Your desktop RAM cost eight times more this year than last, and no AI system will ever touch it. This is how a decision made inside three companies repriced every device you own, and what it means for the businesses caught underneath it.
Samsung, SK hynix and Micron produce roughly 95% of the world's DRAM. In 2024 they began shifting wafer capacity toward high-bandwidth memory, the stacked DRAM that sits beside an AI accelerator. That decision was rational and, at the time, uncontroversial.
It has since raised the benchmark DDR4 chip price roughly eightfold, approximately doubled what a consumer pays for a retail memory kit, put 8GB laptops back on shelves at $1,299, added $100 to a games console, taken the Mac Studio's largest memory option off the market entirely, and made a hosting provider raise memory add-ons by 575%. None of those products run AI workloads. They simply share a supply chain with something that does.
The 8.3× below is the bare DDR4 chip, the industry's benchmark commodity part. It is not what a shopper pays. A memory module carries eight or sixteen of those chips on a board with a PCB, heatspreaders, binning, packaging, distribution and retail margin — costs that did not move eightfold. Channel inventory built from older, cheaper chips also takes months to clear.
Checked against retail: 64GB DDR4 kits that sold for around $160 two years ago are near $300, and 64GB DDR5 kits that were about $480 are near $1,000 — both roughly 2×, against 8.3× on the chip. Two independent observations landing on the same multiple suggests the gap is structural rather than a quirk of one product. Anyone quoting the chip figure as a consumer price increase is overstating it by about four times.
HBM consumes roughly 2.3 times more wafer area per bit than ordinary DRAM. That ratio is the entire story.
High-bandwidth memory will take 30% of DRAM wafer input by end-2027 while producing only 13% of the bits. Every point of capacity moved to HBM therefore destroys more than twice that much commodity supply. This is why a product AI never touches gets repriced by AI demand.
Data centres will consume around 70% of all memory manufactured in 2026.
Everything else — every phone, laptop, console, car and television on earth — competes for the remaining 30%. Order fulfilment for smaller buyers ran at about 70% as early as October 2025, and pricing has since moved to hourly quotes.
HBM stopped being the most profitable use of a wafer, and the shortage still got worse.
TrendForce reports that HBM profitability fell below DDR5 64GB RDIMM in Q1 2026. Commodity server memory is now the better business per wafer — and HBM prices are expected to rise substantially in 2027 simply to stay competitive for allocation. The shortage has become self-reinforcing.
The energy bill is already arriving, and it has started changing elections.
PJM's capacity price rose eleven-fold, and PJM itself attributes nearly 5,100 MW of a 5,250 MW load increase to data centres. Two Democrats won Georgia Public Service Commission seats in November 2025 campaigning on power bills — the first Democratic statewide non-federal wins there in nearly twenty years.
A large share of the demand driving all of this may not exist.
When AEP Ohio required financial commitment, requested capacity fell from roughly 30,000 MW to 5,642 MW. ERCOT's 2030 pipeline shows 410,618 MW against 5,778 MW actually energised. In January 2026 PJM cut its own forecast, with executives saying demand projections had “bloated to levels that are irrational.”
Headline capex now overstates capacity, because a growing share of it is just the memory bill.
Microsoft's roughly $190B 2026 guidance “includes approximately $25 billion from the impact of higher component pricing.” Meta cited the same when raising its range. Spending is rising faster than compute is, which cuts against every model that treats capex as a proxy for capability.
The common assumption is that AI demand competes with consumer demand for the same chips. It does not. An H100 does not contain a stick of DDR5. The competition happens one level up, at the wafer.
DRAM fabs produce wafers, and those wafers are allocated to products: HBM stacks for accelerators, RDIMM for servers, LPDDR for phones, UDIMM for desktops. It is one pool. And HBM is a remarkably inefficient way to turn a wafer into bits — stacked dies with through-silicon vias, lower yields because a single bad die spoils a stack, and a large base die that stores nothing.
By the end of 2027, HBM will absorb 30% of wafer input to deliver 13% of bits — a penalty of roughly 2.3 times. Shift a tenth of an industry's capacity into HBM and you remove nearly a quarter of its commodity output. That is the whole mechanism, and it is why your desktop kit repriced.
After the 2022–23 memory glut — severe enough that Samsung cut production by roughly half — the industry stopped investing. There was little or no new capacity added through 2024 and most of 2025, and what capex there was went into stacking and packaging rather than raw wafer starts. The discipline lesson from the last bust was applied at precisely the wrong moment.
