The 2026 DRAM shortage: why RAM got expensive and when it ends

RAM got expensive because the three companies that make almost nine-tenths of the world’s DRAM gave their wafers to AI data centres first, as HBM for the accelerators and as server DDR5 for the machines around them, and on the dated supply outlooks available in late September 2026, new fabs ramp meaningfully only from the second half of 2027 and the gap does not close before 2028.

TrendForce told the Future of Memory and Storage conference in August 2026 that it expects DRAM supply to exceed demand by about 2% in 2028, as Forbes reported the session, and Micron’s results presentation of 24 June 2026 says industry supply should “improve gradually in 2028” while admitting the company has no line of sight to when supply will catch up with demand. None of the forecasts collected here dates a return to 2025 prices. IDC wrote in February 2026 that its model “does not point to a reversion to 2025 pricing levels within the forecast horizon”, and Micron has written price floors into five-year customer contracts.

“AI ate the RAM” is the usual explanation, and it is incomplete in four ways that matter to anyone trying to judge where prices go next. The DDR4 squeeze began in mid-2025, with end-of-life notices and US tariff pull-ins, before the surge in AI-server buying reached ordinary memory. From the first quarter of 2026 an ordinary 64GB DDR5 server module earned more per wafer than HBM did, on TrendForce’s estimate, so the suppliers had no reason to hand wafers back to PCs even where HBM demand paused. The estimated deficit is small, 1% to 2% of bits in 2026 on TrendForce’s figures, and it still produced roughly a fivefold rise in contract prices, because the buyers at the front of the queue barely respond to price. And “the price” is really five prices, contract, chip spot, module spot, retail and used, which moved at different times and by very different amounts.

This article follows the money through four layers. The fab: who controls the wafers, the downturn that left no slack, the HBM trade ratio, the server DDR5 that out-earned it, allocation, the squeeze from DDR4 down to DDR2, and a dated timeline. The contract: the quarterly series, its missed forecasts, why a small deficit quintupled a price, and the five prices that do not move together. The shelf: kits, laptops, phones, consoles and servers. And the second-hand market, which nobody publishes. Then long-term agreements, profits and the price-fixing question, the supply response fab by fab, every dated forecast of the end beside the forecaster’s interest, what would shorten or lengthen the shortage, and twenty short answers. Coping advice is kept short, because sibling guides own timing, capacity and DDR4. It is long because each layer has its own evidence, and the predictions that went wrong in 2025 and 2026 were nearly all made from one.

One note on sources, because it changes how the numbers should be read. Where a figure is a supplier’s, an analyst’s or a standard’s, the document is named in the text with its date; where a figure is worked out here, the arithmetic is shown so that it can be checked. TrendForce publishes the only public quarterly series of DRAM contract prices, so the article leans on it, and for that reason also shows how often its forecasts missed. Suppliers, and the companies that buy from them, are interested parties and are labelled so. Prices on this site move every few hours, so the text names the page to read rather than freezing a site price into prose. The knowledge here is dated late September 2026; Micron reports its fiscal fourth quarter on 30 September, after this was written, and nothing below anticipates it.

September 2026 brought a plateau, not a turn. TrendForce’s spot report of 16 September said some DDR5 chip prices had begun to soften and had the DDR4 8Gb chip at $45.54, against $42.11 on 4 August, yet a week later TrendForce lifted its fourth-quarter contract outlook, led by PC and server DRAM.

Three companies make nearly nine-tenths of the world’s DRAM, all of it from the same wafers

DDR4 and DDR5 desktop modules, the LPDDR5X soldered into phones and thin laptops, the GDDR on a graphics card and the HBM stacked beside an AI accelerator are all DRAM. They are different dies with different interfaces, and the differences matter to a buyer, which is what DDR4 vs DDR5 is about. For supply they are one thing. They come off the same kind of 300 mm silicon wafer, in the same few companies’ fabs, through largely the same equipment, and a wafer started as HBM is a wafer not started as anything else. That single fact is why a product sold to data centres can double the price of a desktop kit. IDC put it more bluntly in its December 2025 analysis of the shortage: “This is a zero-sum game: every wafer allocated to an HBM stack for an Nvidia GPU is a wafer denied to the LPDDR5X module of a mid-range smartphone …”

So the unit of account for this whole article is the wafer, and the first question is who controls it. TrendForce’s revenue ranking for the second quarter of 2026, published on 7 September, answers it.

The top three, worked from those figures:

60.98 + 38.59 + 36.00 = 135.57
135.57 / 154.73       = 0.876   -> 87.6% of DRAM revenue, 2Q26
87.6% + CXMT's 9.5%   = 97.1%   -> the top four

A figure of “about 95%” for the big three is widely repeated, and the US class-action complaint of June 2026 puts them at more than 90% of global DRAM sales. TrendForce’s revenue data for the latest quarter says 87.6%, and revenue is the relevant measure because it is what the suppliers are paid for.

The small makers matter out of proportion to their revenue. Nanya, Winbond and PSMC in Taiwan are the companies still making DDR4, DDR3 and DDR2 in volume, and their growth rates in that last column are the legacy-memory story told in one line. CXMT, much the largest of the rest, was reported in late 2025 to have ended most of its DDR4 production, and is raising the share of DDR5 and LPDDR5 in its output (TrendForce’s news desk, December 2025 and September 2026).

A market in which three sellers hold nearly nine-tenths of supply answers a jump in demand with price rather than volume whenever capacity is fixed, and in 2025 and 2026 capacity was fixed, for reasons that began three years earlier. Every module in the DDR5 and DDR4 listings on this site was built from wafers allocated by one of these companies.

The shortage was built in the downturn of 2022 and 2023

Every account of the shortage starts with AI. The precondition started with a glut.

In late September 2022, with demand falling fast, Micron became the first major memory maker to announce an official cut to its DRAM and NAND output (TrendForce, 3 October 2022), guiding 2023 DRAM bit growth of only about 5%. The rest followed. DRAM industry revenue fell 21.2% in the first quarter of 2023, to $9.663 billion, the third straight quarterly decline (TrendForce, 25 May 2023). Samsung deepened its cuts at the end of the third quarter of 2023, mainly on DDR4, and TrendForce estimated that its cut would reach 30% in the fourth quarter. The analyst Patrick Moorhead told PCWorld in April 2026 that the makers had negative gross margins in 2022 and 2023 and stopped investing; Micron’s filings put its losses in 2023, with a gross margin still about 22% in the quarter to 1 December 2022 and minus 32.6% in the quarter to 2 March 2023.

Two habits carried into 2024: caution about new capacity after the losses, and a steering of advanced capacity towards HBM. TrendForce noted both on 20 May 2024, with a warning that reads differently now: “The increased focus on HBM production could lead to a DRAM supply shortage.”

Prices were still falling at the turn of 2025. On 30 December 2024 TrendForce forecast conventional DRAM contract prices down 8% to 13% for the first quarter of 2025, blaming seasonal weakness, PC makers’ inventory cuts and passive buying, with Chinese DDR4 output and cheap spot chips pushing DDR4 down harder than DDR5, and DRAM revenue duly fell 5.5% in that quarter, to $27.01 billion (TrendForce, 3 June 2025). It was the last falling quarter. Then the stock ran out, on TrendForce data reported by Reuters and relayed by TrendForce’s news desk on 4 December 2025.

The report does not say whose inventory the weeks measure, supplier or buyer, so read the column as a direction.

The rally started from a price trough only about a year before the spike, with the industry’s spare stock gone and no cleanroom space standing empty. Why the cleanroom is the constraint money cannot quickly relieve is covered in the section on new fabs. What the downturn left behind shows in the revenue series, worked here from TrendForce’s figures rather than claimed by TrendForce:

2Q26 revenue / 1Q23 trough           = 154.73 / 9.663 = 16.0
  -> quarterly DRAM revenue rose about sixteenfold in 13 quarters
2Q26 revenue / 1Q25, the last fall   = 154.73 / 27.01 = 5.7
  -> about 5.7 times in the five quarters since revenue last fell

Almost none of that is volume, as the contract series below shows.

