Buy SSDs and hard drives now or wait? What the 2026-2028 forecasts say

If you will need the storage within a year, buy it now: no maker or named analyst traced here expects hard drive prices to fall before 2028, and the earliest any of them gives for cheaper SSDs is the second half of 2027. Seagate told investors on 28 July 2026 that the vast majority of its nearline output is allocated into calendar 2028, and Western Digital said on 5 August that it is negotiating supply agreements for 2029, 2030 and 2031. On the flash side TrendForce expects NAND supply to overtake demand in the second half of 2027, while Micron, Samsung, Sandisk and Kioxia all expect flash to stay short through 2027, and Micron said on 30 September that 2028 will be short as well. If the drives you own work, your data has a second copy and nothing is filling up, waiting for an SSD upgrade is defensible. Waiting for cheaper hard drives is a bet against every forecast on the record, and the arithmetic below shows how rarely that bet pays even when it is won.

The answer most buying advice gives this year is one sentence long: if you need it, buy it; if you do not, wait. It is right, and it leaves out nearly everything that decides a real purchase. It treats “storage” as one price when hard drives and flash run on different clocks, with different causes and different relief dates. It treats a forecast as a date when the forecasts in this cycle have been raised, revised and pushed outwards almost without exception. It never says what waiting costs, although a drive that fails while you wait, a cloud bill while you wait and a full disk while you wait all have prices. And it ignores the one discount that survived the shortage: on 2 October 2026 ServerPartDeals listed a manufacturer-recertified 24 TB enterprise drive at a third less per terabyte than a new 8 TB NAS drive cost on Amazon in late September, and for less than any new consumer hard drive in the comparisons traced here.

What waiting is worth can be put in money. On the most favourable path built from a dated forecast - a fall the size TrendForce forecast for 2019, arriving by the end of 2027, as soon as TrendForce expects flash supply to loosen - a buyer who can go without a 2 TB NVMe drive until then saves about €4.70 to €5.60 for each month of waiting. A buyer of two 8 TB NAS drives who waits a year and gets the fall IHS iSuppli said in 2012 it would take to undo the Thai floods, and called remote, saves about $12 a month, and renting the same 8 TB from Backblaze B2 for that year costs $667, more than the pair cost on Amazon in late September. A buyer who waits and gets the path the drive makers describe pays $66 to $131 more for the same pair.

This guide sets out every dated forecast for hard drives and for flash, with who made each, when, and what it measures, because a contract-price forecast, a maker’s allocation statement and a price tracker’s guess get quoted as if they were one kind of thing. It then tests them against the forecasters’ record in this cycle, what is actually being built, why the makers are in no hurry, and seven past storage price cycles. The practical half runs the arithmetic of waiting on four real purchases, prices the failure risk of keeping old drives another year, gives decision rules for nine kinds of buyer, and covers the ways around paying peak prices, from recertified enterprise drives to cloud as a bridge. It is long because the question has four layers - the forecast, the price you would actually pay, your own timeline and the risk you carry while you wait - and most advice answers one of them.

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, and the text says which one is used and why. Where a figure reached print only through a relay - an analyst note quoted by a news site - the text says so. Everything is as of 2 October 2026. Micron’s fiscal fourth-quarter results of 30 September and TrendForce’s fourth-quarter memory forecast of the same day are in; Seagate’s and Western Digital’s September-quarter results, which on past cadence arrive in late October, are not. Prices on this site’s listing pages move every few hours and are never quoted in the text; where the guide needs today’s number, it tells you which page to read.

Buy now or wait: the answer depends on the medium and on when you need it

“Buy now or wait” is really two questions: which medium, and when you will need it. The forecasts disagree about 2028 for flash, agree about the next twelve months for both media, and contain no relief date at all for hard drives. So the decision comes first, before the evidence.

Only two statements in this guide meet the test for a consensus - three or more independent sources saying the same thing - and both meet it with room to spare. No maker or named analyst traced for this guide expects hard drive prices to fall before 2028: Seagate, Western Digital and Morgan Stanley all point the other way in documents dated between January and August 2026, and the most hopeful voice found, the US price tracker DatacenterDisk, says only that relief “is not expected before 2027” and that unnamed analysts “float a modest correction”. And no analyst or maker expects flash prices to fall in 2026: TrendForce, Gartner, IDC, Micron, Samsung, Sandisk, Kioxia and Citi all say so, in documents dated between February and September 2026. Beyond that the sources split, and the split is the useful part: the research firms put flash relief in the second half of 2027, the flash makers put it in 2028 or later, and one bank’s model has the deficit jumping to about 6 per cent in 2027 and staying there in 2028.

The two media run on different clocks for reasons that are physical, not fashionable. Hard drive supply grows when the makers put more terabytes on each platter, which costs them almost nothing in materials and needs no new factory; flash supply grows when a maker builds or converts a fab, which takes years. So a hard drive shortage is a decision and a flash shortage is a construction schedule, and the hard drive makers have said in plain words that they have decided not to build more drive units. Why the two cost structures differ is the first section of HDD or SSD; which one a given job needs is the rest of that guide.

The table assumes you know how much you need. Get that wrong and the timing stops mattering: buying 8 TB you will not fill until 2029 at 2026 prices is the most expensive mistake available this year, and buying 2 TB you will outgrow by spring is the second. How much storage you actually need works it out by use, from a phone backup to a hundred-terabyte office.

Hard drive and SSD prices on 2 October 2026, and why the indices disagree

Before forecasts, the level. Six published series cover retail storage prices in 2026, and they disagree by more than a factor of two, mostly because they start in different months and measure different things.

They are less contradictory than they look once the start date is stated. The German series begin before the rise, in July or September 2025, and capture all of it. The US matched-model index began on 23 March 2026, after most of the rise, and DatacenterDisk says itself that it cannot measure what happened before. eRacks follows the cheapest street price per family, which moves faster than a basket. And Nikkei’s figure is a wholesale reference for 1 TB drives sold in bulk, the kind that goes into surveillance recorders and cheap desktops, which is the corner of the market the makers are abandoning. Base month, region, product mix and new against used each move the answer by tens of percentage points, so the only honest comparison is the same drive, in the same shop, a year apart. ComputerBase publishes exactly that.

