Buy RAM now or wait? What the 2026–2027 forecasts actually say

Do not wait for RAM to get cheaper, because nobody on the record expects it to. If you will need the memory within the next year, buy it no later than you need it: none of the forecasts traced for this article expects RAM to get cheaper before the end of 2026, all three of the largest memory makers expect supply to stay tight through 2027 and two of them say it gets worse, and the earliest year any published model puts DRAM supply back above demand is 2028. If the machine you own works and you have no deadline, waiting is sensible, but for a different reason from the one usually given, and it only pays if you can wait a long time.

The most quotable answer in this year’s buying advice is two sentences long: if you need a PC, build it now with 32GB of DDR5-6000; if you have a working machine and no urgency, hold for the late-2026 platform refresh and re-evaluate then. It traces, almost word for word, to a single buying guide first published in May 2026 and updated in September, and it is half right. The buying half survives the evidence: two modules, 32GB, and DDR5-6000 when it costs no more than DDR5-5600 or within a few per cent of it, for reasons AMD’s and Intel’s own specification pages explain. The calendar half does not. Intel’s next desktop platform is dated anywhere from the end of 2026 to the spring of 2027 depending on whose account you read, AMD’s is reported for 2027, and both are reported to stay on DDR5, so memory bought now carries straight across. The refresh is a reason to wait for a processor, not for memory.

What waiting is worth can be put in dollars. On the most optimistic price path any named analyst has published, a buyer who can go without a 32GB kit until the end of 2028 saves roughly $4.40 to $6.60 for each month of doing without. A buyer who will need the memory within a year and waits anyway pays more: if retail follows contract prices as it has for the past year, the one quarter still ahead in TrendForce’s forecasts, the fourth, adds about $41 to $61 to a $409 kit by December, and Bernstein’s path, the only one that later falls steeply, first takes the same kit $84 to $205 higher at its end-2027 peak.

This article sets out the forecasts one at a time, for the rest of 2026, for 2027 and for 2028 and later, with who made each and what it measures, since a contract-price forecast, a supplier’s outlook and a PC maker’s view of laptop prices get quoted as one. It then tests them: what each relief date needs, the forecasters’ record, the eight numbers called “the price of RAM”, and the arithmetic of waiting on one real purchase. The practical half covers the kit to buy, the platform refresh with each date labelled announced, reported or leaked, holding with an end date, six buyers, prebuilts, nine ways to pay less, enterprise purchasing, the signals of a turn, the used market this site indexes, and eight beliefs the evidence does not support. It is long because the question has four layers - the forecast, the price you would actually pay, the platform, and your own timeline - and most advice answers one of them and assumes the rest.

A note on sources. Every market figure below names its publisher, the document and the date in the sentence, every forecast is labelled as a forecast with who made it and when, and where a figure is worked out here the arithmetic is shown so that you can check it. Where sources disagree, both are given with the range. Where a figure reached print only through a relay - an analyst note quoted by a news site quoting a social media post - the text says so. Everything is as of late September 2026, with one later addition: TrendForce’s fourth-quarter 2026 DRAM contract forecast, published as a press release on 30 September and added here on 2 October. Micron’s fiscal fourth-quarter results, released the same day, are not yet worked in. Prices on this site move daily and are never quoted in the text; where the article needs today’s number, it tells you which page to read.

The answer turns on when you will need the memory

“Buy now or wait” is really “when will you need it”, because the forecasts disagree about 2028 and very nearly agree about the next twelve months. So the decision comes first, before the evidence.

Only one statement in this article meets the test for a consensus, three or more independent sources saying the same thing, and it meets it several times over: there is no relief in 2026, and supply gets no easier in 2027. TrendForce, Micron, Samsung, SK hynix, Winbond, Adata and IDC all say so, in documents dated between February and August 2026 and cited below. The dissent is real. Acer’s chairman and chief executive Jason Chen told reporters in September, as DigiTimes reported, that PC prices will plateau in the first half of 2027 and then fall, and accused memory makers of talking up the shortage to protect their margins. Even he expects PC prices to keep rising until the end of 2026, so the most optimistic executive on record does not put the fall this year.

The table assumes you know two things: how much memory you need, which how much RAM you actually need answers by workload for home and enterprise machines, and what your board will take, for which how to check what RAM fits is the procedure. Get either wrong and the timing stops mattering.

Nobody on the record expects cheaper RAM before 2027

The forecasts for the rest of 2026 are quarterly contract-price calls, and the most cited of them come from TrendForce, the Taiwanese market-research firm whose DRAMeXchange unit publishes the industry’s reference price series. Its release of 3 July 2026 forecast conventional DRAM contract prices - TrendForce’s term for all DRAM except high-bandwidth memory - to rise 13-18% quarter on quarter in the third quarter. TrendForce attributed the slowdown to consumer buyers of PCs and phones “reaching their affordability limit” and to a high comparison base, not to new supply.

For the fourth quarter, TrendForce’s release of 30 September forecast conventional DRAM contract prices up 10-15% quarter on quarter: slower again, but still a rise. It expects server DRAM to stay undersupplied, PC makers to keep buying ahead of a tight 2027, and mobile DRAM’s increases to moderate because large rises are already priced in. Until that release the only public fourth-quarter DRAM number attributed to TrendForce was a PC DRAM range of 3-8%, raised from 0-5%, which reached print on 30 June through KuCoin News relaying the Chinese outlet ChainThink; TrendForce’s research pages of 23 and 24 September had since said it raised its outlook. The claim, made in at least one long buying guide, that TrendForce’s 3 July release forecast 3-8% for all conventional DRAM in the fourth quarter is wrong: that release contains third-quarter figures only.

The spread in that table is the story. The module makers, Adata and Apacer, buy chips and sell the DIMMs found in shops, so they see the squeeze first, and they have an interest in firm prices. At the top sits a single unnamed memory analyst quoted in a Jefferies report and passed from a post on X through Wccftech to TechRadar. That one source is the whole basis of the line “analysts expect another 40-50%”, which at least two buying guides still carry as if it were a consensus. Ten days after it appeared, TrendForce published 13-18%.

Rest of 2026 at contract level
(3Q26 then 4Q26; TrendForce's are conventional DRAM, the others do not say)

  TrendForce, low     1.13 x 1.10 = 1.243    about +24%
  TrendForce, high    1.18 x 1.15 = 1.357    about +36%

  Aletheia            1.30 x 1.15 = 1.495    about +50%

  Jefferies-relayed analyst, low     1.40 x 1.30 = 1.820    +82%
  Jefferies-relayed analyst, high    1.50 x 1.40 = 2.100    +110%

Percentages compound; they do not add. On TrendForce’s figures the contract level at the end of 2026 sits about 24% to 36% above the second-quarter level, and the outlier’s path is about three times that, like for like (+82% against +24%, +110% against +36%). The third-quarter part is largely behind us, since contracts for the quarter were settled by late September and, as the arithmetic section shows, shop prices have kept pace with contracts; the fourth-quarter part is the pressure still moving towards the shelf.

• TrendForce, conventional DRAM
• Aletheia Capital, 23 Jun
• Analyst relayed by Jefferies, 23 Jun
• 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%); TrendForce, 30 Sep 2026 (4Q26 forecast: +10-15%); TechRadar, 23 Jun 2026 (Aletheia: +30%, then up to +15%; the analyst: +40-50%, then +30-40%). Figures as of October 2, 2026. Only 4Q25 and 1Q26 are reported; from 2Q26 on every point is a forecast, because TrendForce has published no realised 2Q26 change. The other two paths start from TrendForce's 2Q26 level, do not say which price they measure, and take Aletheia's 4Q26 at its top, 15%.

Show the figures

DRAM contract prices so far, and three forecasts for the rest of 2026 (Index, 3Q25 = 100, compounded from quarterly changes at the middle of each range)

Period	TrendForce, conventional DRAM	Aletheia Capital, 23 Jun	Analyst relayed by Jefferies, 23 Jun
3Q25	100	-	-
4Q25	147.5	-	-
1Q26	288.4	-	-
2Q26	462.8 (forecast)	462.8 (forecast)	462.8 (forecast)
3Q26	534.6 (forecast)	601.7 (forecast)	671.1 (forecast)
4Q26	601.4 (forecast)	691.9 (forecast)	906 (forecast)

Spot prices, which move first, cooled in mid-September, with a few chip prices edging down but no broad fall. TrendForce’s spot note of 16 September said quote enquiries for DDR5 chips from the major suppliers had slowed and that “prices for certain chips have begun to show signs of softening”, while the mainstream DDR4 8Gb chip still edged up to $45.54. One week of softening is not a turn.