A gigabyte of HBM costs about three gigabytes’ worth of wafer

High-bandwidth memory is DRAM dies stacked vertically, joined by through-silicon vias, and packaged beside an accelerator rather than on a module. The stacking idea is the one used in the 3DS server modules described in ranks and 3DS modules; the scale and the interface are not. Every HBM stack costs more wafer than the bits it holds, for three reasons. The die is bigger: TrendForce wrote in March 2024 that an HBM die is 35% to 45% larger than a DDR5 die of the same process and capacity. The yield is lower: 20% to 30% below DDR5’s, including the TSV packaging. And the process is 1.5 to 2 months longer. On 20 May 2024 TrendForce gave different figures, yields of about 50% to 60% and a wafer area 60% larger; the two releases do not reconcile, and no primary figure from Samsung or SK hynix settles it.

Micron gives a single number instead. Its prepared remarks for the December 2025 results call it “the 3-to-1 trade ratio with DDR5, and this trade ratio only increases with future generations of HBM”. At Hot Chips on 23 August 2026, Micron’s HBM design architecture fellow Raghu Sreeramaneni repeated that HBM3E needs roughly three times DDR5’s wafer supply for the same number of bits, according to MLQ News’ report of the talk; Micron’s June 2024 investor presentation, which first gave the figure, measured it at the same technology node. HBM4 makes it worse. It doubles the interface from 1,024 to 2,048 I/O, which enlarges the die, and some makers are moving to a logic-process base die (TrendForce, 22 May 2025).

That is a supplier’s figure, and suppliers have reasons to describe supply as tight. So here is an independent check, built from TrendForce’s estimates of 2 June 2026: the share of the three largest makers’ DRAM wafer input that goes to HBM, and HBM’s share of DRAM bit supply. The release does not say whether the two shares cover the same makers; the table assumes they do.

• Share of the big three's DRAM wafer input
• Share of DRAM bit supply
• Forecast

Source: TrendForce, 2 Jun 2026. Figures as of June 2, 2026. End-2026 and end-2027 figures are forecasts. The wafer share covers the three largest makers; TrendForce does not say whether the bit share does too.

Show the figures

HBM takes more than twice the share of DRAM wafers that it supplies in bits (% of DRAM wafer input and of DRAM bit supply going to HBM, at year end, TrendForce estimates)

Period	Share of the big three's DRAM wafer input	Share of DRAM bit supply
End 2025	18%	8%
End 2026	22% (forecast)	9% (forecast)
End 2027	30% (forecast)	13% (forecast)

Worked here rather than by TrendForce, the analyst’s own shares imply that a wafer of conventional DRAM yields about 2.5 to 2.9 times the bits of a wafer of HBM, which agrees with Micron’s same-node, same-bits three to one. If the bit share covers the whole industry, CXMT and the Taiwanese makers included, the ratio is lower, about 2.2 in 2025 and 2.5 in 2026 and 2027 (assuming the big three make about 88% of the industry’s bits), so read the range as roughly two to three. On either reading it rises from 2025 to 2026.

Higher figures circulate. Commercial Times, relayed by TrendForce’s news desk in December 2025, quoted experts saying a gigabyte of HBM “consumes 4x the capacity of standard DRAM”; Moorhead told PCWorld that HBM needs “approximately quadruple the memory dies”; IBM’s Jung Yoon told the August 2026 conference, as Forbes reported it, that HBM needs three to four times more wafers, because of larger dies and TSV stacking yields, which counts stacking loss. The other two give no basis, and the TrendForce-derived ratio is the only one whose inputs are published.

The amount of HBM per accelerator keeps rising as well. TrendForce’s June 2026 release has HBM per AI chip going from 96GB and 192GB to 216GB and 288GB during 2026, and Nvidia’s Rubin Ultra expected at 384GB per GPU in 2027, though by August TrendForce reported Nvidia weighing lower configurations because of the 2027 shortage. IEEE Spectrum gave the physical scale on 10 February 2026: Nvidia’s B300 uses eight HBM stacks of 12 DRAM dies each, and SemiAnalysis estimates that HBM is 50% or more of the cost of the packaged GPU. An illustration of what one such accelerator represents, in wafer-equivalents and nothing else:

one Rubin Ultra GPU:  384 GB of HBM x about 2.85  = about 1,090 GB
                      of conventional-DRAM wafer equivalent
1,090 GB / 32 GB per desktop kit                  = about 34 kits' worth of bits

A gigabyte of HBM does not remove a fixed number of desktop gigabytes; the trade is in wafers. What an HBM order removes is wafer starts from a shared pool, and how many desktop gigabytes those starts would have become depends on the node, the die, the yield and what the supplier would otherwise have built. In 2026 the answer to that last question was mostly server DDR5, not desktop kits, which is the next section.

HBM is only half the story: server DDR5 out-earned it per wafer in 2026

Most accounts stop at HBM. TrendForce’s own releases show the other half.

HBM is priced once a year, conventional DRAM every quarter. Nvidia opened its 2026 HBM3E negotiations in May 2025, while buyers still had leverage, and locked prices in before the market turned (TrendForce, 18 December 2025). In October 2025 TrendForce noted that HBM3E had carried more than four times DDR5’s price in the second quarter of 2025, and forecast that “DDR5 profitability is projected to surpass that of HBM3e starting in the first quarter of next year.” In December it expected the price gap between HBM3E and server DDR5 to narrow from four-to-five times to one-to-two times by the end of 2026, and by June 2026 it was reporting the outcome: “HBM wafer revenue was overtaken by DDR5 64GB RDIMM in 1Q26”, with HBM’s profitability below the RDIMM’s from the same quarter. SK hynix, whose bit mix leans furthest towards HBM, grew revenue least of the three in the second quarter for exactly that reason. Micron’s mix moved the same way: HBM and other cloud-related memory went from 17% of its DRAM revenue in 2023 to nearly 50% in 2025 (IEEE Spectrum, February 2026).

Once the commodity server module paid better per wafer than HBM, no supplier had a reason to release wafers from either to make desktop memory. The idea that HBM is the single most profitable product was stale from the first quarter of 2026 on TrendForce’s per-wafer estimate, although IBM’s Jung Yoon told the August conference, as Forbes reported it, that suppliers earn 5 to 10 times higher margins on HBM than on standard DDR5, without saying on what basis. The 64GB registered module that overtook HBM per wafer is the same part listed under DDR5 registered and DDR5 64GB modules on this site. Server RAM vs desktop RAM explains why it will not run in a desktop board.

Demand for ordinary RDIMMs widened at the same time, for three reasons. Cloud providers moved from AI-specific servers to general-purpose ones and extended their buying “to RDIMMs across multiple densities” (TrendForce, February and June 2026). Agentic-AI inference raised the number of CPUs per GPU, from 1:8 towards 1:4 or higher, with Nvidia’s NVL72 rack at 1:2 (TrendForce, 29 May 2026), and every CPU carries its own DDR5. And modules got bigger, at first: SK hynix said in October 2025 that its shipments of DDR5 modules of 128GB or more had more than doubled in a quarter, though in the first half of 2026 some systems stepped back down to 32GB and 64GB modules, as the section on servers shows. TrendForce estimated in August 2026 that HBM and RDIMMs together take 51% of all DRAM bit supply in 2026. A much-repeated line that “AI will consume 20% of DRAM” in 2026 descends from a wafer estimate for HBM and GDDR7 alone, and leaves out the server modules.

Phone-class memory joined the same queue. Micron began sampling a 192GB SOCAMM2 server module in October 2025, and Samsung reported the first mass-product sales of SOCAMM2 for Nvidia’s Vera Rubin in the first quarter of 2026. TrendForce reported in June that, on the suppliers’ preliminary 2027 plans, Nvidia would get only about 60% of the LPDRAM it needs, the same class of memory as a phone’s or a thin laptop’s.

One famous figure needs taking down a peg. On 1 October 2025 Samsung announced a letter of intent with OpenAI, whose memory demand was “projected to reach up to 900,000 DRAM wafers per month”. Reuters calculated that as potentially 40% of global output, and some reference pages turned it into “OpenAI consumes 40% of the world’s DRAM”. It is a non-binding letter with “up to” in front of it; SEMI projects global capacity of about 2 million wafers a month by 2030, so 900,000 would be 45% of the world’s capacity four years from now.