The changes are as ComputerBase prints them, worked from its unrounded prices, so one or two differ by a point from the rounded euros shown.

US prices for the same classes, where a dated figure exists:

Four things in those tables matter for timing, and none of them is the headline percentage.

The biggest hard drives rose most in German shops, and most of their discount per terabyte has gone. In ComputerBase’s basket the Toshiba 20 TB drive multiplied 2.71 times in a year and the 18 TB nearly tripled, against 2.07 times for the IronWolf 4 TB; 3DCenter’s index, which follows the cheapest offer at each size, does not show the pattern (1 TB +264 per cent, 22 TB +131 per cent since July 2025). Per terabyte the new CMR drives of 18 to 30 TB now cost €38.64 to €47.67, about what the 8 TB Red Plus costs (€43.63). Only the 24 TB BarraCuda, a shingled desktop drive that suits a backup copy and not a NAS, still undercuts the 4 TB IronWolf by almost a third, and the cheapest terabyte in the whole basket is a recertified one: the Exos 28 TB at €27.82. The old habit of buying the biggest new drive you can afford has nearly stopped paying in Germany; the discount on large drives that survives, there and in the US, is the recertified one.

The smallest hard drives stopped making sense. 3DCenter’s cheapest 1 TB hard drive cost €120 in September, up from €33 in July 2025, against €140 for the cheapest 1 TB PCIe 4.0 NVMe drive. When a hard drive costs nearly what an SSD of the same size costs, the reason to buy it is gone, and Nikkei reports that Seagate has begun phasing out some low-capacity models. Anyone shopping at 1 or 2 TB should be shopping for flash.

Recertified enterprise drives are the cheapest terabytes traced here in both regions, but the sources split on which new class rose least. eRacks found Seagate’s consumer BarraCuda line up 92% per terabyte in the US between March and August 2026 and the enterprise Exos SATA line up 15%, so that by August the enterprise drive was the cheaper new drive per terabyte, $27.92 against $31.25. ComputerBase’s German basket runs the other way: Toshiba’s enterprise MG drives rose 148 to 198 per cent in a year, against 100 to 121 per cent for the NAS drives of 4 to 16 TB, and the recertified Exos 28TB rose 103 per cent. And DatacenterDisk’s US index has new enterprise drives up 11.4 per cent over the 90 days to 2 October while NAS drives fell 0.8 per cent. On the level the sources agree: in the US table above, recertified enterprise drives sit between a quarter and a half below new NAS and enterprise drives, and in ComputerBase’s basket the recertified Exos is the cheapest terabyte of all. That is a point about who these drives are made for, and it has its own section below.

The two series that measure month to month have not turned down. ComputerBase’s hard drive figure went from +107% in June to +125% in July, +129% in August and +134% in September; its SSD figure from +112% to +118%, +126% and +129%. 3DCenter’s internal hard drive index rose another 7.3 per cent in September alone. Only the US matched-model index shows anything flat: -0.4 per cent over thirty days, with NAS drives -4.1 per cent and enterprise drives +0.7. DatacenterDisk’s SSD “floor”, which fell 14.3 per cent over the same thirty days, is not part of that index: it is the cheapest single listing, on 2 October a used drive, and the tracker itself says to read it “as what the cheapest listing did, not as what the market did”. One region flattening while another region’s basket keeps climbing is what a top looks like and also what a pause looks like; the section on signals says which data would tell them apart.

None of this is this site’s own data, by design: its listing pages rank every drive by price per terabyte, item plus shipping, and move every few hours, so sort the hard drive listings and the SSD listings for today’s median in the class you would actually buy.

Why storage prices doubled, in one page

The long version is the storage price hike explained. The short version takes five steps, and each of them bears on how long it lasts.

1. The 2022-23 bust made every maker cut. Kioxia and Micron each cut wafer starts by about 30 per cent, Samsung’s memory business posted its first loss in fourteen years, and Seagate’s revenue fell to $6.55 billion in the year to June 2024, below any year since at least fiscal 2005. Retail SSDs bottomed around June 2023.
2. A second round of cuts in 2025. After a short rebound in 2024, NAND prices fell again, and the five largest NAND makers reportedly cut output by 10-15 per cent in the first half of 2025 (Commercial Times, as relayed by TrendForce’s news desk on 17 July 2025). Neither they nor the three hard drive makers added unit capacity.
3. AI turned into a storage story in September 2025. TrendForce reported on 15 September 2025 that nearline hard drive lead times had gone from “a few weeks to over 52 weeks”, and that the makers had not expanded production capacity in recent years. Cloud providers who could not get disks bought high-capacity QLC enterprise SSDs instead, which is the reason TrendForce gave in its fourth-quarter NAND forecast of 25 September 2025.
4. Enterprise flash took the bits. Counterpoint’s figures, reported by Igor’s Lab on 13 August 2026, put enterprise SSDs at 48 per cent of all NAND bits shipped in the second quarter of 2026, against 26 per cent a year earlier. TrendForce expects enterprise SSD bit demand to grow by more than 80 per cent in 2026 (30 September).
5. The consumer became the residual customer. Seagate’s edge IoT business, which takes in its consumer drives, was about 21 per cent of its revenue in the December 2025 quarter; Western Digital’s consumer business was 5 per cent of revenue in the June 2026 quarter. Micron left the Crucial consumer business entirely, ending shipments at the end of February 2026; its chief business officer Sumit Sadana’s explanation in Micron’s announcement of 3 December 2025 was that “the AI-driven growth in the data center has led to a surge in demand for memory and storage”.

The point for a buyer is the last step. A retail shopper is buying what is left after long-term agreements with cloud providers have been filled, at a price set by scarcity rather than by cost. That is why retail hard drive prices doubled while the makers’ own contract prices rose by tens of per cent, as the next section shows, and it is why the end of the shortage, when it comes, may reach the shop shelf last. The same mechanism priced memory: the sister site’s DRAM shortage explainer covers the half of the story that runs through high-bandwidth memory for AI accelerators.

Contract, spot and retail: three prices on three clocks

Every forecast in this guide is about a price you will never pay. TrendForce’s quarterly figures are contract prices: what the flash makers charge PC makers, phone makers, cloud providers and the module makers who assemble consumer SSDs, negotiated a quarter at a time. Spot prices are what traders pay for wafers and chips outside those contracts, daily, in a thin market that moves first and furthest. Retail is the shelf, stocked with drives built from flash bought months earlier. Each layer moves on its own clock, and for storage in 2026 the gaps between them point in opposite directions for the two media, which is the single most useful thing to know about timing a purchase.

