DDR4 as an escape route, used memory, and upgradeability

Partly, and each of the three works for a different reason and fails in a different way. For a machine you already own, DDR4 is still the cheapest mainstream memory you can buy, at roughly half DDR5’s price per gigabyte at US retail or less. As the basis of a new build it is an escape route with a toll and an end date: DDR4 kits cost 2.2 to 2.8 times what they did a year earlier, its chips now cost more than DDR5’s, and on the one processor family that takes both generations it gives up 9% to 17% in CPU-bound games. Used memory is safe when its origin is read rather than assumed. And a laptop’s upgrade slot is worth exactly what the module that fits it costs, which in 2026 is sometimes more than buying the capacity up front.

The popular advice splits into two camps that never meet. One says DDR4 is a dead end, from reviews that measure frame rates and leave out the price of the memory. The other says DDR4 saves a couple of hundred dollars, from build guides that price the parts and leave out the frame rates. Neither joins the two into one number, says how long the route stays open, or says where the cheapest second-hand modules come from. The laptop half is answered almost everywhere with a form-factor explainer and no prices, which are what decides whether an upgrade slot is worth paying for.

This article covers all three in order: the price of DDR4 against DDR5 at four layers of the market and why they disagree; who still makes DDR4; the two platforms that came back and what DDR4 costs on them, in frame rate and in dollars per percentage point, for a new build and for an owner; then adding capacity, the Windows 10 deadline, laptops and servers. Then where used modules come from, what “ex-mining”, “overclocked” and “new” can mean, and the warranties that do not transfer. Then three kinds of laptop memory, what LPCAMM2 has cost, and what an upgrade slot is worth in dollars. It is long because each of these topics is usually written up alone, and the decision needs all three at once.

A note on the numbers, because it changes how you should read them. Where a figure is a vendor’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 you can check it and change the inputs. Where an input is an assumption, it is labelled as one. Prices on this site move daily, so where the evidence is this site’s own catalogue the article names the page to read rather than printing a number from it. Everything below describes the market as of late September 2026.

Four neighbouring questions have their own articles and are linked rather than repeated here: why the shortage happened is in the 2026 DRAM shortage, when to buy is in buy RAM now or wait, how much to buy is in how much RAM you need, and how to test and read a used module is in buying used RAM on eBay.

DDR4 more than doubled in a year, and it is still the cheaper gigabyte

Tom’s Hardware’s US price index recorded the slide. Its tracking page, last updated on 14 September 2026, says a 32GB (2x16GB) DDR4 kit commonly sold for $60 to $90 in October 2025 and for $150 to $180 by January 2026, and draws the conclusion that gives this article its caution: “DDR4 isn’t exactly a safe haven for consumers.” Other snapshots agree: PC Gamer found a Crucial 32GB DDR4-3200 kit at $139 in late October 2025, and Resell Calendar, reading PCPartPicker’s charts, put a typical 32GB DDR4-3600 kit at about $90 in May 2025 and $240 in December.

The cleanest year-on-year series is the PCPartPicker averages Tom’s Hardware compiled on 17 August 2026, which the author calls “somewhat approximate” but which cover the same kits a year apart. Per-gigabyte figures and changes are computed here.

Every DDR4 kit in Tom’s Hardware’s comparison rose by between 120% and 177%. A 2019-era Corsair Vengeance DDR4-3200 32GB kit cost $48 a year earlier, and by July 2026, Tom’s Hardware wrote, you would be “lucky to find it on sale for $200”. That is the toll. But DDR5 rose by more, between 355% and 485% across the kits in the same comparison, and that is why the ratio moved in DDR4’s favour while both got dearer:

2x16GB, Aug 2025   DDR4-3200 105 / 32 = $3.28/GB   DDR5-6000 108 / 32 = $3.38/GB   ratio 0.97
2x16GB, Aug 2026   DDR4-3200 281 / 32 = $8.78/GB   DDR5-6000 572 / 32 = $17.88/GB  ratio 0.49
best 32GB kit,     DDR4-3200 190 / 32 = $5.94/GB   DDR5-6000 409 / 32 = $12.78/GB  ratio 0.46
  Tom's index, 2026-09-14

A year ago a gigabyte of DDR4 and a gigabyte of DDR5 cost the same at US retail. By September 2026 the DDR4 gigabyte cost about half. An average across listings and a best price differ in level but give nearly the same ratio, and the other dated sources put it in the same region. XDA Developers priced 32GB of DDR4-3200 at $179 against $402 for DDR5-6000 on 26 August (0.45); the Capital and Compute tracker’s 26 September snapshot put 32GB DDR4 desktop kits at $129 to $233 against $429 to $560 for DDR5 (0.30 low against low, 0.42 high against high); and eRacks, a US server builder publishing its own tracking on 22 September, put the median DDR4 desktop gigabyte at $6.22 against $14.58 for DDR5 (0.43). Five sources, five methods, and a DDR4 gigabyte at between about a third and a half of a DDR5 one.

One September guide from a price-tracking site says the opposite, that DDR4 “can cost the same or more” per gigabyte “depending on the week and the kit”. That is true at the chip layer, the subject of the next section; at the retail kit layer, where people buy, no dated source found for this article supports it.

To read the live ratio yourself, divide the medians printed at the top of 32GB DDR4 kits and 32GB DDR5 kits, never the two cheapest rows; DDR4 vs DDR5 explains why, and the all-generations table sorts both generations on one scale.

The chip now costs more than DDR5’s, and the kit has not caught up yet

As silicon, DDR4 is now dearer per bit than DDR5. As a retail kit, it is still about half the price per gigabyte. Both are true on the same day, and the gap between them is the most useful single thing to understand about whether the escape route lasts.

The chip layer first. DRAMeXchange, TrendForce’s price board, published a session average of $83.784 for a DDR4 16Gb 3200 chip on 24 September 2026, against $57.667 for a DDR5 16Gb chip. A 16-gigabit chip holds 2GB, so per gigabyte of silicon:

DDR4 16Gb 3200       83.784 / 2 = $41.89 per GB of silicon
DDR5 16Gb 4800/5600  57.667 / 2 = $28.83 per GB of silicon
DDR4 as a share of DDR5          41.89 / 28.83 = 1.45

DDR4 silicon cost 45% more per bit than DDR5’s. Read it with its caveat: the same DDR4 line traded between $45.50 and $119.00 that day, a spread that points to a thin market, and TrendForce’s 16 September spot report described branded 16Gb DDR4 chips as seeing “more pronounced price corrections” as trading cooled. Erratic as it is, it is where new supply changes hands, and its series is steep: DDR4-3200 8Gb chips cost $1.75 at the end of April 2025, per ComputerBase data relayed by Tom’s Hardware, and $45.54 on 15 September 2026 by TrendForce’s weekly figure, 45.54 / 1.75 = 26 times.

The same gigabyte at each layer of the market:

The module row is $164.30 for a DDR4 16GB UDIMM and $235.67 for a DDR5 one, each divided by 16. Two things in it should stop you. The first is that a DDR4 16GB module traded at $164.30 while the eight 16Gb chips such a module needs would have cost 8 x $83.78, about $670, at branded chip spot: the modules on the trade board were not built from chips bought at September’s branded spot. The second is the last row read against the module row. On 14 September the best US retail DDR4 kit sold at 5.94 / 10.27 = 58% of the module trade price per gigabyte, while the best DDR5 kit sold at 12.78 / 14.73 = 87% of its own. Retail DDR4 is sitting much further below replacement cost than retail DDR5 is.

Two explanations fit the evidence, and one fact shows the pressure is not new. The first explanation is inventory: TrendForce’s 22 September report on the module market attributed ADATA’s 2025 profit growth partly to “low-cost inventory” bought in the previous cycle and sold as prices rose. The second is that retail reprices slowly, for the reasons set out in the 2026 DRAM shortage. The fact is that the inversion is old. The contract price of 8GB PC DDR4 modules overtook DDR5’s in July 2025, a rare “price inversion” in TrendForce’s words, and TrendForce’s forecast for third-quarter consumer DDR4 contract prices doubled in five weeks, from 40-45% on 7 July to 85-90% on 11 August. Micron’s Sumit Sadana had told DigiTimes in June 2025 that DDR4 could become dearer than DDR5; SK hynix called the spike “a short-term demand surge driven by supply concerns” on 24 July 2025. Fourteen months later the chip-level inversion still held.

