The AI memory crisis: why RAM prices doubled in 2026

01 The price shock in numbers

Walk into a store or open an order page for a memory upgrade in September 2026 and the number next to the same DDR5 kit that cost $60 a year ago is difficult to rationalize. Contract pricing tracked by the memory trade press roughly doubled across mainstream capacity tiers between mid-2025 and late 2026, and spot quotes for some densities moved faster than contract terms could reset. The chart below uses representative mid-range figures to show the shape of the move; exact quotes vary by supplier, volume, and week.

This is not the familiar slow slide of consumer electronics pricing. Memory has always been cyclical — glut and shortage alternate as suppliers misjudge demand — but the 2025-2026 leg of the cycle broke the usual pattern in one respect: the new demand did not come from the customers memory makers traditionally serve. It came from data centers building racks of AI accelerators, each of which needs bandwidth-class memory beside every GPU.

The result is a market where PC and phone buyers are competing, indirectly, with companies spending hundreds of billions of dollars on AI infrastructure. When a hyperscaler's procurement order is measured in hundreds of thousands of modules and is signed a year ahead, the walk-in retail customer simply inherits whatever pricing the contract market leaves behind.

02 How HBM eats the wafer supply

The mechanism that makes AI demand so invasive is high-bandwidth memory, or HBM. HBM stacks DRAM dies vertically and connects them with through-silicon vias, producing a memory package with a massively wider data path than a conventional DIMM. Every AI accelerator sold today pairs its compute silicon with several HBM stacks; the memory is not an accessory to the GPU business, it is half of the physical product.

The catch is yield mathematics. Stacking eight or twelve dies with through-silicon vias multiplies the effect of any single defective die, and the thicker assembled packages take more room in the fab. Industry analyses commonly estimate that producing a bit of HBM consumes roughly two to three times the wafer capacity of the same bit in standard DDR5. Every exabyte of HBM a supplier commits to AI customers therefore displaces several exabytes of conventional memory that could have served PCs and phones.

That displacement is the core of the 2026 shortage. The three commodity DRAM suppliers did not stop making DDR5 — they shifted their most advanced capacity toward the product AI buyers pay a premium for. Conventional memory became the residual: same fabs, less of them allocated to you, at prices that clear whatever demand remains.

03 The three-layer demand stack

Memory demand in 2026 stacks three layers with different price sensitivities. The bottom layer, and the one driving the crisis, is hyperscale AI capex: the major cloud operators' combined infrastructure budgets — several hundred billion dollars a year in aggregate — include memory line items that are effectively price-inelastic, because a shortage of HBM or high-capacity DRAM delays revenue-generating AI capacity.

The middle layer is the device market: PCs, smartphones, and tablets. Device makers cannot drop memory from their products — baseline configurations keep climbing as software and on-device AI features demand more RAM — so they absorb higher prices and pass them through. Several PC OEMs raised configuration prices in 2026 explicitly citing memory costs, and phone makers quietly slowed their march toward bigger base storage tiers.

The top layer is automotive and industrial, which buys on multi-year qualification cycles and cannot switch suppliers quickly. Car makers designing memory into 2027-2028 platforms contracted early at painful prices rather than risk line stoppages. Three layers of demand, none of them willing to yield volume, met a wafer supply that could not expand quickly. Something had to give — and the mechanism that gave was price.

04 Who pays first: OEMs, phones, and the DIY market

The sequencing of who absorbs the increase first follows contract structure. Long-term agreements signed by hyperscalers locked in supply early at prices they can plan around. Device OEMs renegotiate annually, so their 2026 builds absorbed the first full repricing — visible in higher PC configuration prices and in the conspicuously absent price cuts that normally accompany mid-year phone launches.

The do-it-yourself PC market pays the most in percentage terms. Retail and e-tail pricing floats freely with spot sentiment, and kits that sat under $70 in mid-2025 peaked well past $150 during 2026, with premium low-latency kits tripling. Builders who could delay upgrades did; system integrators passed costs through on prebuilt machines, where memory lines moved from rounding errors to line-item disclosures.