New fabs cost upwards of $15B and take eighteen months at best. Micron's Singapore HBM fab arrives in 2027; SK hynix's Cheongju plant in 2027; Samsung's Pyeongtaek expansion in 2028; Micron's New York complex reaches full production in 2030. Intel's chief executive put it bluntly in February 2026: “There's no relief until 2028.”
Spot markets moved further. A DDR4 8Gb chip — the industry's benchmark commodity part — averaged $5.07 in July 2025 and $42.14 in July 2026. Legacy parts fared worst of all as suppliers abandoned mature nodes: DDR4 16Gb now trades above DDR5 16Gb, an inversion nobody would have predicted, and the shortage has reached DDR3 and even DDR2.
The suppliers are candid about the gap. Micron's chief executive said in December 2025 the company could meet only “half to two-thirds of demand.” SK Group's chairman put wafer supply about 20% behind demand and expects shortage until 2030. SK hynix has said 2027 will be the worst year. In June 2026 Micron said plainly that it has no idea when the crisis ends — having just signed roughly $100B in long-term supply agreements.
This is the part that separates this story from most technology reporting: the costs are not abstract or forthcoming. They have already been paid, by people who mostly do not know why.
| What | Move | Detail |
|---|---|---|
| HP bill of materials | 15–18% → 35% | Memory cost roughly doubled sequentially; CFO expects further increases |
| Hetzner cloud servers | +30–43% | Memory add-ons up to +575%, applied to existing customers |
| PlayStation 5 | to $649 | Two increases; memory is ~35% of console cost |
| Xbox | +$100–150 | Third rise in two years; 2TB model discontinued |
| Meta Quest 3 | to $600 | Explicitly attributed to the AI-driven RAM shortage |
| Apple, across the range | +15–20% | Mac mini $599 → $799; Mac Studio $1,999 → $2,499; iPad, HomePod, Apple TV all raised |
| Laptop memory | 16GB → 8GB | Dell XPS 13 at $699, and a $1,299 Surface for Business, both 8GB |
| Framework laptop memory | nearly doubled overnight | CEO said absorbing supplier increases would put “our ability to operate at real financial risk” |
| Micro Center | abandoned price tags | Moved memory to spot pricing in November 2025 |
The market-level damage is now visible in shipments. IDC expects global PC shipments to fall 11.3% in 2026, with average selling prices up 17% and no meaningful relief before the end of 2027. Smartphones face their largest decline on record. Counterpoint reports that memory prices have risen more than fourfold since Q4 2025 and that this “single component has eroded the profit margins of most consumer electronics players.”
The market has split into roughly 100 buyers with allocation leverage — hyperscalers, Apple, Samsung, major automakers — and more than 190,000 small and mid-size firms competing for what is left. Suppliers increasingly require prepayment before confirming orders, which is a working-capital test small integrators fail. Quotes now move hourly, and a firm that cannot pay immediately can see its price rise within minutes.
IDC expects white-box builders and lower-tier vendors to bear the heaviest burden. Apple raising prices 18% is a headline; a two-person system integrator being quoted a different price every hour is the actual story.
One consequence nobody planned for: the shortage created its own demand shock, in a product category that had nothing to do with AI when it was designed.
Apple's unified memory architecture — adopted in 2020 for power efficiency and bandwidth, and criticised at the time for soldered, non-upgradeable RAM — means the GPU addresses the entire system memory pool. A Mac with 64GB of unified memory can therefore run models that would otherwise require several discrete graphics cards. The criticism was correct on its own terms and the tradeoff turned out to favour Apple for reasons nobody was arguing about at the time.
By April 2026 the effects were unmistakable. Apple's own store showed the 512GB Mac Studio removed entirely, the $599 Mac mini unavailable for the first time, and lead times of ten to twelve weeks on most higher-memory configurations. The diagnostic detail is what makes it conclusive: iMacs with 32GB shipped in one to two weeks and MacBook Pros with 128GB in two to three, while the Mac mini and Mac Studio — the two machines people buy as inference boxes — did not. This was never a general Apple supply problem.
Tim Cook, on Apple's earnings call at the end of April 2026: the Mac mini and Mac Studio are “amazing platforms for AI and agentic tools, and the customer recognition of that is happening faster than what we had predicted… the Mac mini and the Mac Studio may take several months to reach supply-demand balance.” Apple's silicon product manager later described them as “the machines of choice for running AI agents.”
Apple's response was not only to raise prices but to withdraw capacity. The 512GB Studio went first, then the $599 mini, then the 128GB Studio — capping Apple's maximum unified memory at 96GB. For a local-inference buyer that is worse than a price rise: the configurations that made the machine interesting are no longer available at any price.