The practical reading is a mechanism, not a forecast. When the cheap inventory runs out, retail DDR4 reprices toward what a new module costs to make, and on the September figures that is well above where retail DDR4 sits now. Nobody can say when, and this article does not guess. “DDR4 is the cheap option” is a statement about stock on shelves, not about the technology.

Two errors in 2025-2026 reporting are worth correcting once. Spot tables list chips as “16Gb (2Gx8)”, sixteen gigabits, and several articles printed that as “a 16GB DDR4 chip” or “16GB DDR4 modules hit a record of $60”, confusing a 2GB chip with a 16GB module that holds eight of them. And the 58-63% one April article called a DDR5 price forecast was TrendForce’s second-quarter forecast for conventional DRAM contract prices (Tom’s Hardware, 10 April 2026).

Nobody is building the DDR4 you would want to buy

An escape route is only as good as the supply behind it, and the DDR4 still being made is less of it, for other customers, on old processes.

Micron’s position is the clearest because it is on the record. On its 25 June 2025 earnings call chief executive Sanjay Mehrotra said Micron had sent end-of-life notices for DDR4 and LPDDR4 to customers in the high-volume segments “several months ago with final shipments occurring in two to three quarters from now”, and that it was already on allocation for those products. Micron has since started 1-alpha DDR4 at its Manassas, Virginia fab, and its June 2026 earnings deck says what for: “the legacy product needs of our customers in auto, industrial, medical, aerospace and defense markets.” TrendForce’s reading on 26 May 2026 was that the move “does not indicate a renewed focus on supplying DDR4 components for consumer electronics”, and it expects DDR4 to be around 7% of Micron’s DRAM output in 2026, a forecast.

Samsung and SK hynix are harder, because the reports conflict and neither company published dates of its own that this article could find.

That table corrects two lines still quoted in 2026. “No new consumer DDR4 is being manufactured” is contradicted by Winbond, which expanded its DDR4 shipments, and by Nanya, which is putting more capacity into DDR4 than into DDR5; TrendForce reported on 1 June 2026 that Winbond’s DDR4 and LPDDR4 shipments lifted its first-quarter DRAM revenue 91.4% quarter on quarter, to nearly $568 million. And the July 2025 exit dates were revised within two months and are contested now.

The DDR4 still being made is not being made for you. Nanya’s president said on the July call that customers can get DDR5 from several suppliers but not enough DDR4 or LPDDR4 from anyone, and Winbond’s said in February 2026 that the DDR4 gap is “so large that it’s hard to see how it can be filled.” TrendForce said on 22 June 2026 that the big three had cut wafer allocations for DDR4, pushing buyers to the Taiwanese makers, and forecast on 3 July that their expansion could not fully offset the cut. The shortage has cascaded, too: some manufacturers are redesigning DDR4 products to use DDR3 and DDR3 products to use DDR2, whose contract prices TrendForce estimated on 22 June would rise about 55-60% in the second quarter, forecasting a further 35-40% in the third. The route below DDR5 is crowded all the way down to DDR3.

No leading-edge process will make DDR4 cheaper. Tom’s Hardware reported in August 2025 that no DDR4 would be made on the industry’s newest 10nm-class nodes, such as 1c or 1-gamma. The node moves in the table stop short of that: Micron’s DDR4 and SK hynix’s Wuxi line are on 1-alpha, behind the newest nodes, and Winbond’s 16nm DDR4 goes to television, networking, industrial and embedded buyers whose orders fill its capacity through 2027. A generation gets cheaper when its chips move to the leading edge in volume, and DDR4 is not going there.

The best DDR4 is no longer made at all. At Computex in June 2026 more than half a dozen module and board makers told Tom’s Hardware they were shifting production back toward DDR4, with a catch: “high-performance DDR4 dies, such as the famous Samsung B-dies, are no longer in production. As such, most of the revamped DDR4 kits will top out at a rather pedestrian DDR4-3600.” Fast, tight DDR4 kits are now a finite second-hand pool. Why the wafers went to high-bandwidth memory and server DDR5 instead is the subject of the 2026 DRAM shortage.

Both DDR4 platforms came back from the dead, and so did their prices

Two sockets make up the whole new-build escape route: AMD’s AM4, which takes only DDR4, and Intel’s LGA1700, whose 12th to 14th generation Core processors take either generation while the board takes one. Both were written off in early 2026: Tom’s Hardware’s January build guide said “Intel abandoned DDR4 as it sunset the LGA1700 socket”, and BGR wrote in April that “AMD stopped supporting it with the end of its AM4 socket”. By September AM4 had new processors and boards on sale, and Intel had said it would keep LGA1700 in its line-up for years.

On AM4, AMD announced on 31 May 2026 a Ryzen 7 5800X3D “10th Anniversary Edition”, on sale from 25 June at a suggested $349, which Tom’s Hardware noted was $100 below the original launch price at a time when resellers had been asking up to $800 for the old stock. AMD told Tom’s Hardware it would keep selling it “as long as it makes sense”, which the outlet read as not a limited run. A $99 Ryzen 5 5500F followed in September, according to TechSpot. And Gigabyte quietly launched two new B450 boards in August, eight years after the chipset, with four DDR4 slots and support for up to 128GB at DDR4-2667; Tom’s Hardware found no announced US price. The other platform limit people raise, PCIe, is the smaller one: XDA Developers’ 26 August 2026 case for AM4 cited GamersNexus’s RTX 5090 tests, in which PCIe 5.0 against 3.0 “amounted to a minuscule 1-4%”.

AMD’s specification page for the 5800X3D, read on 26 September 2026, gives the only official AM4 population figures found, in the table below, with 128GB at most and ECC listed as supported where the motherboard supports it. They apply to that processor only.

On LGA1700, Intel told Tom’s Hardware in June that it “will continue to make sure that there are products which can take care of older memory technologies”, and in August Intel’s Robert Hallock called Raptor Lake “a core part of the portfolio that I want to offer to people for years to come”. Tom’s Hardware also reported from Computex that Intel plans a third Raptor Lake refresh, “Raptor Lake Next”, for the first half of 2027; Intel declined to comment, and the outlet warned it could amount to little more than an infusion of stock. At least two board makers said they were ramping DDR4 boards into 2027, with board sales down by as much as 37% at some vendors, and Gigabyte had introduced a new LGA1700 DDR4 board days before Tom’s Hardware’s 14 August interview.

The LGA1700 rows come from Intel’s ARK entries, read on 26 September 2026: the i5-14600K lists DDR4-3200 or DDR5-5600, two channels and 192GB; the i3-12100F and i3-14100F list DDR4-3200 or DDR5-4800, with 128GB and 192GB respectively. On that socket the board, not the processor, fixes the generation for good, which DDR4 vs DDR5 explains; confirm what a particular machine takes with how to check what RAM fits or the CPU index.

The revival had a price. Tom’s Hardware reported in August that the i5-14600K, “$200 or less for the better part of last year”, had “jumped to around $250, if you can find it in stock at all”, and found the i7-14700K at $380 against about $330 expected. The platforms are open; they are no longer the cheap part of the build. The line in Tom’s Hardware’s own price index that it “makes little sense to build a DDR4-based system now due to processor and motherboard availability” predates the June and August launches. The availability argument has weakened. The performance argument, next, has not.

On the same processor, DDR4 gives up 9% to 17% in CPU-bound games

Two controlled tests are worth reading, and their benches explain their different answers.

Tom’s Hardware retested eight LGA1700 processors in July 2026 with the memory as the only variable: an RTX 5090, 1080p, fifteen games, XMP enabled on both. Its summary was that “you’re giving up around 11% to 14% of your gaming performance overall when the memory is the only variable that changes”, and the per-CPU table shows the loss growing with the processor.

Single games went further: Marvel Rivals lost 25.3% on the 14700K, and Spider-Man 2 about 18%. DDR4 also flattens a processor upgrade: on DDR4 the 14900K was only 1.7% faster than the 13900K, against 4.6% on DDR5, and the processor drew a few watts more on DDR4. Read the bench before the headline. It paired DDR5-7200, a fast kit, with a mainstream DDR4-3200 grade in four 8GB modules rather than two. It measures the gap for that pairing, not the cost of every DDR4 build.