Allocation, not price, decided who got memory first

The queue formed in the autumn of 2025. TrendForce reported on 24 September 2025 that some US cloud providers planned to start 2026 procurement “as early as 4Q25, adopting a more flexible stance on pricing”, and by January 2026 they were pulling orders forward and taking a larger share of the growth in bit supply. SK hynix said on 29 October 2025 that customers had already taken its entire 2026 DRAM and NAND production. Micron’s December 2025 remarks were blunter: “Despite significant efforts, we are disappointed to be unable to meet demand from our customers, across all market segments.” How much it could meet is reported only second-hand: about 55% to 60% of core customers’ demand (TrendForce’s news desk, December 2025), or half to two-thirds of key customers’ needs (chief executive Sanjay Mehrotra to Bloomberg in May 2026, as relayed by TechTimes).

Further down the queue the fill rates fall, as the table below shows, and TrendForce said in January 2026 that suppliers had tightened supply to PC OEMs and module makers at once, “forcing some OEMs to procure memory at higher prices through modules makers”.

The pricing cadence broke as well. Samsung largely halted DRAM contract offers in October 2025, with pricing moving to monthly quotes (TrendForce’s news desk), and Hankyung reported in January 2026 that Samsung and SK hynix had asked Microsoft and Google for first-quarter server prices 60% to 70% above the fourth quarter’s.

The second half of row five is an executive’s view, not a supplier disclosure: Apacer’s chief executive C.K. Chang said at the company’s investor conference on 24 July 2026, as DigiTimes and Tom’s Hardware reported, that big-maker supply to independent module makers could fall in 2027 to about 30% of the 2026 volume. It matters because every retail DIMM and SODIMM comes from a module maker, and Kingston alone held 62% of module-maker revenue in 2025 (TrendForce, 22 September 2026). The retail kit is the last link in a queue whose first buyers barely look at price.

The best-known casualty was a brand. On 3 December 2025 Micron announced that it would leave the Crucial consumer business, shipping through the consumer channel until February 2026, “to improve supply and support for our larger, strategic customers in faster-growing segments”, in the words of its chief business officer Sumit Sadana. It did not leave consumer memory altogether: warranty service continues, and it still supplies PC makers. What the exit means for a Crucial warranty claim is covered in DDR4, used memory and upgradeability.

DDR4 was squeezed first, by end-of-life notices, and the squeeze ran down to DDR2

The oldest mainstream generation went up first, and for a while it was dearer than the new one. This is the supply history; whether DDR4 is still worth buying belongs to DDR4 as an escape route.

In mid-2025 the three big makers told customers DDR4 was ending. TrendForce’s release of 5 June 2025 gives the reason in one clause: “Given the significantly higher profit margins for HBM, DDR5, and LPDDR5(X), suppliers have laid out EOL plans for DDR4, with final shipments expected by early 2026.” Korean and US suppliers had also announced cuts to LPDDR4X from 2026. Tariffs added a rush: OEMs front-loaded shipments before the 90-day grace period on April’s US reciprocal tariffs expired on 9 July, and TrendForce raised its second-quarter forecast for server DDR4 contracts from +5-10% to +18-23%.

No press release with the exit dates was found; the dates below come from press reports relayed by TrendForce’s news desk, and they moved.

Prices moved before the dates arrived. DDR4 and DDR5 chips reached spot parity in early June 2025, and on 23 June the DDR4 16Gb chip averaged $12.00 against $6.01 for DDR5 16Gb, “a first in DRAM history”, as TrendForce’s news desk put it, with SK hynix and Micron already in last-time-buy talks with DDR4 customers. In July 8GB PC DDR4 modules cost more than DDR5 ones at contract, and TrendForce’s forecast for third-quarter consumer DDR4 contracts doubled in five weeks, from +40-45% to +85-90%. SK hynix called it “a short-term demand surge driven by supply concerns” on 24 July 2025.

Fifteen months later the chip inversion still held, on TrendForce’s spot table of 24 September 2026:

DDR4 16Gb 3200 chip  $83.784       DDR5 16Gb 4800/5600 chip  $57.667
83.784 / 57.667 = 1.45   -> DDR4 chips about 45% dearer per bit than DDR5 chips
DDR4 8Gb chip:  $2.963 (early June 2025) -> $46.107 (24 Sep 2026) = 15.6 times

Sources: TrendForce weekly spot reports: 11 Jun 2025; 2 Jul 2025; 6 Aug 2025; 3 Sep 2025; 1 Oct 2025; 5 Nov 2025; 3 Dec 2025; 14 Jan 2026; 4 Feb 2026; 4 Mar 2026; 1 Apr 2026; 6 May 2026; 3 Jun 2026; 1 Jul 2026; 5 Aug 2026; 2 Sep 2026; 30 Sep 2026. Figures as of September 29, 2026. One reading per month: the first in TrendForce's weekly reports, taken between the last day of the month before and the 7th. June 2025 and January 2026 are the previous week's readings given in the 11 June and 14 January reports. The latest reading, on 29 September 2026, was $46.32.

Show the figures

A DDR4 8Gb chip cost about 15 times as much at spot in September 2026 as in June 2025 (US dollars per chip: TrendForce's average spot price for DDR4 8Gb (1Gx8) 3200MT/s)

Month	DDR4 8Gb chip, average spot price
Jun 2025	$2.963
Jul 2025	$5.073
Aug 2025	$5
Sep 2025	$4.878
Oct 2025	$6.359
Nov 2025	$10.629
Dec 2025	$16.514
Jan 2026	$25.407
Feb 2026	$30.88
Mar 2026	$33.02
Apr 2026	$33.96
May 2026	$32.4
Jun 2026	$34.8
Jul 2026	$36
Aug 2026	$42.11
Sep 2026	$44.54

The cushion went too. CXMT, whose extra DDR4 TrendForce had expected to push DDR4 prices down harder in early 2025, was reported to have left most of it. Micron’s new Virginia DDR4 serves automotive, industrial, medical, aerospace and defence customers, and TrendForce said in May 2026 that it reflects internal reallocation, with DDR4 at about 7% of Micron’s output and “no plans to increase Micron’s total production capacity for LPDDR4 and DDR4”. Nanya and Winbond took the business, which TrendForce said in July cannot “fully offset the reduction in supply”.

Then the cascade. In March 2026 DDR4 4Gb prices rose more than 20% in a month, and DDR3 and DDR2 prices rose 20% to 40% (TrendForce, 7 April 2026). By June some OEMs were redesigning DDR4 products onto DDR3, and DDR3 products onto DDR2. Winbond was cutting DDR2 in favour of higher-margin parts while ESMT made as much as its PSMC wafer allocation allowed, and TrendForce forecast DDR2 contracts up 55% to 60% in the second quarter and 35% to 40% in the third. The DDR3 4Gb spot chip averaged $13.756 on 24 September 2026.

DDR2 contract, 2Q26 x 3Q26 forecasts at midpoints:  1.575 x 1.375 = 2.17
  -> more than doubled in two quarters, if the forecasts held
DDR3 4Gb chip = 0.5 GB:   13.756 / 0.5 = $27.5 per GB, close to DDR5's $28.8

DDR3 and DDR2 rose last and sometimes hardest, because nobody is building capacity for them. That matters on this site, where the DDR3, DDR3 server, DDR2 and original DDR listings are priced by the same per-gigabyte measure as DDR4 and DDR4-3200, and the history of the per-gigabyte crossover is in DDR4 vs DDR5.

From glut to shortage in four years, in one table

One row per event, each from a named document. Read it as three phases: the downturn and the cuts (2022 to 2024), the legacy squeeze and the turn to AI servers (2025), and the record contract quarters, device price rises and first supply announcements (2026).

Two of those rows are often misreported, and so is one figure the table leaves out. The OpenAI row is the non-binding letter of intent taken apart in the section on server DDR5, a projection and not a measure of consumption. The Micron groundbreaking of July 2026 was an expansion of its existing Hiroshima plant, not a new fab, with commercial shipments expected around the summer of 2028. And TrendForce’s November 2025 revision, sometimes given as “the PC market cut to -2.6%”, was a cut to 2026 notebook production, from +1.7% to -2.4%.

The contract series compounds to about five times, and the forecasts kept missing low

Contract prices are what PC makers, phone makers and cloud providers pay the suppliers. TrendForce’s “conventional DRAM” means all DRAM except HBM, and its forecasts and later measurements are the only public quarterly record. The table sets each quarter’s first forecast beside its revisions and the outcome.