On finished machines the direction is the same. Acer’s Jason Chen expects PC prices to rise 5-20% towards the end of the year (DigiTimes, relayed by Tom’s Hardware on 21 September). And PC Games Hardware, reading TrendForce’s September update on 25 September, concluded that waiting for cheaper DDR5 in the fourth quarter is not worth it.

No named forecast traced for this article expects retail RAM prices to fall before 2027, and on this point the sources pass the consensus test with room to spare. Why contract prices are rising at all - the wafers that high-bandwidth memory for AI accelerators takes from ordinary DRAM, and the suppliers’ decisions about who gets what - is the subject of the shortage explainer.

In 2027 the makers expect worse, and one PC maker expects better

For 2027 the forecasts become statements about supply rather than quarterly price calls, and who is speaking matters as much as what is said, so the table gives each forecaster’s role.

The sellers speak with one voice, and they have reasons to. SK hynix’s chief executive spoke on the day the company listed on Nasdaq, as Tom’s Hardware noted on 11 July. Samsung’s written release of 30 July commits only to the second half of 2026; the 2027 and 2028 statements come from the analyst call.

Apacer’s warning matters most to retail buyers, because shop DIMMs are built by module makers from bought chips; Tom’s Hardware stressed on 29 July that it concerns allocation to module makers, not global output, and it is one executive’s view from a company with inventory to sell.

The buyers do not contradict the sellers: Microsoft, relayed by TechRadar on 26 June, said console memory and storage prices had risen more than 2.5 times, and IDC’s February outlook has the PC market falling 11.3% in 2026 and rebounding only in 2028.

The dissent deserves weight. Acer’s Jason Chen says only some high-end parts are genuinely short and that memory makers want “to keep their margins as high as possible for longer”; Acer buys memory and has an interest in lower prices, exactly as the suppliers have in higher ones. Bloomberg Intelligence’s view that the shortage may already have peaked, with possible oversupply in 2028, reached print only second-hand, through 24/7 Wall St on 6 July and TechTimes on 18 August, and the research itself was not available to check. James Sanders’ call for a 2027 settling is the most optimistic named forecast found and also the oldest, made before the steepest quarter of the cycle.

Talk runs both ways; behaviour is better evidence. Winbond said in August that customers are seeking long-term agreements covering capacity beyond 2028, some already negotiating for 2029 and 2030. Micron said on 24 June it had signed 16 strategic customer agreements, typically running from 2026 to 2030, the largest with a ceiling at the calendar second-quarter 2026 market price and a floor for the whole term. SK hynix said on 29 July it had finalised long-term agreements with about ten customers. People who hold order books do not sign five-year supply agreements with price floors when they expect a glut the following year. The optimistic views on record put improvement in the last quarter of 2027 at the earliest, and the shortage explainer covers why the suppliers can hold their position.

2028 is the first year any model puts supply above demand, and not every model does

Three models on the record put DRAM supply back above demand in 2028. TrendForce’s, presented at the Future of Memory and Storage conference in August and reported by Tom Coughlin for Forbes (republished by VDURA, 18 August), has supply and demand “unbalanced until 2028 when there would be 2% more supply than demand for DRAM”, and 4% more for NAND flash. JP Morgan’s, as reported by MakeUseOf on 25 September, has DRAM supply growth beating demand growth by roughly three percentage points in 2028 after shortfalls in 2026 and 2027. Bloomberg Intelligence’s July research, reported second-hand, has a sufficiency ratio that may tip into oversupply in 2028. The same MakeUseOf report says Citi’s demand model keeps DRAM undersupplied in both 2027 and 2028. So 2028 is a date three models share and one major bank disputes, and none of the three is published in full. It is not a consensus.

Bernstein’s price path, the only named forecast of a steep fall, puts it in 2028 too. As reported by PC Gamer on 7 July, Bernstein expects SK hynix’s DRAM average selling price to reach $1.85 in 2026, peak at $2.23 at the end of 2027 and fall to $1.05 in 2028. PC Gamer gives those figures per gigabyte; ComputerBase, reporting the same note on 8 July, gives them per gigabit and describes the 2027 step as a rise of almost 50%, which implies a 2026 level nearer $1.50. They cannot both be right about the unit, so only ratios, which hold either way, are used here: 1.05 / 2.23 = 0.47, a fall of 53% from the peak. PC Gamer’s own reading is that the 2028 price would still be close to double the 2025 level.

The 2017-2018 boom gives a fall of that size a scale. Micron’s annual reports give the change in its average DRAM selling price per gigabit each fiscal year, which ends around 1 September: up 18% in fiscal 2017 and 36% in fiscal 2018, then down 30% and 34% in the two years after (Micron’s 10-K for fiscal 2021). Compounded, 1.18 x 1.36 = 1.60 on the way up and 0.70 x 0.66 = 0.46 on the way down, so by 2020 the price stood about a quarter below where the boom began, 1.60 x 0.46 = 0.74. That 54% fall from the peak is the size of Bernstein’s 2028 path, spread over two years rather than one. This cycle is of another order: in the first nine months of fiscal 2026 the same measure was up about 140% on a year earlier (Micron’s 10-Q of 25 June 2026), nearly four times the largest annual rise of that boom.

Sources: Micron 10-K, fiscal 2016 (2012-2016); Micron 10-K, fiscal 2021 (2017-2021); Micron 10-K, fiscal 2023 (a decline in the high-40s percent range); Micron 10-K, fiscal 2024 (a rise in the low-teen percentage range); Micron 10-K, fiscal 2025 (a rise in the low-40% range); Micron 10-Q, 25 Jun 2026 (about 140%, first nine months). Figures as of October 2, 2026. Micron gives 2023 to 2025 as ranges, plotted near their middle; 2026 is the nine months to 28 May against the same months of 2025; 2022 has no figure, since Micron put that year's 12% rise in DRAM sales down to bit shipments up slightly over 10%. The figures cover all of Micron's DRAM, HBM included.

Show the figures

The 2017-2018 DRAM boom lasted two years, and the two years after it took the rise back (% change in Micron's average DRAM selling price per gigabit on the previous fiscal year (2026: first nine months))

Fiscal year	Change on the previous fiscal year
2012	-45%
2013	-11%
2014	+6%
2015	-11%
2016	-35%
2017	+18%
2018	+36%
2019	-30%
2020	-34%
2021	+8%
2022	-
2023	-48%
2024	+12%
2025	+41%
2026	+140%

Suppliers, buyers and analysts outside the 2027 table cluster around the same year, and some push well past it.

The reason is physical, and TrendForce stated it plainly on 30 July: “construction schedules, equipment installation, and raw material preparation will delay meaningful production ramp-ups until the second half of 2027”, and “substantial output contributions are not expected to materialize until 2028”.

That gives the one rule needed to read any fab announcement: a cleanroom opening is not output. After the building come equipment installation and a ramp measured in quarters. The dates on the record, limited to the ones that bear on relief:

The M17 and Y2 rows correct the most repeated date error in this year’s coverage: the December 2028 cleanroom is M17, which SK hynix describes as a NAND fab, and the DRAM fab, Y2, opens its first cleanroom in June 2029.

NAND is expected to loosen first. TrendForce’s 30 July release projects NAND flash entering “a looser supply environment in 2H27” while DRAM stays tight, so SSD prices and RAM prices may part company during 2027. Until then SSDs are on the same path: TrendForce forecast NAND flash contract prices up 10-15% in the third quarter (3 July) and overall NAND up a further 15-20% in the fourth (24 September), so for a build the SSD argues for waiting no more than the memory does. The sister site tracks SSD prices by price per terabyte. The full account of the supply response, fab by fab, is in the shortage explainer.

Each relief date needs something specific to happen first

A date is only as good as the conditions behind it.