TechSpot’s December 2025 retest used speeds closer to what people buy, though its DDR4 was a premium CL14 kit of the kind no longer made: DDR4-3600 against DDR5-6000 on the same Core i5-12400F, twelve games, again an RTX 5090 at 1080p. It found DDR5 “typically delivers a ~20% performance bump over DDR4 (on average)”. That is DDR5’s gain; on the July article’s basis, DDR4’s loss against DDR5, it is 1 - 1/1.20 = about 17%. TechSpot concluded that “if you’re building a new system or refreshing an older one, DDR4 shouldn’t be on the table anymore.” A larger gap from a smaller speed difference is not a contradiction: the games and the processor differ, and both move the answer, which is why this article gives a range.

The figure to discard is Tom’s Hardware’s own January 2026 estimate of “only around 5% (or less) in most cases”, which was not tied to a stated test and is contradicted by the same outlet’s July data.

The most awkward result for the escape route came three days after the gap test. Tom’s Hardware’s single-DIMM test found that one 16GB DDR5 module, which loses 8-11% against two on most processors (Ryzen 5 7600X -10.9%, Core Ultra 7 270K Plus -8.7%, Core i5-14600K -8.3%) and only 2.8% on a Ryzen 7 9800X3D, “offers superior performance to dual-channel DDR4 on CPUs that support both memory generations.” On the 14600K a single DDR5-7200 module beat dual-channel DDR4-3200, run as four 8GB modules, by 6.8% in 007 First Light and 6.9% in Starfield. The article cautions that “With slower speeds, the performance will equalize”, and its single-module runs had half the capacity of the dual runs.

Two limits apply to all of it. Every test above is CPU-bound by design, a fast graphics card at 1080p, and no test fetched for this article measured 1440p or 4K with the memory as the only variable. The nearest evidence is a whole-system comparison. Club386 (“DDR4 vs DDR5: a performance showdown of two gaming PCs”, 6 February 2026) ran two machines with the same RTX 5070 at 1080p, a Ryzen 7 5700X with 32GB of DDR4-3200 CL16 and a Ryzen 7 9700X with 32GB of DDR5-6000 CL30. Where the processor was the limit the DDR5 machine was far ahead, by 132-135% in Rainbow Six Siege X at medium settings; where the graphics card was the limit the gap nearly vanished, 1-3% in Assassin’s Creed Shadows and 3-4% in Cyberpunk 2077 with ray tracing. The processor changed too, so those figures bound the platform rather than the memory, but they show the shape: once the graphics card is the limit, the memory generation stops mattering much. A 2026 explainer from the module maker OSCOO put the DDR4 gap at 1-3% at 1440p and 4K without citing a test, and it is not relied on here. And AMD’s claim that the 5800X3D runs 10% faster than a Core i9-14900K when both use DDR4-3600 is a vendor claim, noted and not relied on. Why games are the sensitive case is worked through in DDR4 vs DDR5, the speed guide and the CAS latency guide.

Dollars per percentage point: the sum the reviews leave out

Tom’s Hardware’s July verdict held both sides in one article: “Dollars spent for performance gained, DDR4 makes complete sense”, and also “It’s a dead end, both for performance and upgrade potential.” The way to decide between them is to put the price and the penalty in one number. On a new LGA1700 build, with the memory as the only difference:

Tom's Hardware pairing, Core i5-14600K
  memory difference  32GB DDR5-6000 409 - 32GB DDR4-3200 190  = $219   (index, 2026-09-14)
  DDR4 penalty                                                = 13.1%  (test, 2026-07-28)
  per point          219 / 13.1                               = $16.7

TechSpot pairing, Core i5-12400F
  memory difference  DDR5-6000 2x16GB 572 - DDR4-3600 307     = $265   (PCPartPicker, Aug 2026)
  DDR4 penalty       DDR5 about 20% faster: 1 - 1/1.20        = about 16.7%
  per point          265 / 16.7                               = $15.9

One asymmetry in the first sum should be stated. The 13.1% was measured against DDR5-7200, while the $219 prices DDR5-6000. Against the slower kit the penalty would be smaller, by an amount no fetched test measures, so $16.7 a point, and every figure in the table below, is a floor on what a DDR5 point costs, not an estimate of it.

The same $219 spread across the eight processors in the July test gives a result that is more useful than either verdict: the weaker the processor, the more each point of DDR5 costs, because a weaker processor loses less to DDR4.

That is the memory alone. A DDR5 board and a DDR4 board for the same socket do not cost the same, and no LGA1700 board price for September 2026 was sourced for this article, so the board enters as the reader’s own quote rather than as an invented figure:

with boards    (219 + B5 - B4) / 13.1     B5, B4 = your DDR5 and DDR4 board quotes
               every $50 of board difference adds 50 / 13.1 = $3.8 per point

For scale only, Tom’s Hardware’s January 2026 DDR4 build used a $160 B760 DDR4 board; that is a January price, not a quote for today.

There is a third route, and it is Tom’s Hardware’s own advice: “bite the bullet on DDR5 and buy a weaker CPU.” A 16GB DDR5-6000 kit was $239 on the 14 September index, $49 more than the $190 32GB DDR4 kit, and on the single-DIMM evidence even one DDR5-7200 module outperforms dual-channel DDR4 in CPU-bound games, a margin Tom’s Hardware expected to narrow at slower speeds. The trade is half the capacity; whether 16GB is enough is answered in how much RAM you need. The DDR5-only Arrow Lake Refresh chips carried Intel recommended prices of $219 to $349 in July.

One detail decides the last row. Tom’s Hardware’s index does not say whether its $239 16GB kit is one module or two, and the usual 16GB kit is two: TechSpot’s June 2026 buying guide specified “16GB (2 x 8GB) DDR5-6000 CL36” at $245. Two 8GB modules run dual-channel, better than the single-module result above, but growing to 32GB then means replacing both or filling four slots, the replace-not-add trap the laptop sections below find in LPCAMM2. Read the module count on the listing; the same applies to the 16GB line in the AM4 sum below.

The rule that comes out of the arithmetic: the escape route pays best when the graphics card is the limit, when capacity matters more than frame rate, or when the board is already owned, and it pays worst on a new high-end processor, where DDR4 caps what that processor can do. Prebuilt machines, bundles and the timing of the next platform generation are the business of buy RAM now or wait.

If you already own the DDR4 machine, the drop-in beats the rebuild

This is the case the evidence favours most, and the reviewers who call DDR4 a dead end mostly agree with it. Tom’s Hardware’s June re-review of the $349 5800X3D concluded that it “really only makes sense if you already have an AM4 motherboard and DDR4 memory”, and was otherwise “about $70 to $100 too expensive”. TechSpot’s June buying guide said the same for both sockets: “If you already own a capable AM4 or LGA 1700 system, a targeted CPU or graphics card upgrade may extend its useful life without forcing you to replace the motherboard and memory.”

To see why, price an AM4 owner’s two options with September’s figures. B is the reader’s quote for an AM5 board and R is what the old processor, board and DDR4 would fetch; both depend on the reader, so both stay as letters.

drop-in          Ryzen 7 5800X3D                                     $349
switch, 16GB     Ryzen 5 7600X3D 230 + DDR5-6000 16GB kit 239 + B    $469 + B
extra cost       469 + B - R - 349                                   $120 + B - R
switch, 32GB     230 + DDR5-6000 32GB kit 409 + B                    $639 + B
extra cost       639 + B - R - 349                                   $290 + B - R
gain             7600X3D against 5800X3D in games                    about 20%
per point, 16GB  (120 + B - R) / 20

The gain comes from Tom’s Hardware’s July article, which found the Ryzen 7 7700X3D “nearly 20% faster in average gaming performance” than the 5800X3D and said “The $230 Ryzen 5 7600X3D shows similar gains”. The same outlet calculated in July that a 7600X3D plus 16GB of DDR5-6000 was “only $80 more” than a 5800X3D; that sum left out the board and used July memory prices, and at the September index it is $120.

One dated reference point for B exists. Tom’s Hardware’s best-motherboards guide, last updated on 4 August 2026, listed its B650 pick for AM5, the Gigabyte B650 Aorus Elite AX Ice, at $149.99, and warned that “B650 has been discontinued, so stock on those more affordable options will soon become scarce.” At B = $150 and R = 0 the switch costs $270 more than the drop-in, the middle row. The same guide’s DDR4 pick for AM4, the Asus ROG Strix B550-F Gaming Wi-Fi, was “Priced under $240”, more than its AM5 pick. Both are one outlet’s picks, so use them to check your own quote.