The fourth quarter of 2025 carries two TrendForce figures: +45-50% for conventional DRAM in the February 2026 revenue release (+50-55% with HBM), and “53-58%” in a January release whose basis is not explained. The same release put the historical peak for a single quarter at about 35%, which makes the first quarter of 2026, at 93-98%, nearly three times the old record.

The third-quarter slowdown came from the demand side and from contracts, not from new supply: “customers from consumer markets, such as PCs and smartphones, are reaching their affordability limit”, in TrendForce’s words, and it also cited the high comparison base and the share of buying now fixed by long-term agreements. A July report that TrendForce had “shown” a 15-18% rise misread a 13-18% forecast, and the only public fourth-quarter percentage comes from a crypto-news relay no TrendForce page confirms.

Two larger figures circulate with no published basis: a European component distributor’s advice in June 2026 to plan for rises of “10 to 20 percent per month through the end of 2026”, and a consumer explainer’s April warning of “another 100%”. Compounded over a quarter, 10% a month is 1.10 x 1.10 x 1.10 = 1.33, a rise of 33%, and 20% a month is 1.73, a rise of 73%, against the only public quarterly forecast, TrendForce’s 13% to 18% for the third quarter.

Multiplying the quarters out is the step nobody publishes:

4Q25 and 1Q26 measured, 2Q26 and 3Q26 forecast, as multipliers
midpoints    1.475 x 1.955 x 1.605 x 1.155 = 5.35
low ends     1.45  x 1.93  x 1.58  x 1.13  = 5.00
high ends    1.50  x 1.98  x 1.63  x 1.18  = 5.71

retail check: PCPartPicker average, DDR5-6000 2x16GB (via Tom's Hardware)
             August 2025 $108 -> August 2026 $572     572 / 108 = 5.30

Two of the four quarters are forecasts, so 5.35 is a forecast-weighted estimate, not a measured index. The retail check is independent and lands in the same place, and Counterpoint’s figure, reported by CNBC in June 2026, was that memory and storage prices had “quadrupled in the past three quarters”. From the seller’s side, Micron’s DRAM revenue went from $10.8 billion in its fiscal first quarter of 2026 to $31.3 billion in its third, 2.9 times in two quarters, with prices up in the low 60s per cent in the third quarter and bit shipments up by a low single digit.

• Forecast

Sources: TrendForce, 26 Feb 2026 (4Q25: +45-50%); TrendForce, 1 Jun 2026 (1Q26: about +93-98%); TrendForce, 31 Mar 2026 (2Q26 forecast: +58-63%); TrendForce, 3 Jul 2026 (3Q26 forecast: +13-18%, repeated 7 Sep). Figures as of September 30, 2026. 2Q26 and 3Q26 are TrendForce forecasts. Each quarter uses the midpoint of TrendForce's range: 1.475 x 1.955 x 1.605 x 1.155 = 5.35.

Show the figures

Compounded, TrendForce's figures put DRAM contract prices at about five times their 3Q25 level by 3Q26 (Index, 3Q25 = 100, conventional DRAM, compounded from TrendForce's quarterly changes)

Period	Conventional DRAM contract price
3Q25	100
4Q25	147.5
1Q26	288.4
2Q26	462.8 (forecast)
3Q26	534.6 (forecast)

In both quarters where TrendForce’s first forecast can be checked against its own measurement, the forecast was too low, once by a factor of more than four, so read any upward forecast from this period as a floor rather than a central estimate. What that means for timing a purchase is the subject of buy RAM now or wait.

A deficit of one or two per cent was enough to quintuple the price

Readers meet several different “shortage” numbers, and they measure different things. Set side by side:

IDC’s 16% and Micron’s low-to-mid 20s disagree about supply growth, and they were made six months apart. The gap between bit growth of about 20% and wafer growth of 4.8% is mostly node migration, which puts more bits on each wafer, net of HBM’s growing share of wafers; the section on new fabs returns to it. An illustration shows the sizes are at least coherent. It is not how TrendForce computes its ratio:

demand +25% (SK hynix, "mid-20%")    supply +23% (Micron, "low- to mid-20s")
1.23 / 1.25 = 0.984   -> a deficit of about 1.6% of bits

The puzzle is why so small a gap produced so large a price. The answer is who sits at the margin. TrendForce estimated on 3 August 2026 that the nine largest cloud providers would spend more than $886.7 billion on capital projects in 2026, about 90% more than in 2025; a release three weeks later gave 98% for a group it called major CSPs. On 25 August it put DRAM and NAND together at 47% of their capital spending in 2026 and 68% in 2027. In February 2026 TrendForce set out why this cycle differs from the last: “Rather than being led primarily by end-device manufacturers, the current wave is driven by CSPs. CSP procurement volumes are growing exponentially, and their relatively lower sensitivity to pricing has enabled record-high price increases exceeding those seen in the previous supercycle.”

A PC maker facing a doubled memory bill can cut the specification, raise the price or ship fewer machines, and in 2026 it did all three. A cloud provider racing to build AI capacity has one alternative to paying, which is not building, and through 2026 it chose to build. When the buyer at the margin will pay almost anything, the price is set by the last bit, not the average one, and everyone further down the queue pays the price that buyer set.

Contract, spot, module, retail and used are five prices that do not move together

Every price quoted in this shortage belongs to one of five layers. Contract prices are negotiated between makers and OEMs or cloud providers, quarterly or monthly, and carry most of the volume. Chip spot is daily broker trading in single chips, “only a small fraction of total memory sales” in TechSpot’s words; no primary figure for its share was found in this research. Module spot is weekly trading in finished modules, retail is new kits on shelves, and used is the layer with no published series at all.

The units trap comes first. TrendForce’s spot item “DDR4 16Gb (2Gx8)” is a 16-gigabit chip, which holds 2 gigabytes. A widely repeated April 2026 figure for “a 16GB DDR4 chip” at $74.10 was that 2GB part. Throughout this article Gb means gigabit, for a chip, and GB means gigabyte, for a module or kit.

Every layer normalised to dollars per gigabyte, using the latest figure published for each as of September 2026:

per GB = price / gigabytes; a 16Gb chip holds 2 GB:  57.667 / 2 = 28.8 $/GB
module spot, DDR5 RDIMM 32GB:  2,050 / 32 = 64.1 $/GB
Dell's quote:  2.39 $/Gbit x 8 bits per byte = 19.12 $/GB

The table contains an apparent absurdity: a branded DDR5 chip on the spot market costs more than twice as much per gigabyte as the cheapest retail kit, and a DDR4 chip seven times as much. Three things explain it. Chip spot is thin and disorderly: on 24 September the DDR4 16Gb chip traded between $45.50 and $119.00 within one session. Modules are not built from spot chips at the day’s price: TrendForce’s module ranking credits module makers’ profits to low-cost inventory built in 2024 and 2025, and the lower-grade eTT chips on TrendForce’s table trade at a fraction of the branded price. And retail sells stock bought months earlier; Lexar’s regional manager for Australia and New Zealand told Tom’s Hardware in June 2026 that costs take up to eight or nine months to reach consumer prices, one executive’s account rather than a measured lag.

A price quoted without its layer and its date is not information. Every listing page on this site adds a sixth vantage point, listed modules at the moment you look, with the cheapest and median price per gigabyte for the view; every generation by price per gigabyte, DDR5 16GB modules, DDR4 16GB modules and DDR5 registered correspond to the rows above.

Retail lagged, overshot, dipped in the spring and climbed again

Shelf prices followed the contract series with a delay, ran ahead of it, then paused. The 32GB DDR5 kit shows the shape; each row says what it measures, because a best price, an average and one kit’s price are different things.

The spring dip is the part that misled people, and the explanations given at the time were about inventory, not supply. IDC’s Tom Mainelli told PCWorld in April that “some SKUs saw pricing rise above what the market would support”, adding that “we are still far from normalization”. HotTech’s Dave Altavilla pointed to “localized inventory pressure in distribution channels, particularly in secondary markets”, and, according to DigiTimes as relayed by TechSpot, Google’s TurboQuant announcement prompted some distributors to clear stock. Advice to hold off until August followed; on 24 August the cheapest 32GB DDR5-6000 kit crossed $400 (Tom’s Hardware).