The shape of the entry-level increase is the most revealing part, and it is usually reported as a simple price rise. It was not. The $599 mini was 16GB of memory with 256GB of storage; the $799 replacement is 16GB with 512GB. The buyer pays $200 more and receives twice the storage and not one additional gigabyte of memory.
The substitution is the point. The binding constraint in this cycle is DRAM, driven by the HBM conversion described in section 02 — the 2.3× wafer-per-bit penalty that pulled capacity out of conventional memory. Apple met a price increase by handing the buyer more of the component that was not the constraint, and none of the component that was.
We are not claiming the swap was costless to Apple. NAND flash prices also rose across this period, and we have not sourced a NAND series comparable to the DRAM contract and spot data used elsewhere in this report, so the relative movement of the two is left open. The claim here is narrower and does not depend on it: the buyer paid $200 more and received nothing additional of the thing that was scarce.
24GB was the upgrade this machine's buyers actually wanted, because memory capacity is what determines which models fit. It is also the upgrade that was hardest to give away in a DRAM shortage. The entry price rose while the specification moved along the axis the buyer cared about least — on a product line whose demand spiked for reasons that make it harder, not easier, to build more of.
It also removes an option rather than adding one. The 16GB/256GB configuration cannot be bought at any price now, so the cheapest route to the base chip requires buying storage the local-inference buyer has no particular use for. The longstanding criticism of Apple's bundled configurations was that they are a margin decision rather than a technical constraint. Under genuine cost pressure the bundle moved in the direction that protects the seller, which is the clearest available evidence for that reading.
What the premium on Nvidia buys is prompt-processing throughput, which matters enormously for long-context and agentic loops and barely at all for batch work. The honest trade: a Mac runs models a $2,000 graphics card cannot fit, and runs them slowly.
The arbitrage this opens is worth naming, because it is available now and will close. A 128GB AMD Strix Halo machine at roughly $2,000 puts 96GB of GPU-addressable memory on a desk and appears to be in stock — the same ceiling Apple now charges considerably more for, and cannot deliver for ten weeks.
Memory prices are a supply story. Electricity is a political one, and it is the part of this build-out that arrives on ordinary households' bills before it shows up anywhere else.
PJM runs the capacity auction for thirteen states and Washington DC — an annual market that pays generators to guarantee they can supply power on the worst day of the year. That price is the cleanest available signal of whether supply is keeping up with demand, and it has done this:
That is an increase of more than ten times in two years. And PJM does not leave the cause open to interpretation. In its own release on the 2027/28 auction it stated that forecast peak load rose about 5,250 MW year over year, and that nearly 5,100 MW of that increase is attributable to data centre demand. Its executive vice president put it plainly: the auction “leaves no doubt that data centers' demand for electricity continues to far outstrip new supply.”
PJM's own independent market monitor went further, finding that across two auctions data centres accounted for $16.6 billion of $30.8 billion in total capacity revenue — roughly half — and concluding that current conditions are “almost entirely the result of large load additions from data centers.”
PJM told regulators the 2026/27 result implied retail bill increases of 1.5% to 5% depending on the state. A peer-reviewed study published in Environmental Research Letters in May 2026, modelling 26 grid regions hourly, put the range further out at 6% to 29% nationally by 2030, and up to 57% in the worst-affected regions — Virginia, Pennsylvania, Maryland, Ohio, west Texas and New Jersey among them.
In November 2025 two Democrats won seats on the Georgia Public Service Commission, campaigning explicitly on power bills and data centres. Those were the first Democratic statewide non-federal wins in Georgia in nearly twenty years. Whatever one makes of the politics, it is the clearest evidence available that this has stopped being an industry story.
The legislative response has been fast. More than 300 bills across 30-plus states in the first six weeks of 2026, against roughly 200 bills in all of 2025. At least eighteen states are proposing special rate classes for very large electricity users. New York has passed a one-year moratorium on data centres above 20 MW, and similar measures are moving in Delaware, Georgia, Michigan, Pennsylvania, Vermont and Virginia.
Regulators have converged on a common instrument: a large-load tariff that makes the data centre commit. 77 such tariffs have been filed across 60 utilities in 36 states, 51 of them approved — 29 approved in 2025 alone, against 14 in total across the six years before. The typical shape is a minimum billing demand around 85% of contracted capacity, a contract of twelve to fourteen years, and financial security for network upgrades.
This is where the story stops being simple, and it is the finding most likely to be missing from whatever else you have read.