For SSDs, retail has risen about half as much as contracts

Compound TrendForce’s quarterly NAND contract forecasts, each in its latest published form, and set the result against what German shoppers paid over the same year:

NAND flash contract prices, TrendForce, latest forecast for each quarter
  4Q25  +20-25%  (3 Dec 2025, raised from +5-10% on 25 Sep 2025)
  1Q26  +85-90%  (3 Mar 2026, the third version)
  2Q26  +70-75%  (31 Mar 2026)
  3Q26  +10-15%  (3 Jul 2026)
  4Q26  +15-20%  (30 Sep 2026)

  calendar 2026      low   1.85 x 1.70 x 1.10 x 1.15        = 3.98
                     high  1.90 x 1.75 x 1.15 x 1.20        = 4.59
  3Q25 to 3Q26       low   1.20 x 1.85 x 1.70 x 1.10        = 4.15
                     high  1.25 x 1.90 x 1.75 x 1.15        = 4.78

Retail SSDs, Germany, September 2025 to September 2026
  ComputerBase basket     +129%                             = 2.29

  retail against contract, same year   2.29 / 4.15 = 0.55
                                       2.29 / 4.78 = 0.48

The compounding is mine, and it multiplies forecasts rather than settled transaction prices, because no realised quarterly NAND change from TrendForce turned up in its press releases for this guide. The direction of the error is known, though: TrendForce’s first forecast for the fourth quarter of 2025, 5-10 per cent, was overtaken within weeks. It reported on 1 December 2025 that November wafer contract prices alone had risen between 20 and more than 60 per cent in a single month, and on 3 December it raised the quarter to 20-25 per cent, the figure used above. Every revision in this cycle has gone upwards, so if anything the contract multiple is understated.

On that reading, a German SSD basket sits at roughly 48 to 55 per cent of where it would be had it followed contract prices one for one. Three explanations fit, and this guide cannot say which dominates. Shops are still selling drives built from flash bought at older prices. Many consumer SSDs use QLC flash, which puts less silicon in each terabyte than the blended contract figure assumes. And the suppliers chose to go easy on client SSDs: TrendForce wrote on 3 July that “elevated OEM inventories have significantly reduced buyers’ willingness to accept another round of price increases, pushing suppliers to adopt a more flexible pricing strategy for client SSDs”, and repeated the point for the fourth quarter on 30 September. Each of those explanations closes upwards, not downwards: old stock runs out, and flexibility lasts only as long as the suppliers choose. The one force that could close the gap the other way is consumer demand collapsing first, and the evidence that it is collapsing is in the section on what would change the outlook.

The same arithmetic run the other way gives the upper bracket for a single drive. A Samsung 990 EVO Plus 2TB that cost €122 in September 2025 would cost €506 to €583 had it followed contract prices to the third quarter of 2026, and €582 to €700 with the fourth quarter’s forecast added, against the €280 ComputerBase found on 17 September. Nobody forecasts that it gets there, but the gap is one reason no analyst expects it to fall soon.

For hard drives, retail has risen several times as much as contracts

The hard drive makers’ own prices rose by tens of per cent, not by multiples. Western Digital’s chief financial officer said on its 5 August results call that the company’s blended price per terabyte was up by “high teens” per cent year on year, from “high single digits” the quarter before. Seagate’s price per exabyte was up about 10 per cent year on year in the June quarter, with an analyst on its call floating something closer to 20 per cent ahead (Blocks & Files, 30 July). TrendForce reported on 14 October 2025 that the average hard drive price per gigabyte had risen from $0.012-0.013 to $0.015-0.016, roughly a quarter on midpoints (15 to 33 per cent across the ranges). Nikkei’s bulk reference price for 1 TB drives rose 15 per cent in the July-September quarter, the sixth straight rise.

Retail more than doubled. The cleanest single comparison is one drive:

Seagate BarraCuda 24TB, Germany (ComputerBase)
  Sep 2025    308 EUR / 24 TB  = 12.83 EUR per TB
  Sep 2026    804 EUR / 24 TB  = 33.50 EUR per TB
              at the ECB reference rate of 1.1298 (1 Oct 2026) = 37.85 USD per TB
              without 19% VAT   33.50 / 1.19 = 28.15 EUR per TB = 31.80 USD per TB

Backblaze's purchase price, Nov 2022             14 USD per TB
Nearline contract, Morgan Stanley, Jun 2026      below 15 USD per TB

  retail against contract, Sep 2026   31.80 / 15 = 2.1 times or more

A year ago the biggest consumer drive in Germany sold at about the price per terabyte a cloud storage company paid at the bottom of the last cycle. Now, with VAT taken out, it sells at more than twice what Morgan Stanley says the makers charge their largest customers, and the gap is a scarcity premium on the residual channel, not a cost.

So the two media carry opposite risks for a shopper. SSD retail is below its contract path and has room to rise before it has reason to fall. Hard drive retail is far above its contract path and has room to fall the moment scarcity at the margin eases - but no maker or named analyst expects that before 2028, and the makers can keep the margin scarce simply by continuing not to build units. The practical reading is that the hard drives priced closest to contract - recertified enterprise drives above all - are where the scarcity premium is smallest, which is why they keep coming up in the rest of this guide.

This cycle’s forecasts missed low, and one quarter’s was raised twice in eight weeks

Before weighing any date, check the forecasters’ record. In this cycle it is poor, and the errors run almost entirely one way.

How far the first-quarter 2026 forecast moved, on midpoints
  5 Jan    +35.5%
  2 Feb    +57.5%
  3 Mar    +87.5%       87.5 / 35.5 = 2.46 times the first forecast rise

The 33-38 per cent figure deserves a line of its own, because it is the most misdated number in this year’s coverage. It was TrendForce’s first forecast for the first quarter of 2026, published on 5 January; a good many buying guides quote it as the fourth-quarter 2025 figure, and some quote it as the final first-quarter figure. TrendForce’s fourth-quarter 2025 forecast was 5-10 per cent, raised to 20-25 per cent in December, and its final word on the first quarter of 2026 was 85-90 per cent. A guide that compounds the wrong quarters can still land near the right answer - four to four and a half times - by accident, which is how the error survives.