Relief by late 2027 needs demand to break, and TrendForce names the most plausible way: DRAM and NAND will take 47% of the largest cloud providers’ capital spending in 2026 and 68% in 2027 (25 August), some of those providers report negative free cash flow, and whether memory’s rising share changes their plans is “a key variable to watch” (30 July). Consumer demand is already falling: TrendForce expects notebook shipments down 9.4% in 2026 (4 September), and Mercury Research found desktop CPU shipments down more than 20% year on year in the second quarter (Tom’s Hardware, 22 August). Chinese capacity is the weakest route on current evidence. Commercial Times, relayed by TrendForce on 24 September, reported investors’ view that CXMT’s new output will mainly serve China. CXMT is not on the US Entity List but is on the Pentagon’s 1260H list of Chinese military companies (Tom’s Hardware, 27 June), which it sued to leave on 28 August (Reuters, relayed by AppleInsider); for now it is a reputational risk for American buyers. And Lexar’s 32GB DDR5-7200 kit on Chinese DRAM was listed in August at 3,999 yuan, about $592, in line with or only slightly below comparable US kits (Tom’s Hardware, 3 August).

Relief in 2028 needs the new fabs to arrive on time and HBM not to eat them. Micron’s June presentation says technology transitions add fewer bits over time, greenfield fabs are slowed by construction lead times, trade-skill shortages, permitting and energy, and “HBM’s growth and increasing trade ratio with every new generation further pressures non-HBM supply”. It also expects its blended DRAM cost per bit to rise from current levels, which argues against a return to old price lows even if supply catches up.

The counter-case deserves to be stated fairly. Analysts quoted without names by The Register on 7 July warned that the fab build-out could eventually produce an oversupply if AI companies slow their plans, and The Register wrote on 12 July that an AI shortfall would leave memory makers at “the bottom of a bust cycle to end all bust cycles”. Model-efficiency work cuts inference memory: Google claims its TurboQuant method cuts it at least sixfold, although The Register judged on 1 April that it “won’t end the memory crunch”. And TrendForce’s own September bulletins note that price rises are curbing notebook output, that channel restocking “dims 2027 demand outlook”, and that server buyers are “aggressively accumulating inventory” - demand pulled forward now can leave later demand weaker. TrendForce still expects prices to rise. None of the bear cases on record carries a date or a probability, and this article does not invent either.

The forecasts in this cycle have missed low, and the one dip reversed

Before weighing any date, check the forecasters’ record. TrendForce wrote on 22 January 2026 that historical peak quarterly DRAM price rises were “around 35%”, in the same release that put the fourth quarter of 2025 at 53-58%. On its later conventional-DRAM basis, contract prices rose 45-50% in that quarter (TrendForce, 26 February; the January release did not say whether HBM was included) and 93-98% in the first quarter of 2026. TrendForce’s explanation, on 9 February, is that this cycle is led by cloud providers, whose “relatively lower sensitivity to pricing” allowed rises beyond the device-maker-led cycle of 2017-2019. When the researcher Daniel Lemire wrote on 5 August that RAM had become “about as expensive as it was in 2007”, PC Gamer found DDR5 at about $15-16 per gigabyte against roughly $25-30 for DDR2 in 2007: half to two-thirds of the 2007 price, so still cheaper, as PC Gamer concluded.

How far out, on midpoints

  4Q25   outcome 47.5 / first forecast 10.5   = 4.5 times the forecast rise
  1Q26   outcome 95.5 / first forecast 57.5   = 1.66 times
  3Q25 consumer DDR4, revision 87.5 / 42.5    = 2.06 times, within five weeks

  Counterpoint, "+40% through Q2 2026"
    1Q26 actual x 2Q26 forecast = 1.955 x 1.605 = 3.14, i.e. +214%
    (the second factor is TrendForce's 2Q26 forecast; no realised
     2Q26 contract change had been published by late September)

• First forecast
• Second forecast
• Third forecast
• Reported
• Forecast

Sources: TrendForce, 24 Sep 2025 (4Q25: +8-13%); TrendForce, 29 Oct 2025 (4Q25: raised to +18-23%); TrendForce, 26 Nov 2025 (4Q25: +45-50%); TrendForce, 26 Feb 2026 (4Q25 reported: +45-50%); TrendForce, 5 Jan 2026 (1Q26: +55-60%); TrendForce, 2 Feb 2026 (1Q26: raised to +90-95%); TrendForce, 1 Jun 2026 (1Q26 reported: about +93-98%). Figures as of October 2, 2026. TrendForce revised its 4Q25 call twice and its 1Q26 call once; none of its later quarters has a reported figure yet.

Show the figures

TrendForce's first call for 4Q25 and for 1Q26 was far too low, and was raised as the quarter went on (% change in conventional DRAM contract prices on the previous quarter, middle of each range)

Period	First forecast	Second forecast	Third forecast	Reported
4Q25	10.5% (forecast)	20.5% (forecast)	47.5% (forecast)	47.5%
1Q26	57.5% (forecast)	92.5% (forecast)	-	95.5%

The errors run almost entirely one way: quarterly forecasts revised up within the quarter, relief horizons moving outwards, and every dated stabilisation call failing. Two calls are running high so far, the Lexar manager’s doubling and PC Games Hardware’s €550-600, which keeps the table honest, though the year is not over. None of this proves the 2028 models wrong, but in this cycle a relief date has been better read as the earliest it could happen than as a central estimate.

The spring 2026 dip was a pause, and the people who called it that were right

The one real fall came in March 2026. Supply-chain sources in Asia, reported by TechSpot on 10 April, said DDR5 prices dropped nearly 30% in March and DDR4 about 5%, the first monthly decline since February 2025, which DigiTimes put down to distributors selling stockpiles; the DDR5 16Gb chip was down to $37.20 in early April. Tom’s Guide reported on 1 April that TrendForce saw US DDR5 retail prices fall more than 20% in a month, with a Corsair Vengeance 32GB kit down from a $500 high to $369 at Amazon.

Contract prices never paused. TrendForce’s 31 March forecast for the second quarter was +58-63%. By 24 August the cheapest 32GB DDR5-6000 kit in the US had crossed $400 (Tom’s Hardware), ComputerBase’s check on 16 September found German desktop RAM kits 370% dearer than a year earlier and at a new high, and the DDR5 16Gb spot average was $57.667 on 24 September, 55% above April’s low. The analysts who called it correctly at the time were quoted by PCWorld on 20 April: IDC’s Tom Mainelli described “a pause or even some reversal” on over-priced SKUs, “far from normalization”, and HotTech’s Dave Altavilla tied the softness to “localized inventory pressure in distribution channels”. The arithmetic on the one kit shows what a dip was worth: (500 - 369) / 500 = 26%, to a buyer who bought in April and not to one who waited for a second leg down.

Treat any relief date as the earliest it could happen, and a dip as a chance to make a purchase you were going to make anyway, not as the start of a trend.

Eight different numbers are called the price of RAM

In September alone, the price of a 32GB DDR5 kit was quoted anywhere from about $390 (a DDR5-5200 kit with a coupon, Tom’s Hardware, 8 September) to about $604 (Capital and Compute’s tracker average on 25 September), and nobody was wrong. The sources were measuring different things: the cheapest kit of any speed, the best price at one speed, a tracker’s median, an average across all kits, or the modules’ spot price. Each is a real number, and each answers a different question. The table sets those five between the three that bracket them: contract and chip spot upstream, used downstream.

A buyer should use the cheapest and the median for the exact thing they would buy, and treat every forecast in this article as a contract percentage that has to be translated before it means anything at a till. How the layers move, and why, is in the shortage explainer.

A year apart, on one consistent measure, the kits most buyers compare moved like this (PCPartPicker average prices, compiled by TechTimes on 18 August):

These are averages across every listed kit, premium ones included, so they sit above the best prices quoted here ($409 for 32GB of DDR5-6000 and $190 for 32GB of DDR4-3200 on Tom’s Hardware’s index of 14 September). The ratio is the point: DDR5 kits multiplied about twice as fast as DDR4 kits, and DDR4 still costs about half as much per gigabyte, 8.78 / 17.88 = 0.49 on the averages and 5.94 / 12.78 = 0.46 on the best prices.

Two same-day comparisons show which way near-term retail risk points.

