AI infrastructure has bought up the memory industry's output. Consumer DRAM and NAND were the residual claimant, and they repriced — 3–4× in eighteen months. Finished PCs absorb only part of that, but the direction is set: the base case is roughly +30% on a like-for-like machine by 2027 versus 2025, with the first real relief arriving in 2028–2029, not next year.
Two different things are being confused in the coverage. The rate of memory price increase has collapsed since Q1. The level has not come down at all — and everything built after Q1 2026 is priced off that level.
| Quarter | DRAM contract, QoQ | NAND / SSD contract, QoQ | Source |
|---|---|---|---|
| Q1 2026 | +110% | +147% | Counterpoint Memory Price Tracker (8GB DDR4 SoDIMM / 1TB NVMe) |
| Q2 2026 | +58–63% | +70–75% | TrendForce, 31 Mar 2026 |
| Q3 2026 F | +13–18% | +10–15% | TrendForce via Tom's Hardware |
| Item | Multiple vs mid-2025 | Note |
|---|---|---|
| 32 GB DDR5 kit | 4.00× | aftermarket desktop memory |
| 32 GB DDR4 kit | 3.00× | older platforms, also squeezed |
| M.2 NVMe SSD | 2.00× | mainstream 1-2 TB |
| Finished notebook | 1.25× | like-for-like config, EU retail |
Components are up 200–300%. Finished computers are up 20–30%. That gap is vendor margin and configuration cuts absorbing the shock — and both of those buffers are nearly spent. Omdia measures the added memory and storage cost of a mainstream PC at $90–165 per unit since Q1 2025.
A memory fab has one wafer budget. HBM for AI accelerators consumes far more wafer per usable bit than ordinary DRAM, and it sells at a far higher margin. Every wafer moved to HBM removes several wafers worth of consumer memory from the market.
Hyperscaler capex up 98% in 2026 and another 50% in 2027.
DRAM and NAND move from 47% of that capex in 2026 to 68% in 2027.
HBM plus server RDIMM take 51% of all DRAM bit supply in 2026.
Suppliers price catch-up increases into PC, phone and graphics memory.
Vendors cut configurations first, then raise list prices 15–30%.
Cloud providers have moved to long-term supply agreements, booking capacity years ahead. Server DRAM contract prices rose a cumulative 64% in 2H25 and roughly 270% across 2026.
SK hynix's Yongin Y2 starts mid-2029; its M17 NAND line late 2028; Samsung's Yongin site in 2029. Capex is up nearly 340% between 2024 and 2027 and it still cannot arrive sooner.
The Q3 slowdown to +13–18% came from device makers refusing to pay, not from more supply. IDC expects Q4 2026 PC shipments down about 20% year on year.
Index of what the same specification costs at EU retail, 2025 = 100. 2026 is effectively settled; the fork is 2027–2029 and it turns entirely on when new wafer capacity meets AI demand.
| Scenario | Probability | 2026 | 2027 | 2028 | 2029 | Cumulative vs 2025 |
|---|---|---|---|---|---|---|
| Relief | 25% | 115 | 120 | 110 | 97 | −3% |
| Base | 50% | 118 | 130 | 132 | 124 | +24% |
| Prolonged | 25% | 122 | 144 | 155 | 155 | +55% |
Hyperscaler spending decelerates in 2027, NAND surplus arrives on schedule, and the $500bn+ of new Korean and US fab investment lands sooner than guided. Prices round-trip to roughly 2025 levels by 2029.
DRAM stays tight through 2027, NAND eases from 2H27, and new DRAM output only bites in late 2028. Peak PC pricing in 2028, then a partial retracement. Prices never return to 2025 — IDC says the same.
AI demand keeps absorbing every incremental wafer; suppliers pause expansion to protect pricing. Samsung and SK hynix have warned publicly that the shortage could last until 2027 and beyond, and SK hynix sees no quick recovery before 2028. The sub-€500 laptop disappears entirely.
In the base case a computer bought in mid-2027 costs about 10–12% more than the same machine today, and about 30% more than in 2025. Waiting only pays in the relief branch, and even there the payoff lands in 2029 — not next year.
Memory is a much larger share of the bill of materials on a cheap machine than an expensive one. That inverts the usual intuition: the budget end is hit hardest, and in places it is simply being withdrawn.
| Segment | Cumulative 2025→2027, base case | Driver |
|---|---|---|
| Self-build desktop | +55% | RAM + SSD + GPU all hit |
| Entry notebook (under €500) | +42% | tier is being discontinued |
| Gaming laptop (32 GB) | +36% | GDDR7 + high memory count |
| Mainstream notebook 16/512 | +30% | the volume segment |
| Business ultrabook | +18% | memory a smaller share of BOM |
| Apple / integrated premium | +14% | long-term supply contracts |
| What €800 bought / buys | Mid-2025 | August 2026 | 2027, base case |
|---|---|---|---|
| 15" mainstream notebook | 16 GB / 512 GB, current CPU | 8–16 GB / 512 GB | 8 GB / 256 GB, prior-gen CPU |
| Same machine, 16/512 | ≈ €700 | ≈ €880 | ≈ €960 |
| 32 GB DDR5 desktop kit | ≈ €85–95 | > €340 | ≈ €370 |
| 2 TB NVMe SSD | ≈ €110 | ≈ €220 | ≈ €235 |
| 8 TB SSD with DRAM cache | ≈ €450 | €1,029 minimum | still rising fastest |
Component prices for mid-2025 and August 2026 are anchored to European retail tracking; the 2027 column and the notebook rows are modelled. Prices include VAT and reflect euro-area listings.