None of this makes TrendForce a bad forecaster. Quarterly contract forecasts are anchored on the last quarter, so they trail a turning market in both directions, and the record shows the same lag on the way down. In August 2022, as the last downturn gathered pace, TrendForce had to widen its third-quarter NAND forecast from a fall of 8-13 per cent to a fall of 13-18 per cent, and warned that “if manufacturers’ production capacity planning is not curtailed, this decline is likely to continue”. In 2022 the forecasts underestimated how fast prices would fall; in 2025-26, how fast they would rise. Either way the market got there first.

Two lessons follow for a buyer. A relief date is better read as the earliest it could happen than as a central estimate, because in this cycle every horizon on record that has moved has moved outwards, not inwards. And a forecast of deceleration is not a forecast of a fall: the July call that the rise was slowing was correct for one quarter and then reversed. When a forecaster you trust does finally publish a flat or negative quarter, that is news; until then the direction of revision is the more reliable signal than the number.

When will SSD prices go down? Every dated forecast through 2028

For flash, the forecasts come from three kinds of source with three kinds of interest: research firms, who sell their reports to buyers and sellers alike; the makers, who are negotiating long-term agreements while they speak; and banks, whose models feed investors. The table gives each one’s role.

Three things stand out once the table is read across rather than down.

Nobody on it expects flash to get cheaper in 2026. TrendForce’s own fourth quarter is a rise of 15-20 per cent at contract level. Gartner says up for four quarters. Even the most optimistic named voice, Gartner’s, puts stabilisation in the first quarter of 2027 and “a low level of decline” in the second half of 2027, and PC Games Hardware, summarising the analysts on 16 August, reports Gartner with Bernstein naming 2028 as the year of a possible real fall, so Gartner’s date reaches print in two versions.

The research firms date the turn by supply balance, and supply balance is not a price. TrendForce expects supply to overtake demand in the second half of 2027, which is the source of nearly every “relief in 2027” headline. It puts no figure on prices then: its release of 21 July contains no price forecast for 2027 at all, and its outlook of 30 July goes only as far as “increasing downward pressure on prices” once new capacity arrives in 2H27. Club386, reporting the 21 July release the same day, added the obvious caveat: retailers would still have to clear stock bought at higher prices. Contracts signed in 2026 carry floors, as the section on margins shows, so a positive balance in late 2027 is the start of a turn, not the bottom of one.

The makers all say 2027 stays short, and three say it gets worse. Micron (“much tighter”), Samsung (“even more severe in 2027 than 2026”) and Phison, whose chief executive told Bloomberg Television that 2027 may bring the worst NAND shortage yet, expect worse; Kioxia expects demand to exceed supply and Sandisk expects bits to stay on allocation beyond 2027. Citi’s model, built by a bank rather than a maker, agrees and puts a number on it: a deficit of 6.1 per cent in 2027 after 0.8 per cent in 2026, and 5.5 per cent in 2028. Only TrendForce, among the sources that publish a supply balance at all, has 2027 ending in surplus.

Gartner’s 2026 price figure for NAND is quoted in two versions in this year’s coverage: +202 per cent, attributed to a February research note behind its paywall, and +234 per cent, attributed to its press release of 26 February. They cannot both be the same forecast, neither could be checked, and neither is used here; the figure CRN Australia attributes to Gartner directly is a 130 per cent rise in combined DRAM and SSD prices by the end of 2026. The memory site treats Gartner’s DRAM numbers with the same caution in its buy-now-or-wait guide.

When will hard drive prices go down? Every dated forecast, and none sees a fall before 2028

No research firm publishes a public contract-price series for hard drives, so the hard drive forecasts are the makers’ own statements about allocation, one bank’s model, and a Japanese wholesale reference. The absence matters: there is no equivalent of TrendForce’s “balance turns positive in 2H27” for disks, and nothing to hang a relief date on.

Read together, the hard drive record says four things.

The makers have sold the next eighteen months and are selling the eighteen after that. Seagate went from “fully allocated through 2026” in January to “the vast majority” allocated into 2028 in July. Western Digital went from agreements reaching 2028 in January to talks reaching 2031 in August. In each case the horizon moved out by more than the calendar did.

Supply growth is a choice, and the choice is density. Morgan Stanley’s 30-35 per cent supply growth is not new factories; it is the makers putting more terabytes on each platter. Seagate’s phrase “not really increasing the box count” and Western Digital’s “does not require spending CapEx to add unit capacity” describe the same strategy. Each maker grows exabytes by shipping bigger drives to the customers who buy the biggest drives, which is the cloud.

The bank’s target price is still below today’s retail. Morgan Stanley’s $25-30 per terabyte is a contract price two to three years out. A new 8 TB IronWolf already cost $41 per terabyte on Amazon on 20 September, and a recertified 24 TB Ultrastar was listed at $27.46 on 2 October. So a contract price rising 70 to 100 per cent does not have to mean retail rising by as much; it can also mean the scarcity premium at retail narrowing as the two converge. That is the most hopeful reading of the hard drive data, and it still contains no fall.

A large, informed buyer chose to buy early. Backblaze, which runs more than 350,000 hard drives and publishes its purchase history, said in May that it was pulling part of its 2027 capital spending into 2026, and gave three reasons in this order: a large customer it has to serve in 2027, strong demand, and equipment costing 30 per cent more per unit than a year earlier. Two of the three are about its own growth rather than about timing the market, so this is not a bet against waiting; but the buyer with perhaps the best view of drive costs brought purchases forward at prices 30 per cent higher rather than holding off for cheaper ones.

The hard drive side does have one structural reason to ease that flash lacks: when the makers decide to grow, density lets them do it quickly and cheaply. No maker has said it is deciding that.

The research firms say the second half of 2027, the makers say 2028 or later

For flash the disagreement is real, and it is narrower than the headlines make it. On one side TrendForce, Gartner and IDC cluster on relief beginning in the second half of 2027 or at the end of it. On the other, Micron, Samsung, Sandisk, Kioxia and Phison say 2027 stays short, Micron and Samsung put the end in 2028 or later, Citi’s model has the deficit near 6 per cent in both 2027 and 2028, and SK Group’s chairman and SK hynix’s chief executive both say 2030. Four things explain most of the gap.