Resale R reduces every row, and moving 32GB of DDR4 to 32GB of DDR5 instead of dropping to 16GB adds $170 to each. Eleven to sixteen dollars a point is in the same range as the DDR5 premium on a new LGA1700 machine, but it buys a platform change for a machine that already works. What the drop-in gains over the processor a reader owns now depends on that processor, and no source here measures it.

Then a question readers often ask: should you sell the DDR4 now, while it is dear? Retail DDR4 is selling below module spot, so a used kit’s price is being set against cheap old stock, not replacement cost. eRacks’ own tracking showed the median DDR4 desktop gigabyte falling 13% between March and September 2026 ($7.19 to $6.22) and DDR4 server memory 25%, while DDR5 desktop rose 20%. At a fixed capacity the arithmetic is unforgiving: even if a used 32GB DDR4 kit sold for the full $190 of the best new one, 32GB of DDR5 at $409 would still cost $219 more. Selling DDR4 to fund DDR5 is a bet on inventory, not a free trade. The timing of a platform switch is the subject of buy RAM now or wait; the live medians for the sum are on 32GB DDR4 and 32GB DDR5, and the part index shows what an exact kit is worth.

Adding DDR4 capacity: fill the slots once, and read what four modules cost

Often the DDR4 purchase in 2026 is not a new build but more memory for a machine that already works, and three constraints specific to DDR4 decide it.

The first is the module ceiling. DDR4 unbuffered modules stop at 32GB, so four slots stop at 128GB, and a two-slot board at 64GB. There is no 48GB DDR4 module to fill the gap the way there is on DDR5.

The second is what four modules do to speed, on the AMD 5800X3D figures quoted above:

dual-channel peak = data rate x 8 bytes x 2 channels
DDR4-3200   3200 x 16 = 51.2 GB/s                two modules, AMD's figure
DDR4-2933   2933 x 16 = 46.9 GB/s   (-8.3%)      four single-rank modules
DDR4-2667   2667 x 16 = 42.7 GB/s   (-16.7%)     four dual-rank modules

Those are official speeds before any XMP profile; whether a profile holds at four modules is the board’s decision. Whether a 16GB DDR4 module is single- or dual-rank depends on its chips, and the organisation code tells you, as ranks and 3DS modules explains.

Which grade to buy follows from the same specifications. AMD’s 5800X3D page and Intel’s ARK entry for the i5-14600K both stop at DDR4-3200, so anything faster is an XMP profile the board may or may not hold. The premium for 3600 is small: the PCPartPicker averages in the first table put a 2x16GB DDR4-3600 kit at $307 against $281 for DDR4-3200, 307 / 281 = 9% more. What the timings buy can be worked out, because first-word latency in nanoseconds is 2,000 x CL / data rate:

DDR4-3200 CL16   2000 x 16 / 3200 = 10.0 ns
DDR4-3600 CL18   2000 x 18 / 3600 = 10.0 ns
DDR4-3600 CL16   2000 x 16 / 3600 =  8.9 ns

A 3600 CL18 kit has the latency of 3200 CL16 and 3600 / 3200 = 12.5% more peak bandwidth; only 3600 CL16 is quicker on both counts. For a new DDR4 purchase the default is DDR4-3200, the official speed, with nothing left to prove; pay the 9% for 3600 only on a board whose memory support list includes it, and test it inside the return window. How timings and speed trade off is set out in the CAS latency guide.

The third is matching. Tom’s Hardware warned in July that modules bought separately were never validated together, and advised the same brand and specification, tested at JEDEC speed before the profile is enabled. On DDR4 the high-bin chips behind the fastest kits are no longer made, so the part number is the matching tool, and the part index groups every listing of one.

What each path costs, from the two dated price sources and before whatever the old pair fetches:

second 16GB kit            Tom's index, DDR4-3200 16GB, 2026-09-14                $99
2x16GB, replace or add     Tom's index, DDR4-3200 32GB                            $190
64GB as two 2x16GB kits    PCPartPicker average, Aug 2026   2 x 281 = $562    $8.78/GB
64GB as one 2x32GB kit     PCPartPicker average, Aug 2026             $614    $9.59/GB
premium for 32GB modules   614 - 562 = $52, or 9.59 / 8.78 = 9% more per GB

On the best-price index the gap is wider: its 64GB DDR4-3200 kit was $498, 498 / 64 = $7.78 per gigabyte, 31% more than the $5.94 of its 32GB kit, module counts unstated. Larger DDR4 modules cost more per gigabyte, so buying the largest modules is a premium paid for full speed and free slots, not a saving.

The replacement path keeps full speed and leaves two slots free; the addition path keeps the old modules and costs about 8% of peak bandwidth, or 17% if they are dual-rank; the mixed path, at 8% or 17% by rank, buys the most capacity and the most uncertainty. The rule already set out in DDR4 vs DDR5 applies with more force in 2026: new DDR4 is a shrinking pool, so if you will want the capacity, buy it now and buy it once. On most boards the choice is then between two 32GB modules at full speed and that premium, or four 16GB ones at the four-module speed. How much capacity a workload needs is the subject of how much RAM you need, and the mixing rules that decide whether an old and a new kit will train together are in how to check what RAM fits.

Windows 10 ended in October 2025, and the oldest DDR4 machines are on borrowed time

Part of the escape route expires for a reason that has nothing to do with memory: an old DDR4 office PC or workstation bought to avoid DDR5 prices can run into Windows.

Microsoft’s end-of-support page says Windows 10 reached end of support on 14 October 2025, and that consumer Extended Security Updates run until 12 October 2027. Microsoft’s Extended Security Updates page, read on 26 September 2026, offers three ways to enrol: “At no additional cost if you are syncing your PC Settings”, by redeeming 1,000 Microsoft Rewards points, or by a “One-time purchase of $30 USD or local currency equivalent plus applicable tax”.

Windows 11’s supported-processor lists, both re-read on 26 September 2026, start above the first DDR4 generations. Intel’s, dated 16 October 2025, runs by series from 8th-generation Core, with the X299 “Intel Core 7000X Series” of the Skylake and Kaby Lake years as the one earlier Core entry, plus the Xeon Scalable and Xeon W lines in the table below. AMD’s, dated 25 July 2023, names models: desktop Ryzen from the Zen+ 2000 series, such as the 2600 and 2700X, and EPYC 7002 and 7003 models such as the 7252 and 7313. What searches of both found:

“Not listed” is a reading of Microsoft’s generated tables, not a statement that Windows 11 cannot be installed; exceptions for individual models, and Windows Server’s separate requirements, are not covered here.

The consequence for a second-hand buyer is concrete. A cheap Skylake or Kaby Lake desktop full of DDR4 is a supported Windows machine for about one more year, and only on Extended Security Updates, so budget the $30 or the settings sync and confirm enrolment before counting that year; a used Xeon Scalable workstation is not bound by the same date. The used DDR4 machines that stay on the lists are Intel desktops from the 8th generation on and Ryzen 2000 to 5000 on AM4, less the exceptions in the table. Check the processor on the CPU index, which records the DDR4 grade each family takes, and the machine on the server and workstation index before counting its DDR4 as a saving.

DDR4 laptops: about half the retail price per gigabyte and a 64GB ceiling

The laptop version of the escape route got a new product in August 2026. MSI launched a variant of its Katana 15 HX C14, with up to a Core i9-14900HX and an RTX 5070, that takes DDR4-3200 SODIMMs rather than DDR5. Tom’s Hardware priced the difference at the time:

32GB (2x16GB) DDR4-3200 SODIMM kit   about $200 / 32 = $6.25/GB     (Tom's Hardware, 2026-08-28)
32GB DDR5-6000 kit, type not stated  over $400 / 32  = over $12.50/GB
largest DDR4 SODIMM                  32GB, so two slots x 32GB = 64GB ceiling

The article did not say whether its DDR5 kit was a laptop kit, so the second line is a rough guide rather than a like-for-like price.

The same article doubted MSI’s “up to 96GB” figure for the DDR4 model, because no 48GB DDR4 SODIMM exists. A two-slot DDR4 laptop stops at 64GB. DDR5 laptop memory, meanwhile, has risen sixfold by one maker’s account: Schenker, which builds XMG laptops, said in September 2026 that DDR5 SODIMM prices had risen sixfold since July 2025, and Framework’s published price for DDR5 SODIMMs went from $10 per gigabyte in December 2025 to $13 to $18 by March 2026.