When AEP Ohio's tariff forced applicants to actually commit, the queue collapsed. Roughly 30,000 MW of pre-tariff requests became 13,023 MW willing to pay for engineering studies, and finally 5,642 MW signing binding contracts. More than 80% of the demand evaporated the moment it carried a cost.
ERCOT's numbers are starker still. Its large-load pipeline for 2030 stands at 410,618 MW, of which about 88% is data centres. The amount actually energised and running: 5,778 MW. Nearly 294,000 MW of that pipeline has not had a single study submitted. Industry practitioners estimate speculative interconnection requests run five to ten times actual projects, because the same site gets shopped to multiple utilities and a queue position costs almost nothing to hold.
PJM has begun revising downward. In January 2026 it cut its summer 2027 peak forecast by about 4 GW, citing projects without firm service or construction commitments, with executives saying demand forecasts had “bloated to levels that are irrational.” The auction results corroborate it: peak forecast rose 5,250 MW year over year in the December 2025 auction, then only about 2,000 MW in July 2026, and the clearing price fell slightly for the first time.
A fossil-aligned think tank makes the strongest counter-argument, and its statistics are checkable: the top ten data-centre states average 14.46 cents per kilowatt-hour against 14.39 cents elsewhere — statistically indistinguishable. States whose electricity sales grew fastest saw the smallest price increases, because grids are high-fixed-cost systems and more volume spreads those costs over more units.
The reconciliation is that these measure different things. Average historical state prices have not yet moved much. Marginal capacity and transmission costs in constrained regions have moved enormously, and that is what feeds forward into bills. The load-growth-lowers-rates mechanism only holds when new load pays its full incremental cost — which is exactly what the 51 approved large-load tariffs exist to force, and exactly what regulators found was not happening.
The procurement pattern makes sense only when read as a race for speed rather than price. Fuel cells deploy in weeks. Gas turbines now carry lead times of four to seven years, up from two and a half, with GE Vernova holding an 80 GW backlog through 2029 and expecting to be sold out through 2030. Nuclear restarts take three to five years. Small modular reactors, all of which remain announcements rather than operating plants, take five to ten.
| Buyer and plant | Capacity | Term | Status |
|---|---|---|---|
| Amazon — Susquehanna (Talen) | up to 1,920 MW | 17 years, to 2042 | ~$18B; ramps to full capacity by 2032 |
| Meta — Clinton (Constellation) | 1,121 MW | 20 years | Deliveries begin June 2027 |
| Microsoft — Crane, formerly Three Mile Island | 835 MW | 20 years | Restart of a shut reactor; ~$16B |
| Google — Duane Arnold (NextEra) | 615 MW | — | Restart targeted Q1 2029; over $1.6B |
The regulatory precedent that shaped all of these is worth knowing. Amazon's original Susquehanna deal was structured behind the meter — the data centre drawing directly from the plant, bypassing the grid. Federal regulators rejected it in November 2024 on cost-shifting grounds. The restructured arrangement delivers power through the grid with Amazon paying full transmission and distribution charges, and every subsequent co-location deal has had to work around that ruling.
PJM's price has now cleared at a politically negotiated cap for three consecutive auctions, while the system runs a roughly 6,800 MW reliability shortfall and the most recent auction attracted just 525 MW of new generation — less than the year before, at a maximum price signal. The cap suppresses the very signal that would bring supply, while the shortfall widens. Thirteen state governors agreed to it.
The spending driving all of this is real and enormous. Goldman Sachs puts total AI build-out capital expenditure at roughly $765 billion in 2026, rising toward $1.6 trillion annually by 2031. The four largest hyperscalers spent over $416B in 2025, up 66% year on year.
| Company | 2025 actual | 2026 guidance | Note |
|---|---|---|---|
| Amazon | $134.7B | ~$200B | AWS alone $96.5B in 2025, up 81% |
| Microsoft | $118B | ~$190B | Includes ~$25B from higher component pricing |
| Alphabet | $91.5B | $180–190B | Raised mid-year; 2027 to “significantly increase” |
| Meta | $72.2B | $125–145B | Raised, citing higher component pricing |
| Oracle | $55.7B (FY26) | ~$70B | Up 2.6× year on year |
Roughly $25 billion of Microsoft's 2026 capital expenditure is memory inflation, not additional capacity. Meta cited the same factor when raising its own range. Headline capex is therefore growing faster than compute is — which quietly undermines every model that reads capex as a proxy for capability, on both the bull and bear side.