Who is speaking, and to whom. The makers are negotiating long-term agreements as they speak. Sandisk’s “New Business Model” agreements, with variable prices between floors and ceilings, are expected to cover over half of its bit sales in fiscal 2027 and about two-thirds in fiscal 2028 (The Motley Fool, via Yahoo Finance, 2 October); Samsung said it would put 60-70 per cent of capacity on multi-year agreements; Kioxia wants half of its shipments under such agreements by 2028; SK hynix has concluded them with about ten customers. A maker who says on a results call that next year will be tighter is also telling every customer who has not yet signed what the alternative to signing looks like. That does not make the statement false, but Micron’s outlook darkened between June and September, and a reader is entitled to notice that the makers’ horizons only ever move one way. The research firms sell their reports to buyers and sellers alike; their record, set out above, shows they lean low by habit rather than by interest.

Bits are not prices. TrendForce’s 2H27 date is a supply balance: the point at which bit supply overtakes bit demand after a 4-5 per cent deficit in 2026. A balance turning positive by a point or two starts a turn in contract prices; it does not undo a year in which those prices roughly quadrupled. Contracts signed in 2026 carry floors - Sandisk’s chief financial officer said on 5 August that “we expect attractive margins even at floor pricing” - and retail has to sell its higher-priced stock before it can reprice. The makers’ “tight through 2028” and the research firms’ “balance in 2H27” can both be true at once: a market that tips into small surplus in late 2027 at prices several times the 2025 level, with floors under the largest contracts.

Enterprise and consumer are separating. TrendForce’s fourth-quarter release has enterprise SSD as the only category accelerating, while client SSD suppliers adopt “a more flexible pricing stance” and module makers buy weakly. Consumer demand is already falling: IDC expects smartphone shipments down 16.7 per cent in 2026 (26 August) and PC shipments down 11.3 per cent (2 June); Sandisk’s consumer revenue fell 32 per cent in a quarter, to $556 million, in the three months to early July. If relief arrives anywhere before the makers’ dates, it arrives first in consumer SSDs, the corner of the market the makers care least about, and the research firms’ earlier date partly reflects that.

The whole argument turns on a few points of demand growth. Set the numbers side by side:

NAND bit growth, as published
                         2026            2027            2028
  Micron, industry       low-20s %       mid-20s %       mid-20s %     shipments (30 Sep)
  Citi, supply           -               +21%            +25%          (via Yahoo, 19 Sep)
  Citi, demand           -               +29%            +33%          (via Yahoo, 19 Sep)
  Kioxia, demand         high teens %    -               -             (31 Jul)
  TrendForce, balance    -4% to -5%      positive in 2H27              (21 Jul)

Citi’s supply figures are close to Micron’s shipment figures. The disagreement is almost entirely about demand: Citi has 2027 demand growing about eight points faster than supply, and TrendForce must have it growing slower for the balance to turn positive. One swing item is large enough to decide it. Citi estimated, as Gamers Nexus reported on 2 April, that Nvidia’s inference context-memory storage platform needs about 1,152 TB of extra flash per Vera Rubin system, roughly 9.3 per cent of global NAND demand in 2027 - more than the entire gap between the two camps. Whether a single new product category ships at the volume its maker hopes is worth more to the 2027 SSD price than every fab announcement in this guide.

For hard drives there is no equivalent split, because there is no research firm publishing a balance to disagree with. The makers, the one bank with a public model, and the largest buyer that publishes its purchase history all point the same way, and the one US tracker that offers a date says only “not before 2027”.

A cleanroom is not supply: what is being built, and when it counts

Every fab announcement this year has been read somewhere as the end of the shortage. The rule for reading them is the memory site’s: a cleanroom opening is not supply. After the shell come tools, then qualification, then a ramp measured in quarters. The milestones on the record that bear on flash supply:

Three points come out of the table. Most new flash bits before 2028 come from denser nodes in existing buildings, not from new buildings: Samsung’s V10, SK hynix’s 321-layer parts and Kioxia’s BiCS10. Those add bits quickly, but each maker controls the pace, and none has said it intends to outrun demand. The new NAND fabs from the established makers land in late 2028 and 2029 - Micron’s Singapore fab, SK hynix’s M17 and the Kioxia-Sandisk Fab3 - which is exactly where the makers put the end of the shortage. And the one large new fab landing sooner is YMTC’s, which matters less to Western buyers than its size suggests: YMTC has been on the US Entity List since December 2022, and Counterpoint, as reported by Igor’s Lab, found that although it shipped 14 per cent of the world’s NAND bits in the second quarter of 2026 - third behind Samsung at 25 per cent and SK hynix with Solidigm at 22 - it ranked only fifth by revenue, because much of its output goes into consumer and low-priced OEM products. That is the corner of the market a consumer SSD buyer shops in, so YMTC’s ramp is the most plausible early relief for cheap client drives specifically, and TrendForce’s forecast that Chinese suppliers’ share of NAND bit output will rise to nearly 19 per cent, in a passage covering 2026 and 2027 that does not say which year, is the number to watch.

Blocks & Files reported on 28 August that Kioxia and Sandisk both expect the AI-driven rise in NAND and SSD sales to last ten years or more, and that they plan more than ¥5 trillion of investment over six years to 2032, with about a third expected from subsidies. Makers planning for ten years of demand do not build to flood the market in year two.

Hard drives: no new factories, more terabytes per drive

The hard drive side has no construction table, because nobody is building a drive factory. Seagate and Western Digital each said in their latest calls that they are growing exabytes by putting more on each disk, not by making more drives. The density ramp is the supply forecast:

The pattern has a consequence that shows up in shop prices. Density growth arrives as bigger drives sold to cloud providers, and Seagate said on its April results call that it had deferred lower-capacity HAMR products for edge markets because “demand is so strong in the public cloud that we don’t have enough volume”. The consumer sizes, 1 TB to about 8 TB, are not where the new supply goes; Nikkei’s September report that Seagate is phasing out some low-capacity models is the same fact from the other end. A buyer waiting for relief at 4 or 8 TB is waiting for supply the makers have said they will not send that way. The component makers’ expansions are the one sign that unit volume may grow later in the decade, and HOYA’s 2028 date is about when it could start to show.