At the contract layer the laptop gap disappears, as it does for the desktop chip. TrendForce’s contract table for the second half of July 2026 priced a DDR4 16GB SODIMM at $265 and a DDR5 8GB SODIMM at $130, 265 / 16 = $16.56 against 130 / 8 = $16.25 per gigabyte, so a laptop maker buying modules paid no less for DDR4. At retail, Framework’s $13 to $18 per gigabyte for DDR5 SODIMMs in March against about $6.25 for the DDR4 kit in August is between a half and a third, though from two sellers five months apart. The retail laptop gap rests on stock, as the desktop one does.

The larger DDR4 laptop population is second-hand. The Register reported on 16 February 2026, from figures compiled by the analyst firm Context, that refurbished PC sales through distribution rose 7% in the fourth quarter of 2025 across five large European markets; no equivalent US figure was found.

On those machines the upgrade worth most is often the one people skip: the empty second slot. A two-slot laptop with one module fitted runs its memory single-channel, and a laptop without a graphics card takes its graphics bandwidth from system memory. Intel’s own branding rules make the point in writing: its product specification footnotes, read on 26 September 2026, say Iris Xe branding requires dual-channel memory, and Arc branding on H-series Core Ultra systems requires at least 16GB in a dual-channel configuration. Tom’s Hardware, whose July 2026 single-DIMM test used a discrete graphics card, made the same point in passing: “dual-channel memory is vital if you’re using an iGPU.” Adding a matching second module turns single-channel into dual-channel, which is a larger change than replacing the one module with a bigger one.

Three checks before buying. The Windows 11 cut-off applies to 6th and 7th-generation Intel laptops, and to Ryzen 2000-series laptops, as it does to desktops. DDR4 and DDR5 SODIMMs look almost identical in a photograph, so read the generation off the part number, as DDR4 vs DDR5 explains. And identify the machine before the memory, using how to check what RAM fits. The two generations are listed separately here, as DDR4 SODIMMs and DDR5 SODIMMs.

On the trackers, server DDR5 costs four to six times server DDR4 per gigabyte

For servers and homelabs the DDR4 case is strongest and the sources are weakest: there is no neutral price series for used server memory, only four sources on four bases, which disagree on level and agree on direction.

The two trackers put DDR5 at 31.05 / 5.56 = 5.6 times and 41.09 / 9.34 = 4.4 times DDR4. PCSP’s comparison, which pairs its own catalogue price of $160 to $235 for a refurbished 32GB DDR4 RDIMM with a new-market quote of $400 to $500 for a new DDR5 one and calls itself “the shape of the gap, not a price list”, gives a gap nearer two to two and a half times, and its DDR5 figure is about half of eRacks’ or less. Both eRacks and PCSP sell servers and neither method has been audited, so the table states their spread rather than choosing. PCSP reports refurbished DDR4 up 30-50% from late-2025 lows while eRacks shows it down 25% since March: direction depends on the window.

Now a worked host, because a per-gigabyte ratio is not what a server costs.

DDR4 host: two EPYC 7003 sockets, 8 channels each, one 32GB module per channel
  16 modules x 32GB = 512GB
  at $5.56/GB (eRacks)              512 x 5.56  =  $2,847
  at $9.34/GB (DatacenterDisk)      512 x 9.34  =  $4,782
same 512GB in DDR5
  at $31.05/GB                      512 x 31.05 = $15,898
  at $41.09/GB                      512 x 41.09 = $21,038
  at $64.06/GB (module spot)        512 x 64.06 = $32,799

The DDR5 lines flatter DDR5, because 512GB is the wrong comparison. The largest current DDR5 server processors have twelve channels per socket (Intel lists twelve for its Xeon 6980P and AMD the same for its EPYC 9965; Intel’s Xeon 6700P series has eight), so a balanced two-socket host of that class at one module per channel wants 24 modules. Keeping the 32GB module, that is 24 x 32GB = 768GB, which at $31.05 is 768 x 31.05 = $23,846. 24GB DDR5 RDIMMs also exist, and 24 x 24GB = 576GB is the nearest balanced total to 512GB, but none of the trackers here prices them separately. Filling only some channels to stop at 512GB gives up bandwidth, as server memory capacity limits explains with its population tables.

What the cheap gigabyte buys less of is bandwidth. AMD lists the EPYC 7763 at DDR4-3200 on eight channels, 204.8 GB/s per socket (3200 x 8 bytes x 8), and the EPYC 9965 at DDR5-6400 on twelve channels, 614 GB/s (6400 x 8 x 12 = 614.4). The DDR4 socket has a third of the bandwidth. For a capacity-bound workload, a file server, a lab of idle virtual machines, a cache, that does not matter; for anything that streams memory it matters a great deal. Running cost is the other term, and the method is in buying used RAM on eBay: extra watts times 8,760 hours times your price per kilowatt-hour, with the watts as your own measured input because no source here gives them for these hosts. For scale, every 100 W of continuous difference is 100 x 8,760 / 1,000 = 876 kWh a year, and PCSP’s June 2026 used-server guide warns that “Older platforms draw more watts per unit of work”.

The live market is on registered DDR4, load-reduced DDR4, 32GB and 64GB modules; which type a machine takes is on the server index, and how much a server needs is in how much RAM you need.

A DDR4 RDIMM fits a desktop slot and will not start the machine

The homelab trap is specific to DDR4, and a September 2026 news story got it half wrong. Reporting a Reddit user’s haul of 30 Samsung 64GB DDR4 RDIMMs from a retired work server, Tom’s Hardware wrote that such modules “will be both physically and electrically incompatible” with a home PC. For DDR4 only “electrically” is right. DDR4 unbuffered and registered modules share the same key notch, as this site’s guides to registered and unbuffered memory and ECC set out, so a server RDIMM seats in a desktop slot and the machine simply does not start. DDR5 moved the registered module’s notch; DDR4 never did. That is exactly why the mistake is easy to make with DDR4 and why the listing has to be read for the word “registered” rather than judged by the photograph.

Three routes follow for someone who wants cheap capacity at home, with or without error correction. A used server or workstation that takes RDIMMs, looked up by model on the Dell, HPE, Lenovo or Supermicro pages before buying the memory. An AM4 desktop with ECC unbuffered modules, where the board’s manual decides; Intel’s ARK entry for the i5-14600K also lists ECC, but whether a DDR4 LGA1700 board enables it was not established here. Or no ECC at all and the cheapest unbuffered gigabyte.

Within servers the types do not mix: RDIMMs, LRDIMMs and 3DS modules must each match the machine and each other. And lots need reading module by module. Big Iron’s July 2026 homelab guide warned that “Broker lots pulled from decommissioned server fleets are frequently a mix of speed grades, ranks, and sometimes manufacturers even when sold as a matched set”, practitioner advice that the label-reading method in buying used RAM on eBay is built to catch. On this site, the ECC DDR4 page includes registered modules, so for AM4 confirm “unbuffered” on each listing; the server DDR4 page is the start for the other routes, and rank arithmetic is in ranks and 3DS modules.

Used memory comes from five places, and the server stream is starting to keep its own

Whether memory wears out has a short answer: not the way an SSD does. DRAM has no write-endurance budget. What ages is the solder under each chip, which fatigues with thermal cycles, the spare rows consumed by post-package repair, and the edge contacts, as buying used RAM on eBay sets out; which of those a module’s history makes likely is the subject of this part.

Used memory reaches the market by five routes, each leaving a module with a different history.

The first stream supplies most cheap registered memory, and the certified end of that trade, R2v3 and e-Stewards, is covered in buying used RAM on eBay. The second is how the Reddit user in the previous section came by 1,920GB: an employer let them keep a retired server, which, Tom’s Hardware suggested, may have saved the employer some recycling fees.

The stream that is changing is the first. Meta described in June 2026 a custom CXL 2.0 memory-expander chip, Vistara, that attaches DDR4 recovered from decommissioned servers to new DDR5-only EPYC servers: each carries 768GB of local DDR5-6400 plus 256GB of CXL-attached DDR4-2400 built from 32GB modules, per Tom’s Hardware’s report of 30 June. The Register, reading the paper on 29 June, reported that the expected service life of Meta’s servers is three to five years, “but memory is useful for seven to ten years”. Marvell’s Khurram Malik told EE Times in August that “The first and foremost important use case within CXL is the recycling of DDR4”, and Tom’s Hardware observed on 24 August that a module pulled from a decommissioned 2021 server “is now worth a multiple of what it was when Structera X was announced”, Marvell’s CXL memory-expansion chip of July 2024.