There is also a quieter substitution effect. Buyers priced out of 32GB configurations dropped to 16GB; buyers priced out of 16GB dropped to 8GB or deferred entirely. The industry describes this as demand destruction, and it is the classic mechanism by which memory cycles eventually end — but in 2026 the AI layer's demand was large enough that PC-side demand destruction barely dented the shortage.

05 Why suppliers are not rushing to add capacity

The intuitive response to shortage is to build more fabs, and memory suppliers are indeed expanding — but deliberately slowly, and their discipline has structural reasons behind it. The 2022-2023 downturn was brutal: DRAM prices fell below cost for several quarters, Samsung posted memory losses, and every supplier emerged committed to capacity discipline. The memory business had spent a decade proving that uncoordinated expansion produces gluts that destroy profitability for everyone.

HBM changes the incentive calculus further. It carries margins well above commodity DRAM, and suppliers have committed their leading-edge capacity to it for years. SK hynix, the HBM leader, reported effectively sold-out advanced capacity into 2027; Samsung and Micron publicly prioritized HBM qualification cycles with GPU customers over commodity bit growth. From the suppliers' perspective, the shortage is not a problem — it is the pricing environment they spent two decades trying to create.

Greenfield fabs take two to three years to ramp memory production, and equipment lead times extended through 2026. Even a supplier that broke ground today would land capacity into a 2028 market whose demand profile nobody can confidently predict. Under those conditions, conservative capital expenditure is rational for each supplier individually — even though it extends the shortage collectively.

06 Second-order effects: SSDs, consoles, and prebuilts

The squeeze is not confined to DRAM. NAND flash rides the same fab-capacity logic: suppliers rebalanced production toward enterprise SSDs for AI data storage, and consumer NVMe prices, which spent years in free fall, turned sharply upward in 2026. The per-terabyte price of consumer storage — a number that moved in only one direction for a decade — rose for four consecutive quarters.

Console economics came under pressure because both major platform holders sell hardware near cost and monetize software. Higher memory and storage bills either compress margins further or force price increases late in a console generation, and both Sony and Microsoft faced that decision publicly during 2026. Prebuilt PC vendors, meanwhile, began itemizing memory costs in marketing copy — a tell that component pricing had become a customer-facing issue rather than an internal procurement detail.

The phone segment shows the subtler effect. Flagship base storage tiers had climbed steadily from 128GB to 256GB and beyond as NAND got cheap; that march paused in 2026, with several launches holding last year's storage at last year's price while quietly raising prices on higher tiers. Consumers experience this not as a price increase but as the absence of the upgrade they expected — the least visible and most widespread cost of the memory crisis.

07 What could normalize prices in 2027

Crisis narratives are easy to write and hard to falsify, so the bear case deserves its limits stated. Several forces could normalize memory pricing through 2027. Capacity added during the 2024-2026 investment wave reaches volume production on a lag; suppliers have announced node transitions and new fabs whose output lands precisely in the 2027-2028 window. If AI infrastructure growth decelerates even modestly from its 2025-2026 pace, the marginal bit of capacity would land on a far softer market.

Demand-side elasticity also cuts both ways. The PC and phone markets demonstrated in 2026 that buyers defer purchases above certain price thresholds, and deferred demand becomes a cohort of future supply absorption. Memory is durable; a buyer who skips 2026 returns in 2027 to the same shelf. Historical memory cycles have consistently overshot in both directions, and the same suppliers now rationing capacity have a record of colliding with each other's expansions at the top.

The honest forecast is a plateau with wide error bars: contract prices likely stay elevated through early 2027 as HBM commitments hold, then face real downside risk as conventional-capacity additions land and AI procurement normalizes from breakneck to merely large. Buyers planning upgrades in late 2027 may find the market swinging back toward the oversupply side of the cycle with little warning — which is, paradoxically, exactly what memory suppliers fear most.