Record margins and no new drive factories: why the makers are in no hurry

A seller with these margins does not need to build anything to make money.

Set that against 2023. Micron lost $5.83 billion in fiscal 2023. Samsung’s memory business lost money for the first time in fourteen years in the first quarter of that year. Seagate’s revenue bottomed a year later, at $6.55 billion in the year to June 2024, below any year since at least fiscal 2005. Every executive now setting output plans lived through that, and every one of them is being rewarded for restraint. The flash makers cut output twice in 2025: reportedly by 10-15 per cent in the first half while prices were falling, and again in the second half while prices were already rising, when TrendForce’s news desk reported on 13 November 2025, citing Chosun Biz, that Samsung, SK hynix, Kioxia and Micron had “jointly scaled back NAND flash production”, SK hynix by about 10 per cent.

The hard drive makers have an even stronger reason to wait, because growth costs them so little. Western Digital’s capital spending in its June quarter was modest by the standards of any flash maker, and its chief executive was explicit that new demand “does not require spending CapEx to add unit capacity”. Seagate grows margin every time it moves its mix to more terabytes per disk. Neither has a reason to add a production line that would still be running when demand cools.

History shows what the industry did the last time it had this power. After the 2011 floods the two largest hard drive makers’ share rose from 62 per cent in the third quarter of 2011 to 85 per cent in the first quarter of 2012, after the Seagate-Samsung and Western Digital-HGST mergers (IHS iSuppli, via Computerworld, 6 June 2012), and gross margins went from about 20 to about 30 per cent (IEEE Spectrum, 1 November 2012). Prices stayed high long after the factories were dry. A buyer hoping for a price war needs a maker who wants one, and the only candidates on the record are the Chinese flash makers, whose output is aimed mostly at their home market and, in YMTC’s case, sits behind US export controls.

Seven storage price cycles, measured

Storage has been through enough shortages and gluts to say something about how they end. The record is uneven - there has never been a public hard drive contract price series, and the 2021 episode left no primary price data at all - but the shape of each cycle is clear enough.

What each cycle actually teaches

The 2011 floods are the cycle people cite and the one least like this one. A flood is a supply shock with a known end: the water recedes, the factories are cleaned, and Western Digital’s first post-flood drive came off the line 46 days after the flood, with normal capacity a couple of months later (IEEE Spectrum). Prices did not follow the factories. Backblaze, which bought drives right through the crisis, did not get back to its September 2011 price until September 2013, and its own reckoning was that the crisis would have cost it $1.14 million more than the pre-flood trend had it bought as before; shucking external drives, switching makers and “drive farming” through retail limits saved about $610,000 of that while it added about 50 petabytes. Two years to regain the old price, and never the old trend - that is the lesson of a supply shock in a concentrated industry, and the concentration rose during the crisis, as the margins section showed.

The 2016-17 shortage and the 2018-19 crash are one cycle, and its end came from technology, not buildings. Flash got tight in 2016 because the move to 3D NAND cut output while yields were poor; it got loose in 2018 because the same technology, once yields improved, added bits faster than demand grew. TrendForce forecast prices down 25-30 per cent across 2019. The modern parallel is the denser nodes in the build table above: Samsung’s V10, SK hynix’s 321-layer parts and Kioxia’s BiCS10 are this cycle’s equivalent of the 64-layer ramp, and if anything ends the shortage early it is them rather than a new fab.

Chia is the template for a retail demand spike, and this is not one. A speculative crowd bought large drives in a hurry in the spring of 2021, the coin’s value collapsed, and the demand went with it. A shortage driven by consumers ends when consumers stop. This one is driven by cloud providers on multi-year contracts, which is why it has lasted eighteen months already.

The 2022-23 crash is the reason this shortage was so sharp. The makers cut deeply to stop it, Micron by nearly 30 per cent of wafer starts and Kioxia by about 30 per cent, and then cut again in the first half of 2025 after the short 2024 rebound faded. They entered the AI demand surge with no spare capacity and no inventory, and they remember what overbuilding cost them.

The 2024 rebound is the warning about reading one good quarter. Contract prices rose 15-20 per cent in the first quarter of 2024 because suppliers pushed, not because demand pulled, and the rise faded within the year. The last NAND price trough was the first quarter of 2025, eighteen months before this guide was written. TrendForce expected prices to stabilise in the second quarter of 2025, and every quarterly NAND forecast it has published since September 2025 has been a rise, the fourth-quarter 2026 one included.

What the record says about this cycle, and what is different this time

The history cuts both ways, which is the honest point. Flash shortages have ended in crashes before, and steep ones: the 2018-19 fall and the 2022-23 fall each took prices down by double digits a quarter. Anyone who says storage never gets cheaper again has not read 2019. But every crash in the table needed one of two things - capacity that overshot demand, or demand that collapsed - and the evidence on both is different this time.

Demand is contracted for years, by buyers who are not price-sensitive. TrendForce wrote on 31 March that cloud providers “are willing to accept higher prices and sign LTAs to secure a stable supply”. In 2016-17 the marginal buyer was a phone or notebook maker buying quarter to quarter; in 2026 it is a cloud provider buying to 2028 and beyond, with hard drive agreements reaching 2031. Demand of that kind does not vanish in a quarter. It can be cut, by a decision to spend less on AI, and that is the bust case below - but it has to be decided first, and it would be announced.

Capacity is not overshooting, and the makers are telling you so. The technology steps that ended the 2016-17 shortage are under way again, but the makers are pacing them, and the new fabs land in late 2028 and 2029. On the hard drive side the unit count is flat by policy. An overshoot needs somebody to build ahead of demand, and every maker on the record has said it will not.

The consumer is the residual customer. In earlier cycles consumers were much of the demand; now consumer sales are 5 per cent of Western Digital’s revenue and a share of the 21 per cent Seagate earns outside the data centre, and they are the corner of the flash market where the makers are being “flexible”. So consumer prices are the last to benefit when contract scarcity eases, and the first to soften if consumer demand alone breaks.