Modules kept in service are modules that never reach a broker. That is an inference, not a measured flow: no source gives a volume, and one operator’s reuse is not every operator’s. But the direction runs against the buyer, and demand points the same way: Apacer’s chief executive forecast in July 2026 that chip supply from the major makers to independent module makers could fall in 2027 to about 30% of its 2026 level, and fewer new retail modules would push more demand onto the used market. For DDR4 server memory the used stream is increasingly the market, and the server DDR4 page is where it shows.

Ex-server modules have the best history on the market, and one gap in it

The case for ex-server DDR4 is the one buying used RAM on eBay makes from the fleet studies, summarised here in a sentence each. Google’s 2009 study found memory errors “dominated by hard errors”, which a test finds, and no infant mortality, which the authors put down to burn-in weeding out bad DIMMs before they entered service. Server modules sit at constant temperature, which spares them the thermal cycling that fatigues solder, and are retired on schedule rather than for a fault. The 2026 addition is Meta’s claim, from an operator with an interest in reusing it, that DRAM retired at three to five years has years of service left.

Age is not neutral, though. The same Google study found correctable-error rates starting to climb after 10 to 18 months in the field, and Facebook’s fleet study (Meza, Wu, Kumar and Mutlu, DSN 2015) found that “older machines generally have higher failure rates than younger machines”: in one configuration, with processor count and chip density controlled, the rate rose 2.8% from two to three years of age and 7.8% from three to four, though with no clear trend when only density was controlled, on DDR3-era servers one to four years old. An ex-server module is a tested survivor, not a new part, and Alibaba Electronics’ 15 August 2026 used-RAM guide, which calls a five-year server module “statistically no more likely to fail”, cites no source for it.

The gap is the error history. Google’s study found that 70-80% of uncorrectable errors were preceded by a correctable error in the same or the previous month, so a module’s correctable-error count is the best predictor there is of its next fault. That count lived in the host, not on the module, and it left with the host. Unless a seller kept the server’s logs, nobody can tell you what a module did in service. A long test after purchase is the substitute, and the method is in the used-RAM guide.

What the label does tell you is roughly where a module came from. PCSP, a refurbished-server seller, publishes a platform matrix that ties DDR4 speed grades to server generations:

It is a seller’s summary. This site’s server capacity guide records HPE running EPYC 7002 at 2933 in a DL385 Gen10 and 3200 in a DL385 Gen10 Plus: same chip, different board. A module’s label speed therefore says which machines will take it back at full speed, and its manufacturing date is in its SPD and the chip date codes, which the used-RAM guide explains how to read.

One correction belongs here. A retailer’s April 2026 explainer carried a table row for “DDR4-4800 (server)” at CL40. There is no such server part. DDR4-4800 was sold only as overclocked desktop kits, such as G.Skill’s discontinued F4-4800C18D-16GTRS rated CL18 at 1.5 V; DDR4’s standard rates stop at 3200 MT/s, per the DDR4 standard as summarised on Wikipedia, and so does PCSP’s matrix. The row describes a DDR5-4800 RDIMM with the wrong generation on it. The live market for each grade is on DDR4-2400, DDR4-2666, DDR4-2933 and DDR4-3200, and the PC4 code on a server label is decoded in the speed guide.

“Ex-mining” means three workloads, and only one of them rewarded overclocking the memory

“Mining RAM” is used as though it described one history. It describes three, and there is no failure data for memory from any of them, so the argument has to come from the workloads’ own documentation.

GPU mining was mostly Ethereum, and it ended at the Merge on 15 September 2022, when Ethereum moved to proof of stake. In a GPU rig the graphics cards did the work; the system memory mostly sat there. No source found for this article states how much memory typical rigs carried, so none is claimed. Memory from rigs that stopped at the Merge has been out of that work for four years.

Chia plotting is the opposite. The official BladeBit plotter’s documentation says “416 GiB of RAM are required to run it” for in-memory CPU plotting, and Chia’s documentation gives 256GB for fully in-memory GPU plotting. Plotting is a sustained, write-heavy memory workload; the link from those figures to particular second-hand lots is an inference, and no market source ties them.

CPU mining with Monero’s RandomX needs little capacity, about 2GB per NUMA node (RandomX’s README gives 2080 MiB for its fast mode), but it is limited by memory bandwidth: “DDR4 memory is limited to about 4000-6000 H/s per channel (depending on frequency and timings)”, per the same document and XMRig’s optimisation guide. A RandomX rig keeps every channel saturated around the clock, and the hash rate rises with memory speed, which rewards overclocking the memory. Of the three, it is the one that plausibly ran the memory both flat out and overclocked.

Temperature is the usual worry, and the evidence is narrow: Google’s fleet study found only a marginal effect on error rates within the roughly 20 °C range of normal data-centre operation, which says nothing about an unventilated mining frame. Corsair’s warranty, last edited on 11 August 2026, excludes “Products damaged during crypto mining operations”, which tells you how the maker views the use, not what it did. The practical test is the same as for any module, plus one question: which of the three the seller means. Chia plotting memory is likely to be registered, which changes what to test for.

Every enthusiast kit is rated as an overclock, and no module remembers how it was run

“Was it overclocked?” sounds like the right question about a used enthusiast kit. It is the wrong one, for a reason that 2026’s settlements made explicit.

A DDR4-3600 kit reaches 3600 only through its XMP profile. Without it the kit boots at the JEDEC grade stored in its SPD, commonly DDR4-2133 or DDR4-2400 (Kingston’s datasheet for a FURY Beast DDR4-3600 module gives DDR4-2400 at 1.2 V). G.Skill began paying a $2.4 million US class-action settlement in August 2026 over exactly this, paying claimants about $20 to $25 and agreeing to state on packaging and listings that XMP or EXPO overclocking is needed to reach rated speeds; it denied the allegations. Team Group agreed a $1.1 million settlement over the same claim in May, also denying wrongdoing.

The processor makers call it overclocking too. AMD’s EXPO page says running memory outside AMD’s published specifications “will void any applicable AMD product warranty, even when enabled via AMD hardware and/or software.” Intel’s XMP page warns that changing frequency or voltage may void warranties and may reduce the stability, performance and life of the processor and other components. So a used enthusiast kit ran overclocked by design if its owner enabled the profile, and that is normal use, not abuse. Not every owner did.

The question that matters is whether it was pushed beyond its profile, or run at raised voltage, and no DDR4 module can answer it. DDR4 has no power stage of its own; its supply comes from the motherboard’s regulator, so the voltage history lived in the board and the module records nothing. No study fetched for this article quantifies degradation at XMP voltages, and none is implied. The often-repeated comparison of DDR4’s 1.2 V with DDR5’s 1.1 V is the JEDEC nominal; XMP profiles can run above it, and no 2026 source for typical values was found.

The kits most likely to have been tuned are the high-bin ones, and since Samsung’s B-die left production those are a used-only pool. What a buyer can actually check:

• the part number on the label against the maker’s rated profile for that kit;
• the SPD contents against the label, since G.Skill’s warranty excludes modules with modified SPD firmware;
• stability at the XMP profile, then at JEDEC speed, inside the return window.

All three are explained in buying used RAM on eBay, XMP as a memory-controller overclock in the speed guide, and the most-tuned grade trades on DDR4-3600.

“New” is not a provenance either

In 2026 the line between used and new does not track the line between old and new silicon, so the forms “new” takes matter.

Recovered chips are the first. In 2023, during a DDR4 glut, TrendForce reported that DDR4-3200 chips were being desoldered from old server modules, re-flashed and built into new consumer kits, as Tom’s Hardware relayed; that is the only DDR4-specific report found, and it predates the shortage. In 2026 recovered silicon is a traded commodity again: TrendForce’s spot report of 16 September noted that “reballed chips sustained the upward price momentum seen last week”. Nothing found shows such chips in 2026 DDR4 kits, but the incentive is larger now that DDR4 silicon costs more than DDR5’s.