14 Sep 2026, same day
  module spot, DDR5 UDIMM 16GB       2 x 235.667 = 471.33 USD    14.73 USD/GB
  retail best price, 32GB DDR5-6000                409.00 USD    12.78 USD/GB
  retail / module spot               409 / 471.33 = 0.868        13% below spot
  retail 16GB kit                    239 / 16     = 14.94 USD/GB  level with spot

24 Sep 2026, chip spot per bit
  DDR4 16Gb 83.784 / DDR5 16Gb 57.667 = 1.45     DDR4 chips 45% dearer

On 14 September the best US price for 32GB of DDR5-6000 was 13% below the cost of two 16GB modules on the spot market, and the spot item is slower DDR5-4800/5600, so if anything the gap is understated. The late-July contract table agrees: a DDR5 8GB laptop module at $16.25 per gigabyte, above the $12.78 a 32GB desktop kit cost at retail in September. Retail is selling memory for less than it costs to replace. TrendForce’s 22 September release on module makers shows how: Adata’s 2025 profits rose partly on low-cost inventory accumulated in the previous cycle. When that stock runs out, retail reprices towards replacement cost; Lexar’s Australia and New Zealand manager put the delay between a cost change and consumer prices at eight to nine months (Tom’s Hardware, 6 June), one executive’s view rather than a measured figure.

The second comparison is the stranger one. On 24 September a DDR4 chip cost 45% more per bit than a DDR5 chip on the spot market, while DDR4 kits sold at retail for less than half DDR5’s price per gigabyte. The same inventory lag explains it, and what it means for anyone thinking of staying on DDR4 is covered in DDR4 as an escape route.

Waiting pays only if you can wait past the turn: the arithmetic

One real purchase runs through this section: a 32GB DDR5-6000 kit at the $409 best US price of 14 September. A later reading on a different tracker ran higher - the cheapest 32GB DDR5 kit of any speed on Capital and Compute’s was $429 on 25 September - so the dollar figures below, which scale with the anchor, may be low. It is a method, not a prediction, and the price you would actually pay today is on the 32GB DDR5 kit page.

1. Has retail followed the contract path?
   contract, 3Q25 level to 3Q26 level (conventional DRAM)
     low    1.45  x 1.93  x 1.58  x 1.13  = 5.00
     mid    1.475 x 1.955 x 1.605 x 1.155 = 5.35
     high   1.50  x 1.98  x 1.63  x 1.18  = 5.71
     (4Q25 and 1Q26 reported; 2Q26 and 3Q26 are forecasts)
   retail, PCPartPicker 2x16GB DDR5-6000, Aug 2025 to Aug 2026
     572 / 108 = 5.30

2. Waiting to the end of 2026, kit at 409 USD
   (already at the 3Q26 contract level on step 1's reading)
   TrendForce 4Q26, 30 Sep     409 x 1.10 = 450    409 x 1.15 = 470    +41 to +61
   retail up to module spot    2 x 235.667 = 471                       +62
   Jefferies-relayed 4Q26      409 x 1.30 = 532    409 x 1.40 = 573    +123 to +164
   a spring-2026-style dip     409 x 0.74 = 303                        about -106

3. Waiting to the end of 2027 or 2028, Bernstein's path, ratios only
   2028 against the 2027 peak         1.05 / 2.23 = 0.471   (-53%)
   2027 peak against 2026, two readings   2.23 / 1.85 = 1.205   or about 1.50
   kit at the end-2027 peak  409 x 1.205 = 493    409 x 1.50 = 614    +84 to +205
   kit at end-2028     409 x 1.205 x 0.471 = 232    409 x 1.50 x 0.471 = 289
   saving              409 - 289 = 120  to  409 - 232 = 177
   months, Oct 2026 to Dec 2028 = 27
   saving per month    120 / 27 = 4.4   to  177 / 27 = 6.6 USD
   as a share of 409   120 / 409 = 29%  to  177 / 409 = 43%
   against the Aug 2025 average   232 / 108 = 2.15    289 / 108 = 2.68 times

Step one is the cross-check that makes the rest usable. Compounded, TrendForce’s quarterly contract figures give a rise of 5.0 to 5.7 times over the year to the third quarter of 2026, two quarters of it still forecasts; PCPartPicker’s average for the same kind of kit rose 5.3 times over the year to August (TechTimes, 18 August). Retail has tracked the contract path, which is what justifies applying contract percentages to a kit price at all. It also means a September price already contains the third quarter’s rise, so only the fourth quarter’s is still ahead.

Step two prices a wait to the end of the year. On TrendForce’s fourth-quarter forecast of 30 September, 10-15% for conventional DRAM, the kit costs about $41 to $61 more, close to the $62 that retail catching up with the spot cost of two modules would add on its own; the single relayed analyst’s fourth-quarter call adds about $123 to $164. Beyond December no public quarterly forecast exists. A saving comes only if a dip like the spring’s recurs and is bought, and none is forecast.

Step three is the only route to a saving, and it needs a long wait. Only Bernstein’s ratios are used, for the unit reason the 2028 section gives: the 2028 price is 47% of the end-2027 peak, and the peak is either 21% or about 50% above the 2026 level depending on which report is right. Scaled onto the kit, that gives about $232 to $289 at the end of 2028, a saving of about $120 to $177 spread over 27 months, or roughly $4.40 to $6.60 a month. As a share of today’s price that is 29% to 43% off, not the 40-60% one retailer’s buying guide states without a source. Three things qualify it. Bernstein’s figure is one supplier’s chip selling price, not a retail price. If the eight-to-nine-month lag the Lexar manager described is right, a chip-level fall during 2028 reaches shelves in 2029. And even then the kit costs about 2.15 to 2.7 times PCPartPicker’s August 2025 average, and more against that month’s best prices, since an average sits above them.

The same numbers settle the case most advice leaves out: a buyer who knows the need is coming, such as a build planned around Nova Lake or desktop Zen 6, and could buy the memory first. No named forecast has memory cheaper than today within twelve months. Bernstein’s ratios put the kit at about $493 to $614 at the end-2027 peak, and Gartner’s gentler path, DRAM prices up 2% in 2027 (as relayed by SHI, 14 August), still has no fall, so for a need within about a year, buying early is the cheaper expectation. Against it: the money is spent early; the return window and the warranty start at purchase, so test the kit when it arrives; and the memory support of Nova Lake and desktop Zen 6 is unpublished, so buy a mainstream kit rather than a fast one tuned for today’s boards. Beyond 2027 the case weakens, because the one falling path turns down during 2028.

A hindsight check points the same way. A Kingston manager advised on 16 December 2025 (The Full Nerd podcast, reported by Tom’s Hardware) not to hold off, because prices would be higher 30 days later. A comparable kit cost more than $420 in early December (Tom’s Hardware, 19 December) and $409 on 14 September, with a 26% dip and a new high in between: nine months of waiting gained and lost almost nothing, and the buyer who did best happened to buy in April.

Wait only if you can go without the memory until the end of 2028 or later, and going without is worth less to you than about five dollars a month. Otherwise buy no later than you need it.

Buy two modules, and DDR5-6000 only when it costs what DDR5-5600 costs

The DDR5-6000 half of the advice has real support. AMD’s own EXPO page describes its lab testing, as of March 2026, on a Ryzen 7 9700X with “32 GB DDR5-6000 AMD EXPO Technology Memory”, tested at CL28, CL30 and CL36. AMD’s Robert Hallock called it “roughly” the Ryzen 7000 sweet spot in August 2022. The mechanism is in TechSpot’s test of 7 April 2025: on AM5 the memory controller’s clock can match the memory clock up to about 3,000 MHz, which is DDR5-6000, and above that it drops to half, so DDR5-8000 runs the controller at 2,000 MHz, 33% slower. TechSpot found DDR5-8000 barely ahead of a tight DDR5-6000 CL26 kit, and only about half of the 21 boards it tried ran DDR5-8000 at all.

What the advice leaves out is that DDR5-6000 is not AMD’s specification. AMD’s product pages for the Ryzen 5 9600X and the Ryzen 7 9800X3D list DDR5-5600 with two modules and DDR5-3600 with four. AMD’s AM5 platform page states that running memory “outside of AMD’s published specifications will void any applicable AMD product warranty, even when enabled via AMD hardware and/or software”, and EXPO is AMD’s own technology. G.Skill began paying a $2.4 million US class-action settlement in August 2026 over speeds reachable only by overclocking, and agreed to print on its packaging that XMP or EXPO is required for the rated speed (Tom’s Hardware, 20 August). None of that makes DDR5-6000 a bad buy. It makes it an overclock, which is worth knowing before paying extra for one.