Three arguments deserve attention, and they are not the ones circulating on social media.
Depreciation. Michael Burry has argued that hyperscalers are inflating earnings by depreciating AI hardware over four to six years when the effective cycle is two to three, estimating roughly $176 billion of understated depreciation across 2026–2028. Goldman Sachs, which is not bearish on demand, concedes the structural point: annual release cadences make four-to-six-year schedules “less reflective of the value of the underlying assets.”
Financing structure. The Bank for International Settlements named an AI capex bust, circular financing and sovereign debt as its three top financial-stability risks in June 2026, warning that deal terms are “typically poorly disclosed, with risks of the same asset being pledged multiple times.” Private credit to AI companies went from about $3B in 2010 to over $40B in 2025. Meta moved its Louisiana campus off balance sheet through a special-purpose vehicle carrying roughly $27B of debt.
The customer revised down. In February 2026 OpenAI cut its stated compute commitment from about $1.4 trillion to roughly $600 billion through 2030, and shifted toward renting rather than building. A 600 MW Stargate expansion at Abilene announced in September 2025 was cancelled by March 2026.
The widely repeated claim that “half of 2026 datacentre capacity has been cancelled” does not survive examination. SemiAnalysis traced it to a comparison using a denominator that excluded most of what is actually under construction — satellite imagery shows the top two hyperscalers alone exceed the figure used for the entire industry. Their own forecast revisions over six months were about 1% for North American hyperscaler self-build.
Cancellations cluster in the announcement layer: speculative projects that never ordered equipment. Killing those removes no orders and reallocates no capacity. Meanwhile Microsoft reports an $80B Azure backlog it cannot fulfil because of power — which argues the near-term problem is undersupply, not overbuild.
The hyperscalers bought the capacity. The opportunity for everyone else is in making what already exists go further, and in moving what already exists to where it is wanted.
| Position | Why it is closed |
|---|---|
| Speculating on memory inventory | Q3 2026 DRAM growth already decelerated to +13–18% as buyers hit affordability limits. If enough small buyers exit, tightness becomes glut — the classic DRAM cycle |
| “Replace your LLM with open source” | Menlo Ventures' enterprise survey found open-source share fell to 13% in 2025, down from 19%. The migration narrative is not supported by the best primary data |
| Anything assuming GPU rental gets cheaper | H100 rates rose roughly 40% between October 2025 and March 2026 |
| Competing on hardware price | You are bidding against buyers with allocation leverage and prepayment capacity |
Do not sell “replace your frontier model.” Sell workload triage: identify the 60–80% of a client's call volume that is commodity work — classification, extraction, routing, summarisation — move that to cheap, quantised or local inference, and leave the genuinely hard 20% on frontier APIs. That story survives contact with the enterprise survey data. The stronger claim does not.
And build on services rather than inventory. Services profit from volatility in either direction; inventory only profits if prices keep rising. The people with the most information — the ones building compute futures markets — are constructing instruments to hedge this, not to bet on it.
Roughly now through late 2027, with real probability of extension into 2028. New fabs arrive 2027 and 2028; Intel's chief executive says no relief until 2028; IDC expects no meaningful easing before the end of 2027 and says PC pricing is unlikely to return to 2025 levels even after capacity expands. TrendForce and SK hynix openly disagree about 2027 — one sees NAND supply outpacing demand, the other calls it the worst year yet. That disagreement is itself the signal: nobody knows, and anyone certain is selling something.
Research was conducted across public sources in July 2026: memory market analysts (TrendForce, DRAMeXchange, Counterpoint, IDC, Gartner), company earnings calls and results releases, grid operator filings and auction results (PJM, ERCOT), regulator proceedings, peer-reviewed work, and trade press. Figures are attributed to the publisher that reported them.
Several widely circulated figures in this subject area could not be traced to a primary source and were excluded. Specifically: claims that typical street pricing for high-end consumer graphics cards has doubled are not supported — manufacturer pricing held at $1,999 and measured average increases across all vendors were about 15% between November 2025 and February 2026, with $3,000–4,000 figures reflecting premium variants and resale rather than market price. Claims that Nvidia has announced an exit from gaming could not be verified; what is documented is constrained supply, the withdrawal of a programme that held prices near list, and a product refresh apparently delayed because a 3GB memory module costs three times a 2GB one. HBM vendor market shares for 2026 are unresolved across trackers. The 2.3× wafer-per-bit ratio is our arithmetic on TrendForce's published wafer and bit series; TrendForce publishes both series but does not state the ratio.
Full citation set, including every figure referenced above, available on request.