The first response to a cost shock is not a higher price — it is a quieter configuration. Vendors have cut base RAM and storage, and some are now shipping pre-built desktops with no memory installed at all. A machine at last year's price this year is almost certainly a smaller machine.
Every forecast above reduces to this calendar. Memory capex is up nearly 340% between 2024 and 2027, but a fab announced today produces bits three to four years later.
DRAM contract prices +13–18% QoQ, NAND +10–15%, down from roughly +60% in Q2. Cause: buyers refusing to pay, not more supply.
IDC expects PC shipments down about 20% year on year, with full-year 2026 at −11.3% and ASPs up 17%. Total market value still rises to roughly $274bn.
TrendForce sees the DRAM sufficiency ratio deteriorating further while NAND flips to surplus in 2H27. HBM contract prices could still rise 70–140%. Notebook and smartphone shipments decline again.
The earliest new capacity comes online — but TrendForce is explicit that substantial output contributions are not expected until 2028.
$13.5bn NAND facility begins operations. Storage is the first component to normalise.
SK hynix Y2 and Samsung's Yongin site start production. This is also the year analysts flag as peak risk for an AI-capex correction: enough memory, packaging and energised data-centre capacity arrives at once to test whether demand was real.
The base case gives no reward for waiting through 2027. It gives a large reward for buying the right configuration once, and for not paying the aftermarket premium on memory twice.
Two of three scenarios have 2027 more expensive than today, and the third only pays off in 2029. There is no branch in which mid-2027 is a better moment than mid-2026 to buy a computer.
Aftermarket DDR5 is the single most inflated item on the market at 4× its 2025 price, and adding RAM later now costs more than it did to have it factory-fitted a year ago. On soldered-memory laptops this is irreversible.
The inverse of the laptop rule. On a self-build, the RAM and SSD are now the expensive parts and the CPU and board are the cheap ones. Reusing a working 32 GB kit removes the worst line item from the build.
With the sub-€500 new-laptop tier being withdrawn, the European secondary market has become the default answer at that budget rather than a compromise. A 2023–2024 business ultrabook with 16 GB is the value position of this cycle.
Refresh cycles are the lever, not the purchase price. Extending a fleet from three to four years across this window avoids the peak entirely, and costs less than the 30% premium on replacing it in 2027–2028. If a refresh is unavoidable, contract the configuration and the price now — vendors are repricing quarterly.
The base case is a bet on the AI capex cycle staying intact through 2027. Here is what to watch, in order of how much it would move the answer.
| Signal | What it would mean | Effect on PC prices | Direction |
|---|---|---|---|
| Memory prices fall while bit shipments stall | Scarcity premium and volume demand both breaking — the classic top of a memory cycle | Sharp relief within two to three quarters | Cheaper |
| Hyperscaler capex guidance cut for 2027 | The 50% capex growth assumption fails; wafers return to consumer parts | Relief scenario becomes the base case | Cheaper |
| Fab ramps pulled forward from 2H27 | Korean and US expansion delivering ahead of guidance | Peak arrives in 2027 instead of 2028 | Cheaper |
| Suppliers pause expansion to defend pricing | Memory makers behaving as a disciplined oligopoly rather than racing for share | Prolonged scenario, high prices past 2029 | Dearer |
| HBM contract prices rise another 70–140% in 2027 | Wafer reallocation to HBM accelerates rather than stabilises | Consumer memory squeezed further | Dearer |
| New tariffs or export controls on memory or finished PCs | A cost layer entirely separate from the shortage | Additive to every scenario above | Dearer |
| Euro weakness against the dollar and won | Component costs are dollar-denominated; EU retail carries the FX on top | Amplifies or dampens all of the above for euro-area buyers | Either |
Published estimates of how long this lasts range from six months to a decade. Samsung and SK hynix have warned the crunch could last until 2027 and beyond, with Notebookcheck reporting SK hynix sees no quick recovery before 2028; several analysts expect normalisation between mid-2027 and 2028. The scenario probabilities on this page are the author's weighting of that spread, not a consensus figure.
Figures attributed to IDC, TrendForce, Omdia, Counterpoint, Gartner and 3DCenter are as published. Index levels, scenario probabilities and segment splits are the author's model built on those inputs and are labelled as such wherever they appear.