Intel’s figures conflict with each other. Intel’s specification page for the Core Ultra 9 285K says “Up to DDR5 6400 MT/s”. Intel’s datasheet, as Tom’s Hardware read it in October 2024 and November 2025, rates Arrow Lake at DDR5-5600 with standard modules and DDR5-6400 only with clocked CUDIMMs, and the Arrow Lake Refresh at DDR5-7200 with CUDIMMs at one module per channel; with four modules the rating drops to DDR5-4800 for single-rank and DDR5-4400 for dual-rank modules. TechSpot’s June 2026 buying guide says Arrow Lake CPUs “barely benefit” from memory faster than 6000. What a clocked module is, and why some fast kits quietly become slower ones, is covered in DDR4 vs DDR5.

Then the price test, where the advice holds or fails.

6200 CL36 against 5600 CL40, 17 Sep    (419.99 - 409.99) / 409.99 = +2.4% price
6600 against 6000, 14 Sep              (655 - 409) / 409          = +60% price
6000 against 4800, 24 Aug              (400 - 359) / 359          = +11% price
four modules on AM5                    (5600 - 3600) / 5600       = 36% lower official speed

On 14 September the 5600 and 6000 kits had the same best US price, and on 17 September a faster kit cost 2.4% more, for a gain TechSpot measured at about 5% (6000 CL30 against 5600) on a non-X3D Ryzen. The gain is much smaller on X3D chips: TechSpot’s July test of AMD’s tighter EXPO “Ultra Low Latency” profiles found gains of about 8% to 15% on a 9700X but 1-2% in three of five games on a 9800X3D, and 9% and 11% only in the two most processor-bound. And timings matter less than people pay for: in February TechSpot found a CXMT-based DDR5-6000 CL36 kit essentially matched a premium G.Skill kit in games. Buy DDR5-6000 at CL30 to CL36 when it costs within a few per cent of DDR5-5600; otherwise buy the 5600 kit, and never pay a large premium for anything faster. Buy it as two modules, for the reason the last line of the arithmetic gives.

Two steps turn that rule into a kit that runs at its label. Before buying, check the exact part number against the motherboard maker’s qualified vendor list (QVL) for your processor and number of modules; Newegg’s own 9 July guide listed one of its recommended DDR5-6000 CL30 kits as not compatible with Arrow Lake. After installing, enable EXPO on AMD or XMP on Intel in the BIOS, because DDR5 runs at its standard speed until a profile is on, as G.Skill’s packaging must now say. Today’s prices at each speed are on the DDR5-6000 page and the DDR5-5600 page; what each processor officially supports is in the CPU index; why the label speed and the running speed differ is in the speed guide, and what CL30 and CL36 mean in nanoseconds is in the latency guide.

32GB is the capacity to buy once, because the second purchase is the expensive one

At the 14 September index the 2x16GB kit was the cheapest DDR5 capacity per gigabyte, and both smaller and larger kits cost more.

The optimum moves. In late July Tom’s Hardware’s lowest prices ($209 for 16GB, $379 for 32GB, $699 for 64GB) made 64GB the cheapest per gigabyte, $10.92 against $11.84. Large-kit prices also disagree: 128GB was $2,339 on the 14 September index but $3,399 in Tom’s Hardware’s articles of 17 and 24 August. So read the per-gigabyte sort for the capacities you are choosing between, on the 16GB kit page, the 32GB kit page and the 64GB kit page, rather than a rule of thumb from July.

The buy-once argument comes from the platform. Because four modules drop AM5’s official speed by 36%, capacity is bought as two modules, and a 2x8GB kit bought now is replaced later rather than added to - the point Newegg’s own guide made on 11 September. The detour costs the whole price of the small kit, less whatever it fetches used. At the 14 September prices the step from 16GB to 32GB cost (409 - 239) / 16 = $10.63 per extra gigabyte, less than the 32GB kit’s own average rate.

The same arithmetic decides the next step up. A builder who expects to need 64GB during the machine’s life, for creative work, virtual machines or local AI models, pays less buying it once as two 32GB modules: on 14 September the extra 32GB cost (889 - 409) / 32 = $15.00 per gigabyte, 17% above the 32GB kit’s own rate, against a second full purchase of $889 or more later. The step beyond is dearer: 96GB cost (1,699 - 889) / 32 = $25.31 per extra gigabyte, about twice the 32GB kit’s rate, so 96GB and 128GB are for a measured need, not insurance. Two larger modules are still two modules, which AMD rates at DDR5-5600; it is the second pair that drops the rating to DDR5-3600.

The case for 32GB in a new gaming build is TechSpot’s measurement of 8 January 2026: with an 8GB graphics card, game assets spill into system memory, and going from 16GB to 32GB improved the 1% lows of an 8GB RX 9060 XT by more than 60% in Spider-Man 2. That matters because Valve’s August 2026 survey shows about 69% of Windows respondents with 8GB of video memory or less, while 45.06% of all Windows respondents already have 32GB of system memory or more. The case for 16GB deserves stating too: TechSpot called it “a perfectly viable option” for budget builders in January, and Tom’s Hardware’s best-RAM guide (July) calls it sufficient for gaming. Sizing by workload, home and enterprise, is the whole subject of how much RAM you actually need.

The platform refresh is real, later than advertised, and does not make memory cheaper

“Hold for the late-2026 platform refresh” assumes a refresh arrives in late 2026. For Intel it might; for AMD’s desktops the reports say 2027.

Intel’s chief executive said in the company’s prepared earnings remarks of 22 January 2026 that Nova Lake is “coming at the end of 2026”; the July remarks do not mention it. A slide posted by the leaker @wxnod, which Tom’s Hardware reported on 3 September and believes came from a partner rather than Intel, puts Nova Lake-S mass production in the fourth quarter of 2026, the first 28-core desktop parts in the first quarter of 2027 and the 52-core flagship later in 2027. Tom’s Hardware’s own analysis of 18 September judged a late-2026 launch more likely, with Z990 boards on shelves by October or November. DigiTimes reported, relayed by Tom’s Hardware on 8 September, that Intel will launch “major annual products” in March 2027 after raising CPU prices by about 10%. By the same leaks, Nova Lake needs a new LGA1954 socket and 900-series chipsets and keeps a dual-channel DDR5 controller, reportedly up to DDR5-8000; whether that speed needs clocked modules is not stated anywhere traced. An earlier roadmap leak, reported by VideoCardz and relayed by TweakTown on 17 July, staggers the family through 2027: 28-core K models in March and April, 16-core and 8-core models from late March to May, and 52-core parts between May and September. For a mainstream Nova Lake processor, the leaks point to spring 2027, not late 2026.

AMD launched Zen 6 in servers first, with EPYC “Venice” in July. For desktops, Benchlife reported in February that the Ryzen 10000 “Olympic Ridge” family would arrive in 2027, Tom’s Hardware called 2027 “very likely” in June, and DigiTimes reported AMD launches in June-July 2027. AMD has announced no desktop date. What AMD has said, in its own Computex blog of 31 May 2026, is that AM5 support now runs through 2029, and screenshots of AMD’s documentation site, reported by TechSpot on 27 August, list the Zen 6 desktop parts under AM5. Whether a given board needs a BIOS update for them is not in any source traced.

Both next desktop platforms are reported to stay on DDR5, so memory is the one component that gains nothing from waiting for them. A DDR5 kit bought for an AM5 board today carries into a Zen 6 processor if the reports are right, and one bought for an LGA1851 board carries into Nova Lake with a new board. New processors, meanwhile, arrive dear. Intel raised the recommended price of the Core Ultra 7 270K Plus from $289-299 to $339-349 within four months of its March launch (Tom’s Hardware, 2 July), about 17%. If DigiTimes’ 10% rise lands, it adds about $45 to a $449 processor, an illustration rather than a forecast. The one part of the refresh story that matters for DDR4 owners - Intel keeping LGA1700 alive with DDR4 boards - belongs to the DDR4 guide.