Lower grades are the second. DRAMeXchange quotes a separate DDR4 line labelled eTT, which its page does not define, at $13.65 for a 16Gb chip on 24 September against $83.78 for the standard 3200 grade, and $5.87 against $46.11 for 8Gb chips. A chip at a sixth of the standard price is not the same part, whatever the module’s label says. A cheap new no-name kit is not evidence of first-grade silicon.

The rest are 2025-2026 cases, most of them DDR5, whose lessons carry to DDR4.

Two more show why label and silicon need not agree. In January 2026 a modder built a working DDR5 desktop module from laptop-module chips on a $7.50 blank board, with SPD copied from an ADATA retail kit. And ASRock’s half-width HUDIMM and HSODIMM, announced in April 2026 with Intel and Team Group, use one 32-bit sub-channel, half the bandwidth and capacity of an ordinary module, and can look like one in a listing.

Two cautions keep this from turning into paranoia. A label maker that differs from the chip maker is a flag to test, not proof, because some genuine modules carry another maker’s chips. And a bare module is normal for SODIMMs and server RDIMMs; one 2026 used-RAM guide read a missing heatsink as a sign of “thermal stress”, which is wrong. Tom’s Hardware put it the other way round: fakes are easier to spot on bare modules, and “For desktop memory that does come with heatsinks, you’re basically out of luck until you actually test it.” The checks are in buying used RAM on eBay, and the part index shows what a part number should be.

Between a cheap new no-name DDR4 kit and a used branded one at a similar price, the used kit is usually the more checkable purchase. Its part number names a rated profile and a chip maker, and the label, the packages and the SPD can each be tested against them; the no-name kit’s label promises nothing a test can contradict. Either way, the test inside the return window is what decides.

The warranty stays with the first owner, and G.Skill’s “lifetime” ends with the model

The obvious cost of buying used is the missing warranty. In the US, for G.Skill and Corsair, whose terms say so, it is missing entirely, and in 2026 even a new kit’s warranty can be worth less than it says.

Two terms matter for DDR4 in particular. G.Skill’s “lifetime” warranty ends automatically when G.Skill discontinues that model, so on an older DDR4 kit it lasts only as long as the model stays in G.Skill’s range. And Crucial is gone from retail: Micron announced on 3 December 2025 that it would ship Crucial consumer products only until the end of February 2026, while promising continued warranty support, and Crucial’s warranty page now says that with replacement parts very limited, a refund of the original price, taxes and shipping is “the maximum resolution available during the closure”. Tom’s Hardware reported in August 2026 that Micron first offered one claimant $241.86 for a 48GB DDR5 kit when the cheapest 48GB kit on sale, a faster one, cost $649.99, 241.86 / 649.99 = 37.2%, before sending three 16GB modules instead. Corsair’s page notes that EU, Swiss and UK buyers have two years by statute. Kingston’s terms could not be fetched and are not described.

So a used kit bought in the US normally carries no maker’s warranty, and the protection that remains is the platform’s return path, set out in buying used RAM on eBay.

Ex-server hardware from a refurbisher is the partial exception, because the seller’s terms replace the maker’s. PCSP’s 22 June 2026 buying guide gives its own: returns within 90 days, a server warranty of “1 year, extendable up to five”, and a 20% restocking fee on items ordered in error or “returned as defective but found fully working”. It also warns that “used”, “refurbished” and “recertified” “are not standardized across sellers”. Those are one seller’s terms for whole servers, but they show what to ask any seller for: a window long enough for a full test, and warranty terms in writing. A warranty that outlasts the test also takes the failure probability out of the sum below.

One US price site’s rule of thumb is to buy used only at about 60-70% of a comparable new kit; it gives no basis for the figure. A sum does better:

break-even chance of failure after the return window = discount / replacement cost

new 32GB DDR4-3200 kit, best price (Tom's index, 2026-09-14)    $190
used kit at, for example, $150                                   discount $40
break-even                                                       40 / 190 = 21%

Unless you believe a used kit that passes its tests has more than a one-in-five chance of failing after the window closes, the used kit and its $40 saving are the better bet. The probability is your input, since no source here measures it for DDR4, and the test that sets it belongs inside the return window.

Three kinds of laptop memory, and only two of them can be changed

Upgradeability in a laptop is decided by the form the memory takes, and there are three: a socketed module, soldered LPDDR, and a compression-attached module.

SODIMMs are the open market: 260 pins for DDR4, 262 for DDR5, usually two slots, and an upgrade adds capacity to what is there. DDR5 also has a clocked variant, the CSODIMM, with a clock driver on the module. Soldered LPDDR has no socket at all, and the configuration bought is the one kept.

Between them sits CAMM2. JEDEC published JESD318 on 5 December 2023, and it covers two different modules: DDR5 CAMM2, “intended for performance notebooks and mainstream desktops”, and LPDDR5/5X CAMM2, the one called LPCAMM2, for “a broader range of notebooks and certain server market segments.” The pinouts differ, and JEDEC built “intentional variations in mounting procedures” so that one cannot be fitted where the other belongs. LPCAMM2 uses a dense connector with 644 or 666 contacts. A dual-channel LPCAMM2 fills both memory channels on its own, so an upgrade replaces the module rather than adding to it, which is the fact that decides its economics, worked below.

The last row is worth knowing because it is easy to mistake for a normal two-slot machine: some laptops, as Newegg’s May 2026 upgrade-planning guide describes, combine soldered memory with a single slot, so only the slotted half can grow.

The 2026 machines in this article’s sources that still take SODIMMs are the MSI Katana 15 HX C14 above, Framework’s AMD Ryzen AI 300 models and the Dell Pro Max 18 Plus’s SODIMM option, both below. That is not a census, and no source here counts SODIMM models. The rule that sorts any specification sheet is simpler. “SODIMM”, or a slot count, means socketed; “LPCAMM2” or “CAMM2” means one replaceable module; “LPDDR5” or “LPDDR5X” with neither word means soldered. Lenovo’s PSREF sheets, quoted below, spell out the connector, and they are the model to look for on any maker’s sheet.

Identify the machine rather than the processor, too: Intel’s Panther Lake supports DDR5, LPDDR5X and LPCAMM2, according to ServeTheHome and PC Perspective’s January 2026 launch coverage, so the same chip can arrive in all three forms. How to read the form factor off a running laptop is in how to check what RAM fits, and the two socketed generations are on DDR4 SODIMMs and DDR5 SODIMMs.

Soldered memory is a capacity decision made once, at the till

On a soldered machine, the capacity chosen at purchase is the capacity for the machine’s life. These platforms on sale in 2026 offer nothing else:

Lunar Lake carries its memory on the processor package itself, in 16GB or 32GB versions, and Wikipedia’s account of it notes that Apple’s M-series chips do the same, with the unified memory in the processor’s own package; for the rest the sources say only that the memory is fixed at purchase. The reasons makers give are power and space: the same article reports Intel’s claim that Lunar Lake’s on-package memory achieved “a reduction of 40% in power consumption”, without saying against what, and saved “up to 250 square millimeters”. Micron measures LPCAMM2’s own space saving against a dual-stacked SODIMM, below. Both are vendor claims.

What the up-front decision costs has a price list, and in June it went up. 9to5Mac reported in March 2026 that Apple’s step from 24GB to 48GB on the 14-inch MacBook Pro was still $400, 400 / 24 = $16.67 per added gigabyte. Apple raised prices across its Macs on 25 June 2026, citing memory costs, and Basic Apple Guy’s July 2026 price charts put the 14-inch M5 Pro MacBook Pro with 1TB at $2,499 with 24GB and $3,099 with 48GB, so the same step now costs 3,099 - 2,499 = $600, or $25 per gigabyte. On the M6 Mac mini, Tom’s Hardware found, going from 16GB to 24GB costs $200 and to 32GB $400, also $25 per added gigabyte. What the shortage did to laptop prices generally is in the 2026 DRAM shortage.

The makers are also shipping less memory by default. TrendForce said in December 2025 that notebook brands would hold capacities near the minimum, and that thin notebooks with soldered LPDDR “cannot reduce costs by lowering specifications or replacing modules”. Microsoft’s Copilot+ PC floor, 16GB of DDR5 or LPDDR5, is the only official platform minimum above 8GB found.

The rule on a soldered machine: buy the capacity for the machine’s whole life, because there is no second chance to buy it cheaper. What that capacity is, for a given workload, is in how much RAM you need; how to confirm before buying whether a model is soldered is in how to check what RAM fits.