August 2026:  572 average / 392 best = 1.46  -> the average kit cost 46% more than the cheapest
DDR5 16Gb spot:  $37.20 (early April, TechSpot) -> $57.667 (24 September,
                 TrendForce) = +55% after the dip

Europe had the same shape. 3DCenter’s index of the cheapest German retail offer (July 2025 = 100) put DDR5 at 179 in November 2025, 440 in January and February, 408 in March and 419 in June; a headline calling that “419% more expensive” confused the index with the increase, since an index of 419 is a rise of 419 - 100 = 319% since July 2025. The climb resumed in the summer: 3DCenter’s September index, published on 14 September 2026, had DDR5 at 544, while its DDR3 and DDR4 index had slipped from 334 in April to 292. ComputerBase’s check of 16 September 2026 then found desktop kits 370% dearer than a year earlier, a new high.

• DDR5 desktop memory
• DDR3 and DDR4 desktop memory

Sources: 3DCenter, June 2026 index, 15 Jun 2026 (July 2025 to June 2026); 3DCenter, September 2026 index, 14 Sep 2026 (July to September 2026). Figures as of September 13, 2026. Prices are taken on the weekend nearest the middle of each month, from Geizhals, leaving out eBay and Amazon Marketplace sellers. An August change of method revised some earlier DDR5 months by one point but left June 2026 at 419, so the two tables join there.

Show the figures

German retail: DDR5 reached 5.4 times its July 2025 price by September 2026, DDR3 and DDR4 2.9 times (Index, July 2025 = 100: 3DCenter's average over a basket of desktop memory, each item at its cheapest German shop price)

Month	DDR5 desktop memory	DDR3 and DDR4 desktop memory
Jul 2025	100	100
Aug 2025	102	110
Sep 2025	103	110
Oct 2025	119	127
Nov 2025	179	168
Dec 2025	345	218
Jan 2026	440	319
Feb 2026	440	331
Mar 2026	408	326
Apr 2026	410	334
May 2026	414	317
Jun 2026	419	296
Jul 2026	445	298
Aug 2026	486	301
Sep 2026	544	292

The fivefold figure is DDR5’s, and it depends on the kit. Tom’s Hardware’s compilation of PCPartPicker’s US averages, published on 17 August 2026, has three rows that between them cover most desktop buyers:

572 / 108 = 5.30       1,118 / 191 = 5.85       281 / 105 = 2.68
dollar rise in a year:  DDR5 32GB +$464    DDR5 64GB +$927    DDR4 32GB +$176

DDR4 kits rose about half as much as DDR5 kits over the same year, although DDR4 chips cost more per bit at spot. Two facts fit that gap, though no source measures how much of it they explain: module makers stocked DDR4 in early 2025, ahead of the tariffs and the end-of-life dates (TrendForce, 22 September 2026), and lower-grade DDR4 chips traded at about half the per-gigabyte price of lower-grade DDR5 on 24 September. The discount for a bigger kit has gone too: PC Games Hardware found in August 2026 that German 64GB kits cost about twice as much as 32GB DDR5-6000 ones. Because the shortage behaves as a charge per gigabyte, the buyers who need the most, for workstations, virtual machines or local AI models, carry the largest dollar rises. Whether that capacity is worth its price is the subject of how much RAM you need, and whether DDR4 is a way round it is in DDR4 as an escape route.

A dip in one layer, in one region, for one kit, is inventory moving, not supply arriving. The gap between the average and the cheapest in August is also why this site shows both: the DDR5 kits totalling 32GB and DDR5-6000 pages state the cheapest and the median price per gigabyte for the view. What a dip should mean for a purchase is in buy RAM now or wait.

Used memory is the one price nobody publishes

No analyst, tracker or publisher found in this research quantifies second-hand DIMM prices for 2025 or 2026. What exists is qualitative, or narrow, or both.

The two small indices agree on one thing: in September 2026 DDR4 server memory sold for about a fifth of DDR5 server memory per gigabyte (5.56 / 31.05 = 0.18; 9.34 / 41.09 = 0.23), and eRacks has DDR4 falling while DDR5 rose. That runs opposite to chip spot, where DDR4 is the dearer silicon. Neither is a used-price index in the strict sense, since DatacenterDisk’s DDR5 median is new stock and eRacks does not say which it tracks; neither is eBay, and neither should be generalised to it.

Two forces act on that supply. Meta now reuses 32GB DDR4 modules recovered from decommissioned servers, attaching 256GB of them beside 768GB of DDR5 in new servers through its own CXL chip (The Register, 29 June 2026), and its ISCA 2026 paper reportedly says its servers last three to five years while the DRAM in them is good for seven to ten. Retired DDR4 kept in-house plausibly means fewer ex-data-centre modules reach resale, an inference with no volume figure behind it. And recovered chips are now a traded commodity: TrendForce’s spot report of 16 September noted that “reballed chips sustained the upward price momentum”, which is one reason provenance matters, as DDR4, used memory and upgradeability explains.

This site is the nearest thing to a used-price series, read at the moment you look: pages such as DDR4 registered rank listings by price per gigabyte, show condition per listing and hide bid-only auction rows by default, because a bid is not a price (?auctions=no removes auctions altogether). How to test what arrives is covered in buying used RAM on eBay.

HP says memory went from a sixth of a PC’s cost to more than a third

HP said on its fiscal first-quarter call in late February 2026 that memory had risen to about 35% of the cost of materials for a PC, from 15% to 18% the quarter before, and its chief financial officer expected memory costs to rise about 100% sequentially into the next quarter. Other sources measure different things:

TrendForce’s September figure implies what the rises did to the whole bill:

1Q25:  core parts 45% of the bill, everything else 55%
3Q26:  core parts 68%             -> everything else 32%
if everything else costs the same, the bill = 0.55 / 0.32 = 1.72 x the 1Q25 bill
  -> a 72% higher bill; TrendForce's own estimate of the price rise needed
     to keep 1Q25 margins is about 80%

The gap between 72% and 80% reflects assumptions the release does not detail. TrendForce added that cheaper inventory is running out, so later products will reflect current component prices.

Prices moved accordingly. Dell was expected to raise prices by at least 15-20% from mid-December 2025 (industry sources via TrendForce), and its chief operating officer said in February that DRAM prices had risen nearly 5.5 times in six months, to $2.39 a gigabit. Apple raised Mac and iPad prices on 25 June 2026, the 13-inch MacBook Air from $1,099 to $1,299, saying it had “never seen a component price increase this much, this quickly.”

Mercury Research found desktop CPU shipments down more than 20% in the second quarter of 2026. The makers’ answer was mix: Lenovo shifted “towards premium to improve average unit revenue”, HP spoke of “accelerating product reconfiguration”, and Gartner expects the sub-$500 PC to disappear by 2028. The cheap configuration is what disappears: a laptop at 2025’s price now tends to carry less memory, and one with the same memory costs more. Apple’s cheapest desktop shows it: in May 2026 Apple dropped the $599 Mac mini with 256GB of storage, leaving $799 as the entry model (CNBC, 25 June 2026). High-end ultrathin laptops with soldered memory were the segment most prone to early and significant price pressure, TrendForce said in December 2025, because they “cannot reduce costs by lowering specifications or replacing modules”; the upgradeable alternatives are in laptop upgradeability, the modules themselves under DDR5 laptop and DDR4 laptop, and how much to buy in how much RAM you need.

One US-specific layer is tariffs. IDC wrote in February 2026 that the Supreme Court had struck down the reciprocal tariffs and that a 10% across-the-board tariff had replaced them. That tariff, under Section 122 of the Trade Act, took effect on 24 February and lapsed after its 150 days on 24 July 2026, replaced the same day by Section 301 duties of 10% or 12.5% on imports from about 60 economies, with exemptions that include goods already under Section 232 tariffs (Holland and Knight, 30 July 2026; Global Trade Alert, 24 July 2026). Whether a given module or PC is exempt depends on its classification and origin, which this article has not checked.