If what you own works, wait, and give the wait an end date

The case for holding a working machine is strong. TechSpot wrote on 15 July: “we strongly recommend that you do not purchase DDR5 memory right now, or possibly any DRAM, if you can avoid it”. Its June buying guide advised owners of capable AM4 or LGA1700 systems to make “a targeted CPU or graphics card upgrade” rather than replace the board and memory. Gartner expected on 26 February that PC lifetimes would lengthen by 15% for businesses and 20% for consumers by the end of 2026.

But a wait needs an end date, or it becomes indecision. Three triggers are worth setting now:

1. The launch of Nova Lake or desktop Zen 6, with the memory each supports stated by the vendor rather than a leak, and independent reviews of it.
2. A TrendForce quarterly DRAM contract forecast that is flat or negative, which none has been since its forecast for the second quarter of 2025.
3. The median price per gigabyte on the 32GB DDR5 kit page falling for several consecutive weeks below a figure you write down today.

While waiting, check whether the machine is short of memory at all, and fill an empty channel before replacing modules that work; how much RAM you need covers measuring it. A spare DDR4 kit is worth more than it was a year ago (the eight beliefs give the figures), and what that means for selling or reusing it is in the DDR4 guide.

Six buyers, and the arithmetic that decides each one

Each case gives the evidence that decides it, the dated prices behind the arithmetic, and the one thing that would change the answer.

A first PC, needed now. Buy now. In TechSpot’s June 2026 buying guide, 16GB of DDR5-6000 at $245 was 245 / 786 = 31% of the $786 entry build and 245 / 1,457 = 17% of the $1,457 mid-range build. Take 2x8GB only if the budget forces it, and price the replacement in: at 14 September prices that is $239 now plus a 32GB kit later, against $409 once. What would change the answer: a prebuilt or bundle that already contains what you would have bought.

A gamer on a working AM4 or LGA1700 machine with DDR4. Hold. If the processor is the bottleneck, Tom’s Hardware’s June re-review judged AMD’s $349 Ryzen 7 5800X3D sensible only for someone who already owns an AM4 board and DDR4, and “about $70 to $100 too expensive” otherwise. Its re-review worked out that a Ryzen 5 7600X3D and 16GB of DDR5-6000 cost only about $80 more than the 5800X3D alone for someone who already owns DDR4, a sum that leaves out the AM5 motherboard the move also needs (Tom’s Hardware, 25 June). What would change the answer: a machine that genuinely cannot run what you need. The DDR4 economics in full are in the DDR4 guide.

A gamer with 16GB, an 8GB graphics card, and stutter. Buy the capacity now. TechSpot measured more than 60% better 1% lows at 32GB with an 8GB card in Spider-Man 2. At 14 September prices the 32GB kit that replaces the 16GB one costs $409, less whatever the old kit sells for (the used-market section shows how to price it), and waiting to the end of the year adds about $41 to $61 on TrendForce’s fourth-quarter forecast, about what retail catching up with spot would add, while the stutter continues. What would change the answer: finding that the machine is not actually short of memory, which how much RAM you need shows how to measure.

A laptop buyer. Buy when you need it, and buy the memory in the configuration, because most thin laptops cannot be upgraded later. TrendForce calculated on 4 September that a $900 notebook of early 2025 would need to cost about 80% more, roughly $1,620, to keep its old margin, and that the cheaper component stock holding prices down is running out. If the need falls within the next few months, buying before that repricing is the cheaper expectation: on the 5-20% rise Acer’s chairman expects by the end of the year, a $1,000 laptop becomes 1,000 x 1.05 = $1,050 to 1,000 x 1.20 = $1,200. The price of extra memory at purchase is steep: on Apple’s M6 Mac mini, on sale from 22 September, 8GB more costs $200 (Tom’s Hardware, 25 August), which is $25 per gigabyte, about twice the $12.78 a desktop kit cost per gigabyte on 14 September. What would change the answer: a laptop with socketed memory, covered in the upgradeability guide.

An office with 20 desktops due for refresh. Extend where the fleet is supportable, and buy the rest under a locked quote. Gartner’s expected 15% longer business PC life, applied to an illustrative four-year cycle, is 4 x 1.15 = 4.6 years, about seven months more. The memory in 20 new machines with 16GB each, at the 14 September best price for a 16GB kit, is 20 x 239 = $4,780, which bounds what deferring the memory itself could save. What would change the answer: machines falling out of security support, a cost Gartner itself flags.

That cost has a published price. Windows 10 support ended on 14 October 2025, and Microsoft sells organisations Extended Security Updates at $61 per device for the first year, doubling each year for at most three years, and a device enrolled late pays for the years it missed (Microsoft Learn, updated 17 November 2025). Two years for a desktop that cannot run Windows 11 cost 61 + 122 = $183, three years $427; for 20 machines, $3,660 and $8,540. So “supportable” means machines that can move to Windows 11, plus those whose update bill is smaller than what deferring their replacement saves. Tiers of memory by role are in the capacity guide, and quote terms are in the enterprise section below.

A homelab or small server that needs 256GB or more. The memory generation decides more than the timing.

Server memory per GB, September 2026
  DDR4 RDIMM   5.56   eRacks median of lowest street prices (seller's tracking)
  DDR4 ECC     9.34   DatacenterDisk median (small sample of new listings)
  DDR5 RDIMM  31.05   eRacks median
  DDR5 RDIMM  41.09   DatacenterDisk median
  DDR5 RDIMM  64.06   32GB module spot, 2,050 / 32 (DRAMeXchange, 14 Sep)

256GB of it
  DDR4    256 x 5.56  =  1,423     256 x 9.34  =  2,391
  DDR5    256 x 31.05 =  7,949     256 x 64.06 = 16,399

Measured like for like by each tracker, DDR5 server memory costs 31.05 / 5.56 = 5.6 times (eRacks) and 41.09 / 9.34 = 4.4 times (DatacenterDisk) as much per gigabyte as DDR4, which outweighs anything the timing can do. What would change the answer: a workload that needs a current platform’s bandwidth or core count. Server-memory prices by generation are on the DDR4 registered page and the DDR5 registered page; how many modules a given board takes is in the server capacity guide, and the DDR4 route’s risks are in the DDR4 guide.

Prebuilts, bundles and laptops undercut parts until the old stock runs out

Deals from late August and September show whole systems selling for less than their parts.

CyberPowerPC, Ryzen 7 9800X3D, RTX 5080 FE, 32GB DDR5, 1TB SSD    2,299
  graphics card 1,579 + processor 449                             2,028
  left for memory, SSD, board, power supply and case                271

HP Omen 35L at Walmart against HP's own site
  (3,639.99 - 1,899.99) / 3,639.99 = 48% less

Newegg 9800X3D + X870E board + 32GB DDR5-6000 combo
  memory effectively 489.99 - 248.99 = 241
  against the best standalone 32GB price, 14 Sep  (409 - 241) / 409 = 41% less

Alienware Aurora, July 2025 to July 2026   1,999 / 1,099 = 1.82

The first three items are Tom’s Hardware deal reports of 21 September, 25 September and 20 August; there is, as Tom’s Hardware put it, “no way you will find 32GB of DDR5 memory and a 1TB PCIe 4.0 SSD for $271”.

The mechanism is procurement: system makers buy memory on contract, and some of what they are selling was built from components bought before the steepest rises. TrendForce said on 4 September that this cost buffer “will diminish” as the older inventory is used up, and the repricing has started. The Alienware Aurora that Tom’s Hardware recommended at $1,099 in July 2025 cost $1,999 a year later. Apple raised Mac and iPad prices on 25 June, taking the 13-inch M5 MacBook Air from $1,099 to $1,299. Schenker announced rises of €100-200 on most of its gaming laptops (Tom’s Hardware, 25 September), saying DDR5 laptop modules cost six times what they did in July 2025 and RAM will be short “through 2027 and beyond”.