LPCAMM2 is upgradeable in principle and scarce in practice

LPCAMM2 was supposed to end the choice between thin laptops and upgradeable ones. More than two and a half years after launch it has delivered the first half of that and not yet the second.

Micron introduced it on 9 January 2024 with claims of up to 61% lower power and 64% space savings against SODIMMs. Its own footnotes carry the conditions that most retellings drop: the power figure is “lower active power per 64-bit bus at the same DDR5 speed”, and the space figure is “compared to dual-stacked SODIMM”. Both are vendor claims against a particular comparison.

Machines also run it slower than the rating on the module. Lenovo’s PSREF sheet for the ThinkPad P1 Gen 7 says the “System comes with LPDDR5X-7500 modules and will run at 7467 MT/s due to platform limitations”, Framework runs 8533-rated modules at 7467 on Intel’s Core Ultra Series 3, and Lenovo’s sheets for the ThinkPad T14 Gen 7 and T16 Gen 5 say their 8533 modules run at 6800 or 7467. “8533” on a specification sheet is usually the module’s rating, not the speed it runs at.

That is everything verified in the sources for this article, and it corrects one common error: Framework’s Laptop 13 Pro uses LPCAMM2 only in its Intel version, and the AMD Ryzen AI 300 models keep SODIMMs, as Tom’s Hardware reported in July. No retail desktop CAMM2 board was found either; in July 2025 Tom’s Hardware found none on sale, and Crucial was then the only retail seller of CAMM2-family memory.

One older module belongs in a second-hand buyer’s reckoning. Dell used a compression-attached module before the standard existed: in April 2022 its Precision 7000-series laptops shipped with what Wikipedia’s CAMM article calls “a custom form factor of CAMM for DDR5 SDRAM”, while JEDEC standardised CAMM2 only on 5 December 2023, and the same article names the ThinkPad P1 Gen 7 of April 2024 as the first machine to use it. Nothing fetched for this article says a 2022 Precision accepts a JEDEC module, so match its memory by Dell part number.

Then the prices, which went the wrong way:

Crucial, May 2024             32GB $179.99 = $5.62/GB    64GB $329.99 = $5.16/GB
Framework, before July 2026   16GB $239    = $14.94/GB
                              32GB $439    = $13.72/GB   64GB $849    = $13.27/GB
Framework, July 2026          32GB $800    = $25.00/GB   64GB $1,600  = $25.00/GB
64GB, July 2026 against May 2024          1,600 / 329.99 = 4.85 times

Framework raised its prices in July after its supplier quoted, in chief executive Nirav Patel’s words, “prices that are more than double that of the prior inventory we had brought in”. In September it cut the 32GB and 64GB prices again, with retroactive refunds, to figures the sources did not capture but which Tom’s Hardware reported were still above the original baseline. Crucial left retail in February 2026, Tom’s Hardware’s July review of the Laptop 13 Pro found “LPCAMM2 is hard to get right now (harder, in general, than regular SO-DIMMs, which at least some enthusiasts might have around)”, and a second-hand market barely exists. The same review noted that fitting the module needs care with its interposer and a three-screw cover; it is not a SODIMM clip.

An upgrade slot is worth what the module for it costs, worked in dollars

“Upgradeable” is a selling point only if the module bought later costs less than the capacity bought now. Price one path, 16GB now and 32GB later, on each format:

LPCAMM2, Framework July 2026    16GB now 239 + 32GB later 800 = $1,039
  (16GB at its pre-July $239)   32GB up front                 = $800     staged path +$239, +30%
LPCAMM2, Framework before July  239 + 439                     = $678
                                32GB up front                 = $439     staged path +$239, +54%
SODIMM, Framework March 2026    each 16GB module at $13-18/GB = 16 x 13 to 16 x 18 = $208 to $288
                                16GB now + 16GB later         = $416 to $576
                                32GB up front, same rate      = $416 to $576, nothing discarded
Soldered, Apple M6 Mac mini     16GB to 32GB, at purchase only = $400 extra, no later path

On a SODIMM machine the staged path costs nothing extra at a steady price, because the upgrade adds a module. On LPCAMM2 it costs the whole first module, because the upgrade replaces it, and the replaced 16GB module becomes a spare with almost no resale market. On a soldered machine there is no path at all. The SODIMM figure is Framework’s own March 2026 retail price, and Framework reported rising costs for 16GB DDR5 modules in July, when it also said the $239 16GB LPCAMM2 price would last only until that stock ran out (Tom’s Hardware, 2026-07-22); a second-hand module may cost less, and DDR5 SODIMMs and DDR4 SODIMMs are where to check. One caveat on SODIMMs: 16GB as a single module runs single-channel until the second is fitted, which matters with integrated graphics.

eRacks’ founder put the option value plainly in September: “With slots you can act on that. With sealed memory you cannot.” No regulation is about to force the choice: the EU’s Right to Repair Directive, due for transposition by 31 July 2026, covers only product groups already under EU repairability rules, which include servers and phones but not laptops, according to the campaign group Right to Repair Europe, and no rule found requires replaceable laptop memory as of late September 2026.

How much capacity is enough is the question how much RAM you need answers, and whether to buy a laptop now or later is in buy RAM now or wait.

Filter by the machine first: which page on this site answers which question

Every decision in this article comes down to a comparison at a fixed capacity and type, and that is how the listing pages here are built: each sorts by price per gigabyte and prints the cheapest and the median for the view at the top. Bid-only auctions are left out by default, because a bid is not a price (?bids=show brings them back, ?auctions=no removes auctions entirely). Condition is shown on every listing, so new, used, refurbished and for-parts modules sit on one page, and the part index groups every listing of one part number.

Two habits make the numbers honest. Compare medians, or compare at a fixed capacity, never the cheapest rows. And treat a DDR4 desktop row far below the median printed at the top of its own page as a mismatched listing until proven otherwise: it may be a server module, a single module in a kit-priced title, or a typo. Module spot is no floor for this test: DRAMeXchange’s 14 September trade price for a 16GB DDR4 UDIMM was $10.27 per gigabyte while the best US 32GB retail kit sold at $5.94, for the inventory reasons set out near the top. The cheapest DDR4 rows of all are usually genuine registered server modules, cheap because decommissioned servers supply them in bulk and useless in a desktop because of the shared notch.

How each field is recorded is on the methodology page, and processor support on the CPU index.

What to do with this on a listing page

No page here can tell you provenance, overclocking history or warranty status, and it helps to know what stands in for each before starting.

Then, in order:

1. Identify the machine and its Windows 11 status first. The processor, the board and the Microsoft list decide what the memory is worth, via the CPU index and how to check what RAM fits.
2. Decide keep, add or switch with the per-point sum, using your own board quote and resale figure rather than anyone’s verdict.
3. Compare at a fixed capacity using medians, never the cheapest rows, and on the all-generations table when crossing generations.
4. Read the part number and the rank before the price, and match a second kit by part number on the part index.
5. Fill the slots once. On DDR4, buy the capacity you will want now: two large modules keep full speed at a premium per gigabyte, four smaller ones cost less and run slower.
6. For a server, look up the machine and buy by its channel count, on the server index, not by the price of the cheapest lot.
7. Ask which kind of mining, if a seller mentions it, and test accordingly.
8. Treat a cheap “new” no-name DDR4 kit as a claim to test; a used branded kit is often the more checkable buy.
9. Price the missing warranty with the break-even sum, not a percentage rule.
10. For a laptop, find out whether an upgrade adds a module or replaces one before paying for “upgradeable”.
11. Book the test inside the return window, using the method in buying used RAM on eBay. A listing marked “untested” or “no returns” has no window to book, which is why Windows Central’s May 2026 scam round-up told readers to avoid second-hand RAM sold that way.

As of late September 2026, the three parts of the question have three different answers. DDR4 is still the cheaper gigabyte, at roughly half DDR5’s price at US retail, but it more than doubled in a year, its silicon now costs more than DDR5’s, and its cheapness rests on inventory that cannot be replaced at the price it was bought; the escape route is open, far dearer than a year ago, and least rewarding for people who have not yet bought the board. Used memory is as safe as its origin is legible, and the record that matters most, a server’s error log, rarely travels with the module. And an upgrade slot is worth what the module for it costs: a SODIMM adds, an LPCAMM2 replaces, and soldered memory is bought once, at the till.