Phones, consoles and small computers paid in different ways

One sourced line per segment:

The phone figures differ because segments, dates and counts (production against shipments) differ. Samsung shows both sides of the shortage in one set of accounts: its phone division made an operating loss of KRW 0.7 trillion in the second quarter of 2026, blaming “elevated component cost pressures”, while its chip division made an operating profit of KRW 89.2 trillion.

Raspberry Pi gives the cleanest decomposition, because it published its own memory cost. The Register reported on 25 September 2026 that the company’s average memory inventory cost rose from $3.60 per gigabyte at the end of 2025 to $13.30 at the end of June 2026. The 16GB board cost $145 after Raspberry Pi’s price rise of 1 December 2025, so the comparison can run over nearly the same months:

price rise, 16GB Raspberry Pi 5, Dec 2025 to Sep 2026:   305 - 145 = $160
memory cost rise, end-2025 to end-June 2026:
                    16 GB x (13.30 - 3.60) = 16 x 9.70 = $155
155 / 160 = 0.97  -> memory explains nearly all of the rise since December 2025

That uses an average inventory cost, not the marginal price of the next chip, and the two periods overlap rather than match exactly.

Graphics cards held up comparatively well in Europe, in 3DCenter’s judgement, and in the US memory was one of several pressures on GPU prices. It showed in supply as well: TechSpot reported on 15 August 2026 that “tightening stocks of GDDR7 VRAM” had raised prices of Nvidia’s RTX 5000-series cards and delayed the RTX 5000 Super, and that Nvidia had revived the 12GB RTX 3060 as a stopgap.

Flash is on a different path: TrendForce expects NAND to loosen in the second half of 2027 while DRAM stays tight, though ComputerBase’s check of German shop prices on 16 September 2026 still found SSDs 129% and hard drives 134% dearer year on year. Drive prices are tracked on the sister site’s SSD, NVMe and hard drive pages.

Servers pay the most per gigabyte, and buyers without contracts pay the margin

Server memory rose further and for longer than desktop memory:

contract multiple, mid-2025 to end-2026:  1.64 x 3.70 = 6.07
a memory budget that bought 1,000 GB in mid-2025 buys 1,000 / 6.07 = about 165 GB
  at end-2026 contract prices, if the forecast holds
module spot, 14 Sep 2026:  RDIMM 64.06 / UDIMM 14.73 = 4.35
  -> a DDR5 RDIMM costs about 4.3 times a desktop DDR5 module per gigabyte

The 270% is a forecast. What turns it into a problem for ordinary businesses is who pays it. Several US cloud providers signed multi-year agreements that “restrict suppliers from raising prices for these clients”, and TrendForce expects that from the third quarter of 2026 “the primary source of server DRAM price increases will shift toward customers without LTAs”. TrendForce does not name them; they are, by definition, DRAM buyers who have not signed one, and mid-sized enterprises meet the increase at one remove, through the server makers they buy from. TrendForce also expects server DRAM prices to keep rising every quarter through the second half of 2027. Buyers without a contract, and the businesses downstream of them, now set the price of last resort.

Buyers shrank memory to cope. Cloud providers and server makers cut RDIMM capacities in the first half of 2026, some systems moving from 96GB and 128GB modules to 32GB and 64GB, Micron said its higher server forecast was enabled by “a modest reduction in average server DRAM content growth”, and Nvidia halved the SOCAMM capacity of its Vera Rubin modules. Prices reached finished servers with a lag. In European distribution, the channel-data firm Context found component RAM up 63% between September and December 2025, OEM enterprise RAM up about 28% and whole servers only 5-10% by December. HPE cut quote validity from 30 to 14 days in February 2026, allowed repricing between quote and shipment in March, and by August was honouring quotes up to $1 million until shipment. Meanwhile, TrendForce reported in September that server buyers were “aggressively accumulating inventory”.

The pages to read are DDR5 server, DDR5 registered, MRDIMM, DDR4 registered and memory by server model. Server RAM vs desktop RAM and why your server won’t take its maximum cover the module types and ceilings, and capacity planning under these prices belongs to how much RAM you need.

Long-term agreements turned a spike into a floor

In January 2026 Hankyung reported that Samsung and SK hynix were rejecting two- and three-year agreements in favour of quarterly contracts, expecting prices to keep climbing through 2027. By the middle of the year multi-year deals were being signed, with take-or-pay volumes, deposits and price floors for the sellers, and ceilings at or near 2Q26 prices for the largest buyers.

Micron says 14 of its 16 agreements carry “cumulative revenue at minimum price” of about $100 billion, and that the floor “enables a very robust gross margin for Micron, well above our peak quarterly margins in any past cycle”. As a rough average:

$100 bn of minimum revenue over the remaining term, mid-2026 to end-2030
  (about 4.5 years) = about $22 bn a year, from 14 agreements
all 16 together cover about 20% of Micron's DRAM volume and a third of its
  NAND, so the $100 bn is not DRAM revenue alone

A widely repeated claim that “most” of Micron’s DRAM is contracted to 2030 runs ahead of the signed deals: the 16 agreements signed by June cover about a fifth of its DRAM volume. Micron expects about half or more of its company revenue to fall under such agreements once the programme is complete, and about 40% of revenue to carry fixed prices or ceilings at or close to calendar 2Q26 levels; volume and revenue are different measures, and the figures can all hold. A floor set at the top of a record cycle limits how far contracted prices can fall. Nor did ceilings cap everything: TrendForce says HBM contract prices “could still rise by 70-140% in 2027” despite them, and its bulletin of 23 September put the other consequence plainly: “LTAs shrink negotiable supply as new fabs lag”. Micron also expects its blended DRAM cost per bit to rise, with new product generations (LPDDR6, DDR6, newer HBM) and new greenfield fabs, which argues against a return to the lows of 2023 to 2025.

Record profits, and the price-fixing question the profits invite

The suppliers’ own accounts show where the money went:

Micron revenue:  41.46 / 9.30 = 4.46 times a year earlier
Micron gross margin (non-GAAP):  56.8% -> 74.9% -> 84.9% -> about 86% guided

In June 2026, 17 US consumers filed a proposed class action in the Northern District of California (No. 3:26-cv-06345), alleging, in TrendForce’s summary of the complaint, that Samsung, SK hynix and Micron deliberately restricted commodity DRAM supply from 2022 “under the pretext of expanding HBM production for AI” and drove prices up by about 700% over four years. The plaintiffs argue that in a competitive commodity market at least one producer would have expanded output when prices rose so sharply, and that this “did not happen”. These are allegations, reported in June and untested in court. There is a precedent. In October 2005 the US Department of Justice announced that Samsung had agreed to plead guilty and pay a $300 million fine for fixing DRAM prices between April 1999 and June 2002, the third company to do so after Hynix, fined $185 million, and Infineon. Micron was not among the companies that pleaded guilty, though it is sometimes listed with them.

The case against collusion rests on the record: losses in 2023, a cleanroom constraint that money cannot quickly relieve, capital spending rising rather than falling (Micron’s went from about $18 billion planned for fiscal 2026 to about $27 billion), and a supplier saying publicly that it could not meet demand. The case for suspicion rests on margins, on Acer’s accusation, and on the industry’s own vocabulary: TrendForce’s July 2026 release credited rising prices for consumer DRAM, its category for the legacy chips in televisions and set-top boxes, partly to “disciplined production cuts by leading memory manufacturers” that Taiwanese DDR4 expansion could not offset, and SK hynix’s second-quarter release promised “CapEx Discipline”. What separates disciplined supply from coordinated supply is a question for a court, and nothing public settles it.

New fabs take three to five years, and a cleanroom opening is not supply

Money is not the constraint. Hyundai Motor Securities wrote in February 2026 that new Korean capacity would not come online until the second half of 2027, and SK Securities that “capacity expansion remains restricted by a lack of available cleanroom square footage”. Micron’s June 2026 presentation adds construction lead times, a shortage of skilled trades, permitting and energy infrastructure. TechInsights’ James Sanders told The Register in December 2025 that adding capacity takes about three years “if you’re rushing”, and that the big three are “looking at four or five years to ramp production at a new fab.” The shorter figures readers meet describe an earlier stage: IEEE Spectrum’s “18 months or more” (February 2026) is for building a fab and getting it running, while Moorhead’s “three or four years to build a memory fab and get volume out of it” (PCWorld, April 2026) runs to volume. Meanwhile the big three grow bits mainly by migrating to finer processes, TrendForce says, and in July it wrote that “construction schedules, equipment installation, and raw material preparation will delay meaningful production ramp-ups until the second half of 2027”, with “substantial output contributions” not until 2028. Spending caught up late: in November 2025 TrendForce expected DRAM capital spending to rise only 14% in 2026, to $61.3 billion, and Micron has since raised its own fiscal 2026 budget by half, as the previous section showed, though its figure covers all products and construction and is not comparable with TrendForce’s DRAM total.