Bundles work the same way for memory, but mostly on AMD now: Tom’s Hardware’s bundle page noted on 2 September that Intel deals including memory “have become extremely rare”. Two checks apply before buying any of these. Some prebuilts ship a single memory module, which Tom’s Hardware’s July testing found costs 8-10% of gaming performance on most processors, and IDC expected in February that makers would cut base memory configurations. And most thin laptops’ memory is soldered, so the configuration bought is the configuration kept. Buy the prebuilt or bundle while the old stock lasts, if it contains what you would have bought anyway. How much to buy up front is in the capacity guide; soldered, SODIMM and LPCAMM2 laptops are compared in the upgradeability guide; socketed laptop memory is listed on the DDR5 laptop page.

Nine ways to pay less, and two that only look cheaper

1. Buy the speed the price allows. DDR5-6000 only at or within a few per cent of the 5600 price, as the speed section sets out.
2. Accept CL36. The CXMT-based CL36 kit in TechSpot’s February test matched premium memory in games.
3. Skip the premium brand. Newegg’s 9 July guide priced a Patriot Viper Venom 32GB DDR5-6000 CL30 kit at $399.99, and the store card on the same page showed a G.Skill Trident Z5 Neo 32GB DDR5-6000 CL30 kit at $649.99 when checked on 26 September. The two prices are eleven weeks apart, so the ratio, 649.99 / 399.99 = 1.63, is indicative only; compare brands at the same capacity, speed and latency on the per-gigabyte sort before paying for a name.
4. Buy a bundle, if you need the processor and board anyway. The Newegg combo above priced the memory at $241.
5. Buy a prebuilt, on the terms of the previous section.
6. Buy one DDR5 module now and a matched kit later. Tom’s Hardware measured on 31 July that one 16GB DDR5 module loses 8-10% of gaming performance against 2x16GB on most processors (10.9% on a Ryzen 5 7600X, 2.8% on a 9800X3D) and still beats dual-channel DDR4. Sticks bought separately are not validated together, so it suggests buying a matched kit later and selling the single module; TechSpot’s refusal of a single 8GB stick in its June guide is consistent with a 16GB test. Single modules are filtered on the 16GB module page.
7. Buy used, with buyer protection, against the page’s median. PC Games Hardware noted in May that buyers who test used modules with MemTest86 often get much cheaper kits. The tests and the scams are in the used-RAM guide, and a specific part is easiest to follow on the part-number index. Crucial kits still on sale are old stock: Micron stopped shipping Crucial products to the consumer channel at the end of February 2026 and said it would keep supporting their warranties (Micron, 3 December 2025).
8. Keep the machine and upgrade the processor or graphics card instead, as the hold section sets out.
9. Reuse memory you own. Custom builder Maingear started accepting new, used or pulled DDR5 kits in December 2025, after its own cost for a 32GB kit rose 394%. Mixing rules are in how to check what RAM fits.

Two offers only look cheaper. V-Color’s “1+1” packs pair one real module with a non-functional lit filler; the 16GB DDR5-6000 CL30 version was £210, about $285, in the UK on 4 September (Tom’s Hardware): 285 / 16 = $17.81 per real gigabyte against $12.78 for a 32GB kit. And HUDIMM, a half-width DDR5 module that ASRock, Intel and TeamGroup introduced in April 2026, uses one 32-bit sub-channel instead of two, halving both bandwidth and capacity; it is a saving only if its price per gigabyte is lower, which a per-gigabyte sort shows directly.

Two things are not tactics at all. Sales events have not beaten the shortage so far: on Black Friday 2025 Engadget found “there aren’t any promotions worth even considering” (28 November 2025), with a 32GB Crucial bundle at $322 on Amazon that had cost about $175 a month earlier, and Prime Day 2026 (23-26 June) fell between a $374.97 US floor for 32GB of DDR5 in early June (TechTimes) and $379 in late July (Tom’s Hardware, 28 July). Nothing on the record suggests that waiting for Black Friday 2026 beats buying when you need it; a bundle that happens to be on sale is the likelier saving. And buying more than you will install, to resell later, has to beat selling fees, postage and a market whose turns this cycle’s forecasters have not called: the forecasts support rising prices into 2027 and say nothing about when to sell.

Enterprise buyers are bargaining over quote windows, not prices

TrendForce’s 9 July release expects server DRAM contract prices to rise 13-18% in the third quarter and to keep rising every quarter from the second half of 2026 through the second half of 2027, at a moderating pace, with total RDIMM bit supply growing only 15-20% in 2027. From the third quarter, it says, the increases fall mainly on “customers without LTAs” - the long-term agreements the largest buyers have signed. Its 25 August release estimates that server DRAM contracts rose a cumulative 64% in the second half of 2025 and will rise about 270% more across 2026.

The per-gigabyte measures in the homelab case differ by a factor of two or more. eRacks’ $31.05 is up 27% since March (a server seller’s own figures, via PR.com, 22 September). DatacenterDisk’s $41.09 rests on 16 listed modules (26 September). DRAMeXchange’s $64.06 module spot (14 September) is a thin market that rose 7.9% that week. And a 256GB DDR5-6400 RDIMM that Tom’s Hardware reported retailing at about $19,000 on 16 September works out at about $74.

Vendors’ terms moved faster than their prices. The Register reported that HPE cut quote validity from 30 to 14 days in February, then in March told customers that DRAM and NAND “now make up over half of the bill of material cost of a traditional server” and changed its terms to allow repricing between quote and shipment; Cisco reserved the same right. HPE restored 30-day quotes in June, and on 5 August it began honouring quoted prices on orders up to $1 million until the kit leaves its factories. The analyst firm Context showed how slowly finished systems follow: in European distribution, memory modules rose an average of 63% between September and December 2025 and OEM server memory about 28%, while whole servers rose only 5-10% in December (The Register, 14 January 2026).

Large buyers have two tools a small one lacks. They sign long-term agreements - SK hynix has about ten, Micron sixteen, the largest of them with price ceilings and floors - and they redesign: TrendForce reported on 9 July that cloud providers and server makers have been moving some systems from 96GB and 128GB modules to 32GB and 64GB ones since the first half of 2026. A smaller buyer’s options are to lock a quote, right-size configurations, defer refreshes where the fleet is supportable, reuse older memory - Meta now attaches recovered DDR4 to new DDR5 servers through its own CXL chip, as the DDR4 guide describes - or rent: Nutanix’s chief executive said the fastest way to get a new server was to rent one from a hyperscaler (The Register, 12 August).

Two more routes are open to a buyer who cannot sign with a memory maker. SHI, a large US IT reseller, advised on 14 August that organisations with the budget and the forecasts “may opt to order 6 to 12 months ahead of required deployment dates”, and suggested long-term agreements with memory distributors to secure allocations, adding that enterprises “shouldn’t assume the market will return to pre-shortage conditions”. The lever an enterprise buyer actually has in 2026 is the quote window, not the price.

A worked server shows what the per-gigabyte measure and the module size do to the bill.

AMD's published 2P EPYC 9965 reference configuration
  2 sockets x 12 channels x 1 DIMM per channel = 24 DIMMs
  24 x 64GB = 1,536GB

    at 31.05 USD/GB (eRacks median)          47,693 USD
    at 41.09 USD/GB (DatacenterDisk median)  63,114 USD
    at 64.06 USD/GB (32GB RDIMM spot)        98,396 USD

  24 x 32GB = 768GB: half of each, at the same USD/GB; every channel still filled

  illustrative rises on the 63,114 case
    +10% = 6,311     +20% = 12,623     +30% = 18,934

The spot figure comes from a 32GB module and is applied to 64GB modules as an assumption, and TrendForce has published no public quarterly server percentage beyond the third quarter, so the last line is an illustration, not a forecast. Stepping down from 64GB to 32GB modules keeps all twelve channels filled, and so the bandwidth, while halving the capacity and, at the same price per gigabyte, the bill; whether the workload fits is the question the capacity guide answers, and why a server will not always take its maximum is in the server capacity guide. Module types are explained in registered against unbuffered, and memory for specific server models is in the compatibility index, with current DDR5 server listings on the DDR5 server page.

The turn will show up in a known order, on known dates

If prices turn, the signals should arrive in roughly this order, from the market that moves first to the one that moves last.

The calendar is partly known. Micron reported its fiscal fourth quarter on 30 September 2026, as announced on 26 August. TrendForce’s quarterly DRAM contract releases came on 24 September 2025, 5 January, 31 March, 3 July and 30 September 2026, so a first-quarter 2027 release around the turn of the year would fit the pattern - an inference, not an announcement. Samsung gave preliminary second-quarter guidance on 7 July and full results on 30 July, SK hynix on 29 July, so early October for Samsung’s guidance and late October for both companies’ results are likewise inferred.

