The memory shortage is not a component-cost story, it is a transfer. Fewer PCs will ship in 2026 than in any recent year while total PC market value rises to $274 billion, because DRAM capacity has been reallocated to HBM for AI infrastructure. This piece traces the causal chain from cleanroom allocation to retail price, sizes the forward demand with one number, and explains why a structural shortage inverts the usual advice to wait.
Two facts about the 2026 PC market, both from IDC, published together. Global PC shipments will fall 11.3% this year — from 284.7 million units in 2025 to 252.53 million in 2026, a decline of 32.17 million machines. And the total value of the PC market will rise about 1.6%, to $274 billion.
Fewer computers, more money. Those two numbers are not in tension. They are the same fact viewed from opposite ends of a supply chain that has been quietly reorganised around something that is not the PC.
Everyone has now written the sentence AI is making RAM expensive. It is true and it is not a story. What makes it a story is how badly the people whose job is forecasting this market have been caught out.
IDC's 2026 global PC shipment forecast has moved three times in nine months, always in the same direction: −2.4% in November 2025, −8.9% in January 2026, and −11.3% now. A forecast that gets refined moves in both directions as new information arrives. A forecast that moves one way three times is not being refined. It is an analyst house repeatedly discovering that a shock is larger than it thought.
−2.4% → −8.9% → −11.3%. Three revisions, one direction. The memory reallocation is still being priced in, not absorbed.
That direction of travel is the actionable part. It says the market has not yet found the bottom of this, and that each round of new information has made the picture worse rather than better. Gartner, working independently, has published its own downward revisions on surging memory costs. Estimates differ in magnitude — more on that below — but not in sign.
The mechanism is physical and it is the reason this does not resolve itself. Samsung, SK Hynix and Micron account for the great majority of global DRAM output. High-bandwidth memory — the stacked DRAM that sits alongside AI accelerators and largely determines how fast they can be fed — is manufactured using the same fabrication resources as the conventional DRAM in a laptop. Same fabs, same cleanrooms, substantially the same process steps, with HBM consuming more wafer area per usable bit because of the stacking and the yield loss that comes with it.
So every wafer allocated to HBM is a wafer not allocated to consumer DRAM, and HBM carries margins that consumer DRAM cannot approach. Three companies made a rational commercial decision about where to point their capacity, and the consequence arrives at retail as a price.
TrendForce's contract pricing shows the size of the move. Conventional DRAM contract prices rose a record 90–95% quarter over quarter in Q1 2026, then a further 58–63% in Q2 — deceleration, but from an extraordinary base. PC DDR5 specifically was projected up 43–48% in Q2. NAND flash contract prices rose roughly 60% in Q1 and a further 70–75% in Q2, accelerating while DRAM slowed. Those are contract prices to large buyers, not spot prices; they are what OEMs actually pay, and they show up in retail configurations one product cycle later.
On 26 August, Nvidia disclosed in its Q2 FY2027 CFO commentary that its supply commitments had more than doubled in a single quarter — from $119 billion to $279 billion — primarily related to the procurement of memory, largely tied to the forthcoming Vera Rubin platform. The company's framing is that it secured supply before prices moved, which its CFO has suggested competitors did not.
That number is not this article's subject and the earnings are not the point. It is a measurement. One buyer, in one quarter, locked up $160 billion of incremental forward supply commitments concentrated in memory. Whatever else is true about AI capex debates, that is the scale of demand standing between a fab's output and a consumer's shopping basket, and it is contracted rather than speculative.
Here is the argument the two headline numbers actually support. The AI build-out is not simply raising component costs as a side effect. It is transferring consumer surplus into the AI supply chain in a form that never appears on anyone's AI spending line.
When a data centre operator buys accelerators, that spending is visible: it is in the capex line, the analyst models, the bubble arguments. When a household pays $200 more for a laptop with the same specification as last year's, that is also money flowing to the AI build-out — it is the price of the memory that was not made because HBM was made instead — but it is booked as consumer electronics inflation. Nobody buying a laptop in 2026 thinks of themselves as funding a data centre. In a real and traceable sense, they are.
“Thirty-two million fewer PCs will ship in 2026, and the market will be worth more money. The gap between those two facts is the transfer.”
This is the part that changes what a reader should do, and it turns on a distinction most coverage skips.
A cyclical shortage is a demand spike hitting fixed capacity. It resolves when demand normalises or when new fabs come online, and the correct response is to wait, because waiting is free and prices revert. A structural shortage is a deliberate reallocation of existing capacity by a small number of suppliers toward a higher-margin product. It does not resolve when demand blips, because the reallocation was not a response to a blip. IDC's own framing has the shortage persisting through 2027, and industry analysis broadly expects it to run beyond that — a forecast, not a fact, but a consistently held one.
In a structural shortage, waiting is the expensive choice. That is the opposite of the default instinct, and it produces three concrete consequences for anyone buying hardware in the next twelve months:
One honest caveat, and it strengthens rather than weakens the case. The smartphone side of this is genuinely contested. IDC-attributed coverage has circulated figures around −12.9% and −13% for 2026 global smartphone shipments; an earlier round of analysis had −5%. Gartner has published separately with its own numbers. These are not reconcilable by averaging — they are different houses, different methodologies and different publication dates, and the honest move is to name one and date it rather than blend them.
But the disagreement is itself part of the argument. The reason estimates diverge this much is the same reason IDC has revised three times: this shock is moving faster than the forecasting apparatus built to measure it, and everyone measuring it is chasing. When the analysts cannot agree on the magnitude and only on the direction, the appropriate posture for a buyer is not to wait for consensus. It is to assume the direction and act on it.
Because the same fabrication capacity that makes consumer DRAM also makes high-bandwidth memory for AI accelerators, and the three companies that dominate DRAM production — Samsung, SK Hynix and Micron — have shifted capacity toward HBM and enterprise DDR5, where margins are far higher. Every HBM wafer is a consumer DRAM wafer not produced. TrendForce recorded conventional DRAM contract prices rising 90–95% quarter over quarter in Q1 2026 and a further 58–63% in Q2.
DrafterDaily traces where the money in technology actually goes — from cleanroom allocation to the price on the shelf. One clear read a day.
Get the daily briefCVE-2026-0768 is an unauthenticated root RCE in Langflow. It was disclosed in January, the fix has shipped through seven releases, and attackers are hitting it in September — because low-code AI middleware became critical infrastructure without acquiring a patch owner.
Opposition to data centres is not a referendum on AI. It is a cost-allocation dispute — and PJM's capacity auction is the mechanism that turned an abstract argument into a number on 67 million households' bills.
OpenAI's first published benchmarks for its custom inference chip are real, from a public benchmark, and normalised on nameplate TDP rather than measured power. OpenAI disclosed that itself, in a sentence nobody is quoting.