Each announcement moves through four stages: a cleanroom opens, equipment goes in, the first wafer comes out, and then volume ramps over several quarters. The dates below are the stage each source names.

Two corrections follow. M17’s December 2028 cleanroom is NAND for SSDs, not DRAM relief; it will show on the sister site’s SSD page. And Samsung’s P4 figure is mostly converted lines, adding bits through a finer node rather than new wafers, which is the general pattern:

2026 industry DRAM bit shipments:  up in the low- to mid-20s % (Micron, Jun 2026)
DRAM wafer capacity:  up 4.8% a year on average to 2030 (SEMI, Feb 2026)
1.23 / 1.048 = 1.17  -> roughly 17% more bits per wafer, a rough figure that is
                        already net of HBM's larger share of wafers

The first date on any fab announcement is the start of a ramp, not the arrival of supply. SK Group’s chairman Chey Tae-won gave the long view in June 2026: SK hynix’s wafer capacity should double within five years and triple by about 2034, with the first new phase running as early as 2027, and “there is currently no way to move faster than this.”

China’s new capacity is real and mostly spoken for

CXMT’s share of DRAM revenue rose from 7.6% in the first quarter of 2026 to 9.5% in the second (TrendForce’s news desk, 24 September). It raised about $8.6 billion in a Shanghai listing in July, is considering a sixth fab, according to Reuters as relayed by Tom’s Hardware, and on 20 September put its G5 process into mass production, claiming at least 50% more dies per wafer. How much it can make is estimated, not disclosed, and the estimates conflict:

Most of it is spoken for at home. Institutional investors quoted by Commercial Times expect the new capacity to go mainly to China’s own AI and cloud build-out and to remain “insufficient to reverse the global supply-demand imbalance”, with no meaningful effect before 2027. Nanya’s president said in March that Chinese makers serve domestic, lower-end demand. Policy narrows access further. The US controlled HBM exports in December 2024, and as of June 2026 CXMT was not on the Commerce Department’s Entity List but was on the Pentagon’s 1260H list, while the Financial Times reported Apple lobbying for permission to buy from it. In December 2025 the US Trade Representative also set a new tariff on Chinese semiconductors, aimed at mature-node chips, at 0%, rising on 23 June 2027 to a rate to be announced, on top of an existing 50% tariff (CBS News, 24 December 2025); whether it will reach memory was not established.

Acer, HP and Asus already use CXMT chips in some products, and in February TechSpot found a CXMT-based DDR5-6000 kit matching a premium kit in games. Memory built on Chinese chips was not cheap, though: a Lexar DDR5-7200 32GB kit using Chinese DRAM listed in China in August at 3,999 yuan, about $592, in line with comparable US kits. A headline that CXMT “targets 30% by 2030” reports an investor’s opinion, not a company target. The former Samsung chip head Kye-hyun Kyung argued in May that prices could fall from the second half of 2027 if Chinese capacity delivers. Chinese supply is the most plausible early brake on prices and the least certain one.

When it ends: every dated forecast of the end, side by side, with each forecaster’s interest

“When does it end” is three questions: when supply catches up with demand, when prices start falling, and when, if ever, they return to 2025 levels. The table says which one each forecast measures, and who has what to gain.

Kearney’s widely quoted “until 2030” is left out because its report could not be read at source; the 2030 dates above are SK hynix’s and its parent’s.

Count the rows that answer the supply question. TrendForce, Micron, Samsung, IDC and Nanya, five independent sources of two kinds (suppliers and analysts), place no meaningful relief before 2028, and none of them expects it in 2026; Intel’s chief executive gave the same date, but as what two memory makers had told him. That clears the bar this article sets for the word, at least three independent sources saying the same thing, so, narrowly: the consensus is no relief in 2026, probably none in 2027, and supply catching up in 2028 at the earliest. The dissent comes mostly from buyers, Acer above all, and from TechInsights’ December 2025 call of a 2026 peak. The optimistic 2028 models are thin: TrendForce puts supply about 2% ahead of demand, JP Morgan has supply growing about 3 points faster than demand that year, which is a growth rate rather than a surplus, and Citi keeps DRAM short in both 2027 and 2028.

The steepest named fall is Bernstein’s, and even that leaves prices high. Its units are reported inconsistently (per gigabyte by PC Gamer, per gigabit by ComputerBase), so the figures below are ComputerBase’s, all in one unit, and only their ratios are used:

Bernstein via ComputerBase, SK hynix DRAM price per gigabit:
2025  $0.54    2026  $1.56    2027  $2.23    2028  $1.05
2027 peak:  2.23 / 0.54 = 4.1 times 2025   (PC Gamer: "almost five times")
2028:  1.05 / 2.23 = 0.47 of the peak   -> a fall of about 53%
2028 against 2025:  1.05 / 0.54 = 1.9 times

PC Gamer’s own reading is “only a near 100% increase from 2025”, and ComputerBase puts 2028 between the 2025 and 2026 levels: about twice 2025, either way. No forecaster on the record dates a return to 2025 prices as such. The nearest are AMD’s David McAfee, who said at Computex 2026 that DDR5 was about two years from normal pricing, which points to mid-2028, and Moorhead, who does not expect normality until 2030; neither defines normal as the 2025 level. IDC’s model does not reach 2025 pricing, Lenovo says never, and Micron says its contract floors give it a gross margin “well above our peak quarterly margins in any past cycle”. They say different things, so they are not a consensus, but they are the whole of the evidence. “When it ends” has an answer for supply; for price there are only loose dates for “normal”, and none tied to the 2025 level. The quarter-by-quarter paths for 2026 and 2027, forecaster by forecaster, are in buy RAM now or wait.

What “2025 prices” means is easiest to hold in the unit this site uses. On the PCPartPicker averages, a mainstream 32GB desktop kit cost about $3.3 to $3.4 per gigabyte in August 2025, DDR4 and DDR5 alike. Bernstein modelled one supplier’s chip price, not retail kits, so carrying its ratio to the shelf is an assumption, made only to give a sense of scale:

DDR5-6000 2x16GB, August 2025 average:     108 / 32 = 3.4 $/GB
x 1.94, the Bernstein-implied 2028 level:  3.375 x 1.94 = about 6.6 $/GB
DDR5-6000 2x16GB, August 2026 average:     572 / 32 = 17.9 $/GB

Those three numbers are a yardstick for the medians on DDR5 kits totalling 32GB and DDR4 kits totalling 32GB. Read the page on the day: the distance from about $3.4 per gigabyte is the shortage premium still in a new kit’s price.

What would end it sooner, and what would make it last

The last supercycle shows what a real turn looks like. TrendForce reported DRAM revenue down 18.3% in the fourth quarter of 2018, and in March 2019 revised its first-quarter contract forecast from a fall of 25% to one of nearly 30%, the sharpest quarter since 2011. Prices were still falling at the end of 2019, and only in January 2020 did TrendForce expect contracts to turn flat to slightly up: about five quarters of decline, reading those releases together. As a thought experiment, not a forecast:

undo a 5.35x rise at the 1Q19 rate of fall, about -30% a quarter:
ln(5.35) / ln(1 / 0.70) = 1.677 / 0.357 = 4.7 quarters
check:  5.35 x 0.70^4 = 1.28      5.35 x 0.70^5 = 0.90

The triggers that would shorten or lengthen the shortage, with the signal to watch for each:

A slump in AI spending is the main route to an early end, and it would reach the price layers in order of how loosely they are contracted. Spot and retail would move first, as the spring of 2026 showed on stock clearing alone. Contracted volume would move last: Micron’s agreements are take-or-pay, with floor prices in the largest, on about a fifth of its DRAM volume. TrendForce named the risk in July 2026: some cloud providers had reported negative free cash flow, and memory staying expensive through 2027 may affect their investment plans. Thomas Coughlin of Coughlin Associates, a storage and memory analyst quoted by IEEE Spectrum in February 2026, warned of “another of the boom-and-bust cycles for this industry as demand drops at the same time as supply increases”, and the fabs in the supply table arrive from 2027 whether demand does or not. A slump would lower spot and retail prices before contract prices, and contracted prices no further than their floors.