Put together, the record gives a buyer three working assumptions. A turn, if it comes when TrendForce expects, starts in the second half of 2027 for flash and later for hard drives. After a turn, the old price takes years to come back, if it comes back at all: twenty-four months after 2011. For memory, PC Games Hardware’s RAM buying advice of 16 August attributes to TrendForce and Bernstein the view that prices will stay 30 to 50 per cent above 2024’s average even in the long run; no comparable figure was found for storage. And the scenario that would bring prices down fast is a bust, not a cure, which would come with a public cause - cloud providers cutting AI spending - that a buyer can watch for rather than guess.

What waiting costs: the arithmetic on four real purchases

Four purchases run through this section, each priced from a dated third-party figure, and each tested against three paths per medium. It is a method, not a prediction. Every path is anchored to something on the record, and none of them has a probability attached, because no source gives one and this guide does not invent them.

For hard drives:

• The makers’ path: price per terabyte up 10 to 20 per cent over twelve months, the range that Seagate’s and Western Digital’s own year-on-year figures span.
• Flat: contract prices keep rising but the retail scarcity premium narrows to meet them, the hopeful reading of Morgan Stanley’s target price.
• The fast fall: four consecutive quarters of 6 per cent declines, which IHS iSuppli described in June 2012 as what it would take to get back to pre-flood prices, and which it did not expect to happen. Compounded, 0.94 x 0.94 x 0.94 x 0.94 = 0.781, a fall of about 22 per cent in a year. No forecaster expects it now; it is the best case with a precedent.

For SSDs:

• Catch-up: retail follows TrendForce’s fourth-quarter contract forecast of 15-20 per cent by the end of 2026.
• Flat through 2027: the makers’ view, with prices holding where they are.
• A 2019 repeat: a fall of 25-30 per cent, the size TrendForce forecast for 2019, starting when TrendForce expects supply to loosen in the second half of 2027 and arriving by the end of that year at the earliest. It is the milder reading of 2019: Micron’s own average flash price per gigabit fell 47 per cent in its fiscal year to August 2019, as the Micron chart above shows.

A 2 TB NVMe drive

Samsung 990 EVO Plus 2TB, 280 EUR (ComputerBase, 17 Sep 2026)

  catch-up, by end-2026         280 x 1.15 = 322    280 x 1.20 = 336    +42 to +56 EUR
  flat through 2027             280                                     0
  2019 repeat, by end-2027      280 x 0.75 = 210    280 x 0.70 = 196    -70 to -84 EUR

  months from October 2026 to December 2027 = 15
  saving per month of waiting   70 / 15 = 4.67      84 / 15 = 5.60 EUR

  back to the Sep 2025 price of 122 EUR     (280 - 122) / 280 = a 56% fall
  a 2 TB cloud plan for those 15 months     9.99 x 15 = 149.85 EUR

On the only path that produces a saving, a buyer who can do without the drive for fifteen months saves about €4.70 to €5.60 a month. That is the whole case for waiting on an SSD, and it is a real one for somebody whose current drive works: a gamer with room to spare, a laptop that is not full. It is not a case for waiting to get back to last year’s price, which needs a fall of 56 per cent that no forecast supports. And it collapses if the need is real, because the catch-up path costs €42 to €56 within three months. The cloud line is not a substitute for a local NVMe drive; it is there to show what two terabytes of storage costs per month when somebody else buys the drives, which is about twice what waiting saves.

Two 8 TB drives for a NAS mirror

2 x Seagate IronWolf 8TB, 2 x 328 = 656 USD (Amazon prices via MakeUseOf, 20 Sep 2026)

  makers' path, 12 months      656 x 1.10 = 722    656 x 1.20 = 787    +66 to +131 USD
  flat                         656                                      0
  fast fall                    656 x 0.781 = 512                        -144 USD
  saving per month             144 / 12 = 12.00 USD

  the same 8 TB rented for the year
    Backblaze B2               8 x 6.95 x 12 = 667 USD
    Wasabi                     8 x 7.99 x 12 = 767 USD

  recertified instead, today
    2 x Ultrastar HC550 18TB   2 x 489 = 978 USD for 36 TB     27.17 USD per TB
    against the IronWolf       1 - 27.17 / 41.00 = 34% less per TB

Even the best case with a precedent saves $144, and only for a buyer who has somewhere else to keep the data for the year. If that somewhere is rented, the rent costs more than the drives. If it is the old drives, the next section prices that risk. The recertified line answers the question differently: it buys more than twice the capacity for about half as much again, at a price per terabyte 34 per cent below the new NAS drive today, which is a bigger discount than the fast fall would deliver a year from now.

A four-bay NAS, 80 TB raw

80 TB raw, US prices
  new, enterprise-class median    80 x 46.04 = 3,683 USD  (DatacenterDisk, 2 Oct 2026)
  new, NAS-class median           80 x 55.00 = 4,400 USD  (DatacenterDisk, 2 Oct 2026)
  recertified, 4 x HC580 24TB     4 x 659    = 2,636 USD for 96 TB  (ServerPartDeals, 2 Oct 2026)

  waiting 12 months on the new enterprise price
    makers' path   3,683 x 1.10 = 4,052    3,683 x 1.20 = 4,420    +368 to +737 USD
    fast fall      3,683 x 0.781 = 2,876                           -807 USD

  per TB a year from now, after the fast fall    46.04 x 0.781 = 35.96 USD
  per TB today, recertified                                      27.46 USD
  recertified discount on new enterprise         1 - 27.46 / 46.04 = 40%

  half now, half later
    2 x HC580 24TB now, mirrored, 24 TB usable   2 x 659 = 1,318 USD

The recertified discount today is nearly twice the best-case fall a year from now: 40 per cent against 22. A buyer who waited a year and got the fall the 2012 analysts said it would take to undo a flood would still pay more per terabyte than a buyer of recertified drives today, and would have run without the array for a year to do it. The half-now line is the other way round the problem, and the section on buying smaller covers when a layout allows it.

