit. On December 3, 2025, Micron announced it would wind down Crucial, the consumer memory and storage brand it had operated for 29 years — the brand that pioneered retail DRAM upgrades in 1996. The company's own announcement stated consumer products would ship "through February 2026, with warranty and support continuing," and its commercial chief Sumit Sadana tied the decision directly to the displacement mechanism: data center growth driven by AI created demand that made consumer supply the thing worth giving up. Micron said it would redeploy the resources toward large strategic customers in faster-growing areas.

The same company's earnings materials supply the "toward" in that sentence. Micron's fiscal Q4 2025 prepared remarks — published on its own investor site — state: "In fiscal Q4, our HBM revenue grew to nearly $2 billion, implying an annualized run rate of nearly $8 billion," within a fiscal year where the data center business hit a record 56% of total company revenue. A memory maker looking at ~$8B/yr of HBM demand with pricing agreements covering most of its calendar-2026 HBM3E supply, against a consumer business being squeezed for capacity by its own AI allocation, did arithmetic. The consumer brand lost. That is what "AI is taking the memory" looks like at the corporate level: not a shortage headline, a discontinued checkout page.

Who pays, and how much

Run the displacement through each buyer class (our arithmetic throughout; the underlying multipliers are TrendForce's published series). The cloud buyer pays through the quote. Server DDR5 RDIMMs are a line item in every GPU node's BOM; a ~3.3× server-DRAM move inside one year shows up either as higher rental rates or as thinner margins per accelerated hour — and TrendForce's expectation that memory reaches two-thirds of CSP capex by 2027 means memory is on its way to being the swing cost in AI data centers, competing with the accelerators themselves. The PC buyer pays through the shelf. OEMs are cutting DRAM per machine and reducing SSD capacities to hold BOM costs, per TrendForce's 4Q26 note, because the component inside the laptop rose faster than the laptop's price could. The phone buyer pays through allocations — TrendForce expects mobile DRAM supply to stay tight into 2027 as capacity keeps shifting to server. And the AI infrastructure cost stack gains a new top-heavy layer: inference is already the real compute bill, and now the memory that inference is bottlenecked on has its own repricing running in parallel.

The home-lab reader is not exempt, and the Homelab section's math changes shape because of it. The used RTX 3090 argument assumed platform costs were flat — but a 128GB DDR4/DDR5 server board build, the classic "used enterprise RAM for cheap VRAM-adjacent capacity" play, is precisely the legacy-product segment TrendForce reported as rising fastest (DDR4 nearly doubled in 3Q25). The cloud vs home TCO spreadsheet from this forum now needs a memory line with a date stamp on it, the same way its electricity line does. The honest 2026 version of the home-server argument is: build with what you already own, because the upgrade path just repriced.

When does it end

Two clocks govern the exit, and they point in opposite directions. The bear-clock is supply response: memory makers have announced new capacity, and TrendForce's own framing — that price growth "moderates" through 2H26 as new fab output lands — is the classic end-of-cycle shape. The bull-clock is demand structure: enterprise SSD bit demand up 80%+ in 2026 on AI inference deployments, agentic workloads driving general-server RDIMM demand, and TrendForce's expectation that PC DRAM supply shrinks again in 2027 as more capacity shifts to server applications. A shortage caused by real, contracted, multi-year demand does not mean-revert the way a demand panic does; it clears by new fabs, which take years — the same rhythm this forum documented in the power queue and the transformer queue. Memory is the third queue, and it is the one with no way to substitute around.

That is the wider point. Every layer of the AI infrastructure stack has now had its repricing moment: silicon (packaging), memory (HBM, and now commodity DRAM), power (queues, transformers, water), network (the fabric), and demand-side (the agent token bill). The pattern across all of them is identical — the naive price is the old price, and the gap between the old price and the invoice is where capacity actually went. Anyone building across this stack, centralized or distributed (omc.network), should read the 2026 memory series the way buyers read GPU rates: as a published, sourced, quarterly-updated meter that will not go back to zero when the news cycle moves on. Memory learned what compute learned before it. When a scarce input gets a richer customer, the old customer pays the difference — politely, quarterly, and in writing.