Efficiency is weaker than it looks. Google’s TurboQuant, published in March 2026, compresses the key-value cache used during inference, not the model, and The Register judged it unlikely to end the crunch. The headline “six times” comes from squeezing that cache to about 2.5 bits per value, when 8-bit caches were already common, so against current practice the saving is nearer 8 / 2.5 = 3.2 times, and PCWorld reported in April 2026 that TrendForce expects cheaper inference to raise usage rather than cut memory demand.

Device prices will lag on the way down. The makers have not passed on the whole rise: TrendForce estimated on 4 September 2026 that a $900 notebook would need a price about 80% higher to keep its 1Q25 margin, against the 18% Apple added to the 13-inch MacBook Air and the 15-20% expected of Dell, which suggests margin will be recovered before prices are cut. And memory is only part of the bill:

MacBook Air, 25 Jun 2026:   1,299 / 1,099 = 1.18   -> +18%
materials = 0.65 other parts + 0.35 memory   (HP's share, Feb 2026)
memory price halves:  0.65 + 0.175 = 0.825   -> materials cost down 17.5%

The economist Mina Kim of Mkecon Insights, quoted by IEEE Spectrum in February 2026, put it generally: “Prices come down much more slowly and reluctantly than they go up.” Acer expects PC prices to plateau in the first half of 2027 before any decline. When memory falls, a finished device falls later and by less.

Flash, as the section on phones and consoles noted, is likely to turn first, and NVMe drives should diverge from memory from the second half of 2027. Apart from Chinese capacity, the least certain brake, every path to an early end runs through demand falling, not supply rising. How to act on that is in buy RAM now or wait.

Twenty questions people ask, each answered in brief

The questions that recur across the pages that rank for this subject, answered from the evidence above.

Why is RAM so expensive right now? The big three gave wafers to HBM and server DDR5 first, and a bit deficit of 1% to 2% met buyers who barely respond to price.

Why does HBM affect my desktop RAM? It is made on the same wafers, and a gigabyte of HBM takes roughly three gigabytes’ worth of wafer.

Why is old DDR4 sometimes dearer than DDR5? End-of-life notices in 2025 cut supply. DDR4 chips still cost about 45% more than DDR5 chips at spot, although DDR4 kits cost less per gigabyte at retail.

Is this real, or price-fixing? A class action filed in June 2026 alleges coordination. It is untested, and the proven case dates from 1999 to 2002.

When will prices go down? Supply is forecast to catch up in 2028 at the earliest. Nobody on the record dates a return to 2025 prices as such; the nearest, AMD’s David McAfee, put normal DDR5 pricing about two years out from June 2026.

Prices dipped in the spring. Is it over? No. That was inventory clearing: the cheapest 32GB DDR5-6000 kit crossed $400 in August, and German kit prices reached a new high in September.

Would an AI bubble bursting fix it? It is the main route to an early end. Spot and retail prices would fall first, contracted prices last and no further than their floors, and the 2018-19 bust ran for about five quarters.

Does TurboQuant reduce memory demand? It compresses inference caches, not models. Against the 8-bit caches already in use the saving is nearer three times than six, and TrendForce expects extra usage to absorb it.

Will China bring prices down? Possibly from 2027, but most of CXMT’s output serves China and US policy limits access to it.

How long does a fab take? About three years if rushed, and four or five to ramp a new fab, on TechInsights’ estimate. A cleanroom opening is only the first of four stages.

What happened to Crucial? Micron stopped Crucial retail shipments in February 2026, and warranty service continues; the detail is in DDR4, used memory and upgradeability.

Are SSDs, hard drives, GPUs, consoles, phones and Macs affected? Yes, by different amounts: in German shops SSDs were 129% and hard drives 134% dearer in September 2026 than a year earlier, and GDDR7 supply delayed Nvidia’s RTX 5000 Super. Drive prices are on the sister site’s hard drive page.

Will laptops get cheaper again when memory does? Later and by less: halving memory’s price cuts a PC’s materials bill by about a sixth, the makers have margin to recover first, and the cheaper component inventory is running out.

Is a pre-built cheaper than building now? Sometimes, while the maker’s earlier, cheaper memory lasts, and that inventory is running out. The arithmetic is in buy RAM now or wait.

Should I buy now or wait? Only TechInsights’ December 2025 forecast puts the price peak in 2026, and no supply forecast expects relief this year; see buy RAM now or wait.

How much RAM should I buy while it is expensive? What your workload measures, not a round number: how much RAM you need.

Is used or DDR4 RAM a way round it? Partly. DDR4 kits cost less per gigabyte at retail, and one seller’s index shows DDR4 server memory falling. See DDR4 as an escape route and buying used RAM on eBay.

Why do server and ECC modules cost so much more per gigabyte? RDIMMs are what the cloud providers buy; at module spot in September 2026 a DDR5 RDIMM cost about 4.3 times a desktop DDR5 module per gigabyte, and buyers without long-term agreements carry the contract rises.

What is an LTA, and how do businesses get allocation? A multi-year supply agreement, often take-or-pay with a floor price; Micron’s 16 cover about a fifth of its DRAM volume. Allocation follows those agreements first and module makers last.

Which price number should I trust? The one whose layer and date match your purchase: a retail or listing price for a kit, never a chip’s spot price.

Coping is a question of timing, capacity, generation and fit, at home and in a business

Timing. If the machine is needed now, waiting has no dated reward: no supply forecast above expects relief in 2026, TrendForce lifted its fourth-quarter outlook in September (to +10-15%, a slower climb than the third quarter’s 13-18%, but a climb), and the earliest named falls are in 2027 and conditional. If the current machine works, the case for waiting rests on supply in 2028, not on a dip. Either way, a finished system can carry cheaper memory than a loose kit while the maker’s older stock lasts. The full decision is in buy RAM now or wait.

Capacity. Every step up in capacity is now priced at shortage rates, so the step matters more than it did: how much RAM you need.

Generation and condition. DDR4 is not a safe haven, but it does cost less per gigabyte at retail; a laptop with soldered memory is bought once; used modules are tested, not trusted. See DDR4, used memory and upgradeability and buying used RAM on eBay.

Fit. Confirm the machine before the module, with how to check what RAM fits, the CPU index or memory by server model.

Businesses. Large buyers have already taken five steps: quotes that hold to shipment, which HPE offered from August; longer refresh cycles, with Gartner expecting business PC lifetimes to lengthen by 15% by the end of 2026; modules sized to the workload, as cloud providers did with some systems; reuse of what is owned, as Meta does with recovered DDR4; and a second qualified source, as HP reported in February. Without a long-term agreement, an older platform on DDR4 registered is worth pricing before a DDR5 refresh, and capacity planning is in how much RAM you need.

What to do with this on a listing page

1. Read the view’s cheapest and median price per gigabyte, not one kit. The gap between them is the market’s spread.
2. Compare generations per gigabyte on every generation by price per gigabyte, or on DDR5, DDR4 and DDR3, before assuming the older one is cheaper.
3. Remember that bid-only auctions are hidden by default, because a bid is not a price. ?bids=show brings them back.
4. For servers, expect the RDIMM premium. Compare DDR5 registered with DDR5 desktop per gigabyte for the premium, and DDR5 server with DDR4 registered for the generation gap.
5. Use part pages to compare one part number across sellers and conditions.
6. Check condition on each listing and test used modules as buying used RAM on eBay describes; the methodology page says what the site records.
7. Re-read the page on the day you buy. Nothing in this article freezes a site price.

The wafers went to AI first, as HBM and then as the server DDR5 around it; the supply gap is forecast to close in 2028 at the earliest; and nobody on the record puts a date on prices returning to 2025 levels as such. That is the state of the evidence as of late September 2026. The next dated piece of it is Micron’s fiscal fourth-quarter report on 30 September, and the right way to read it, like every figure above, is by its layer and its date.