What is not yet known, or not yet worked in, and would change this article: what Micron’s fourth-quarter results of 30 September say about its outlook; the launch dates and official memory support of Nova Lake and desktop Zen 6; whether CXMT’s memory is cleared for US buyers; and the realised change in second-quarter 2026 contract prices, which TrendForce has not published. One US-specific unknown belongs on the list: tariffs. The Section 232 proclamation of 14 January 2026 put a 25% duty on a narrow set of advanced computing chips and exempted imports for non-data-centre consumer uses in the US (EY’s summary, 9 February), so memory kits and consumer PCs were not its target; it leaves room for what EY calls “significant broader tariffs” later, undated in any source traced, which is a reason to watch the news, not to buy early.

Used memory is priced by the same shortage, and it is the market you can read directly

No third-party index of used memory prices was found, so the used market is the one a buyer reads directly, on the pages this site indexes.

How to read it here. Every listing page ranks by price per gigabyte, the item price plus shipping over every gigabyte in the listing, and states the cheapest and the median for the view. Bid-only auction rows are left out by default, because a bid is not a price anyone can pay today; ?bids=show brings them back and ?auctions=no removes auctions entirely. Each page takes one filter: a 2x16GB kit is on the 32GB DDR5 kit page, which holds kits and sets totalling 32GB, while the 16GB module page filters on the size of each module, and speed is read off the rows (the DDR5-6000 view is its own page). Part-number pages under the part index follow one part across sellers, and condition - new, used, refurbished, for parts - is on every row. The all-generations page puts DDR5 and DDR4 kits side by side on the same measure, and the methodology page says where the figures stop being reliable.

Track the median, not the cheapest row: note it on a date and compare it week by week, and compare it with the shop prices quoted in this article only with care, because this site’s figure includes shipping. The cheapest row on any page is one listing and can be an outlier; the median moves when the market does. An all-time-low alert is only as good as its history: RamRadar flags a kit within 10% of the lowest price it has itself recorded for that kit, while Tom’s Hardware’s index lists an undated lowest-ever US price of $72 for 32GB of DDR5-6000, and its article of 24 August put one such kit’s low at about $80 “in years gone by”. Against $409 on 14 September that is a factor of 5.1 to 5.7, so a rule anchored to a pre-shortage low is a wait for the pre-shortage market.

The same pages price a kit you would sell. If an upgrade retires a 2x8GB set, open the 16GB DDR5 kit page (or the DDR4 one) and read the used rows near the median, not the cheapest, as a guide to what it may fetch before fees and postage. Selling it the month you buy keeps both sides of the trade in the same market; holding it as a spare is a bet on where used prices go next.

Evidence on used prices themselves is thin and mixed. Refurbished PC sales through distribution rose 7% in the fourth quarter of 2025 across five European markets (Context data, reported by The Register on 16 February), which shows demand moving to second-hand machines rather than what they cost; eRacks’ tracking of the lowest street price per gigabyte, which does not separate new from used, shows DDR4 falling while DDR5 rose between March and September; and Meta’s reuse of retired DDR4 in its own servers suggests, though nobody has measured it, that fewer ex-datacentre modules reach the second-hand market. Fakes have followed the money: counterfeit DDR5 laptop modules with dummy chips circulated on Japanese second-hand marketplaces in May 2026 (Tom’s Hardware, 11 May). How to test what arrives, and what to do when it fails, is in the used-RAM guide; where used modules come from, and the ex-server, ex-mining and overclocked histories behind them, is in the DDR4 and provenance guide.

Eight beliefs about RAM prices that the evidence does not support

“The new fabs will fix it in 2027.” A cleanroom date is not output: TrendForce expects substantial output only in 2028, and the most quoted “December 2028” date belongs to SK hynix’s M17, a NAND fab.

“China will flood the market.” Not on current evidence: CXMT’s new output will mainly serve China, the company is on the 1260H list, and Lexar’s kit on Chinese DRAM was priced like US kits, as the relief section sets out. The strongest counterpoint is real: CXMT announced on 20 September that its G5 platform is in mass production with at least 50% more dies per wafer than its previous generation (Reuters, via TrendForce), and Acer expects Chinese capacity to help by late 2027; Adata expects it to make no material difference to 2027.

“The spring dip means the top is in.” Spot DDR5 fell nearly 30% in March; by late August US retail prices had set a new high, and by mid-September German ones had too.

“DDR4 is a safe haven.” US DDR4 kit averages rose 120% to 177% between August 2025 and August 2026 (Tom’s Hardware, 17 August), and DDR4 chips cost 45% more per bit than DDR5 chips on 24 September. The DDR4 guide covers when it is still the right buy.

“Analysts expect another 40-50%.” One unnamed analyst said so, relayed three times; TrendForce’s figure for the same quarter was 13-18%.

“Gartner says DRAM prices will rise 130%.” The 130% in Gartner’s release of 26 February is DRAM and SSD prices combined. Gartner’s DRAM-only figures for 2026 have since been raised: 80% (with NAND at 202%) is attributed to a February research note, “Emerging Issue: Managing ‘Memflation’ Through 2027”, behind Gartner’s paywall and not checked here; its forecast of 8 April put DRAM at 125% and said “any meaningful price relief is not expected until late 2027” (as reported by Evertiq, 13 April); and August relays gave 199% (SHI, 14 August) and 271% (The Register, 6 August), which disagree. The late-2027 relief date that guides attribute to Gartner is Gartner’s own, and it marks when relief might begin, not a return to old prices.

“Prices will go back to 2024 levels.” No model traced says so. IDC’s February model “does not point to a reversion to 2025 pricing levels within the forecast horizon”; Micron says the floors in its five-year agreements enable a gross margin “well above our peak quarterly margins in any past cycle”, and expects its cost per bit to rise; Bernstein’s 2028 price is still about twice 2025’s; Lenovo calls it a “new normal”.

“It is price-fixing.” A US class action filed in June 2026 in the Northern District of California alleges that Samsung, SK hynix and Micron deliberately restricted commodity DRAM supply since 2022. Those are allegations and have not been tested in court. The precedent people cite is real but specific: Infineon pleaded guilty in October 2004, Hynix in May 2005, and Samsung agreed to plead guilty in October 2005 to a DRAM price-fixing conspiracy that, in Samsung’s case, ran from April 1999 to June 2002 (US Department of Justice, 13 October 2005); Micron was not among them. The history is in the shortage explainer.

The short answers, with where each is argued

What to do with this, in order

1. Decide your horizon first: need it now, within a year, before the end of 2028, or only after it. If the need is within a year and you can test the kit on arrival, buying now is the cheaper expectation.
2. If it is late 2028 or after and going without costs you little, wait, and set the three triggers from the hold section with a date to review them.
3. Otherwise confirm the capacity with the capacity guide and what your board takes with the CPU index and how to check what RAM fits.
4. Buy two modules, DDR5-6000 at CL30 to CL36 only when it costs within a few per cent of DDR5-5600, check the part number against the board’s QVL, and enable EXPO or XMP once it is installed.
5. Price a prebuilt or bundle before the parts, and check it does not ship a single module.
6. Read the cheapest and the median on the 32GB DDR5 kit page, with bid-only rows left out, and compare the used rows with the new ones on the same page.
7. If a dip comes and you were going to buy anyway, buy. Do not wait for a second leg down.
8. For servers and fleets, lock the quote, right-size the modules, and defer only what is supportable.
9. Re-read this later in October, once Micron’s results of 30 September are worked in, as TrendForce’s fourth-quarter forecast of the same day already is.

Everything above is as of late September 2026, with TrendForce’s fourth-quarter forecast of 30 September added on 2 October; the fact most likely to change it next is Micron’s results of the same day, not yet worked in. This cycle’s forecasts have mostly run low and its relief dates have mostly moved out, so the article will be revised as new figures land rather than defended. Until they do, the short version stands: if you need memory, buy it no later than you need it, as two modules, at the speed that costs no more than the slower one; if you do not need it, the wait that pays is a long one, and it pays about five dollars a month.