About 100 TB of pool for a small business

About 100 TB of pool: eight 18 TB drives in RAIDZ2, six of them carrying data
  8 x Ultrastar HC550 18TB, recertified   8 x 489 = 3,912 USD
     usable  6 x 18 TB = 108 TB = 108e12 / 2^40 = 98.2 TiB
     at the 80 per cent planning line   98.2 x 0.80 = 78.6 TiB of data
  8 x 18 TB new, enterprise median        144 x 46.04 = 6,630 USD
  a second, backup copy                   doubles either figure

  renting 100 TB for a year while waiting
    Backblaze B2    100 x 6.95 x 12 = 8,340 USD
    Wasabi          100 x 7.99 x 12 = 9,588 USD

  eight drives at Backblaze's Q2 2026 fleet failure rate of 1.73%
    chance of at least one failure in a year   1 - (1 - 0.0173)^8 = 13.0%

For a business, “wait” means one of two things: rent the capacity, or keep running on what you have. A year of renting costs more than the new array and more than twice the recertified one. Keeping the old array means carrying its failure rate, and the chance that at least one of eight drives fails in a year is about one in eight even at the failure rate of a well-run data centre’s fleet, before any allowance for age. The second copy is not optional at this scale: RAID is not a backup, and the NAS guide’s rebuild arithmetic shows why a wide array of large drives needs one.

The rule the four purchases share

Waiting pays when the price you expect to save is larger than what waiting costs you: the rent on a bridge, the risk carried on old drives, or the value of doing without. For hard drives, no named forecast puts the expected saving above zero in the next year, and the best case with a precedent puts the fall at 22 per cent, less than the recertified discount available now. For SSDs the saving is about €5 a month on a 2 TB drive, and only for a buyer who can wait until the end of 2027. Wait only if you can go without the capacity until the end of 2027 or later, nothing you own is at risk, and going without is worth less to you than about five euros a month per 2 TB of flash. For hard drives there is no waiting horizon with a forecast behind it.

Keeping old drives another year has a failure rate, and it can be worked out

The usual alternative to buying is to keep what you have for another year. That has a price too, and Backblaze publishes the data to estimate it.

Backblaze’s revisit of the failure-by-age curve, published on 15 October 2025, found modern drives far better than their predecessors: barely above 1.30 per cent a year in their first year, then “a pretty even failure rate through the significant majority of the drives’ lives”, and a peak of 4.25 per cent at ten years and three months. Its 2013 curve had peaked at 13.73 per cent at three years and three months, and its 2021 curve at 14.24 per cent at seven years and nine months. Its latest quarterly report, for the second quarter of 2026, published on 29 September, put the fleet’s annualised failure rate at 1.73 per cent across 354,415 drives, “the highest it’s been in quite a while”, with ten of 31 models above 3 per cent, three of them outliers above 7; it retired two HGST models during the quarter at nearly eight and nearly nine years old.

The failure rate of one drive is small. The failure rate of an array is not, because any one of its drives can fail:

chance of at least one failure in a year = 1 - (1 - AFR)^n

  4 drives at 3%       1 - 0.97^4 = 1 - 0.885 = 11.5%
  8 drives at 4.25%    1 - 0.9575^8 = 29.4%

The columns run from a young drive’s rate to the peak of Backblaze’s 2025 curve. A four-drive array of drives approaching ten years old, at the 4.25 per cent peak, has about a one-in-six chance of losing a drive in the next year; eight such drives, close to one in three. Those are fleet averages from air-conditioned data centres, and a fleet failure rate says little about one particular drive, as the enterprise-drives guide explains at length; a home server in a warm cupboard is not a Backblaze pod. The table is an order of magnitude, not a promise, and if anything an optimistic one: the arithmetic treats each drive as independent, while drives of one batch bought on one day share their faults, and the NAS guide cites the field study showing that failures arrive in clusters.

What a failure costs depends on the layout. A single drive with no backup turns a failure into lost data, which has no price on any listing page. A mirror or RAIDZ array turns it into a rebuild, during which the array runs with less protection: a 20 TB drive resilvering at about 100 MB/s takes around 55 hours, and the NAS guide works through what that exposure means at current capacities. Buying one replacement at peak prices is cheap next to that; the expensive outcome is a second failure during the rebuild.

Three conditions make keeping old drives a sound decision rather than a gamble: a current backup of anything you cannot replace; SMART data with nothing moving in the reallocated, pending or uncorrectable counts; and drives young enough to sit on the flat part of the curve, which on Backblaze’s 2025 data means under about seven years. If a drive is already reporting pending or reallocated sectors, or the array has no redundancy left, the decision is no longer about the market. Replace that drive now, at whatever it costs, and keep the rest.

SSDs fail differently. Wear is readable: the NVMe health log reports a percentage used, and most consumer drives retire their warranty long before their flash, as the SSD endurance guide shows from the ratings. A working SSD is rarely a reason to buy a new one; capacity is.

Running out of space costs more than the drive

The third cost of waiting is the one people pay without noticing. When storage fills, something gets deleted, and what gets deleted first is usually the protection.

• Snapshots and old versions go first, because they are the easiest space to reclaim, and they are the thing that would have saved you from ransomware or an accidental delete.
• The backup copy goes next. A backup drive that has filled up stops taking new backups, often silently.
• Recording gets downgraded. Camera footage at a lower bitrate, photos exported smaller, a surveillance retention window cut from a month to a week.
• Pools run too full. A filesystem near capacity slows down, and the NAS guide’s advice is not to plan past 80 per cent: near full, ZFS changes how it allocates (at 90 per cent in TrueNAS’s documentation, at about 96 in a TrueNAS blog post of 14 August 2026), and a resilver needs room it no longer has. A consumer SSD near full loses most of its fast cache, which is why it feels slower than it did new.
• Games and libraries get uninstalled and downloaded again later, which costs time and, on a capped connection, money.
• For a business, the cost is downtime or an emergency purchase at whatever price the day offers, which in this market is the worst price available.

Having no backup is the one situation where waiting is never right. A single external drive for a second copy of irreplaceable files costs a fraction of what a lost photo library or a client archive costs, and the cheapest terabytes on any listing page are good enough for it. How much storage you actually need covers measuring how fast you are filling what you have, which turns “running out” from a surprise into a date.

Nine buyers, and what decides each one

Each case gives the evidence that decides it and the one thing that would change the answer.

A failed drive, a failing drive or a degraded array. Buy now, the capacity you need to restore redundancy, at the cheapest class you can vet. Nothing in this guide argues for running an array with no margin while the market sorts itself out; the failure arithmetic above says the next failure is the expensive one. What would change the answer: nothing.

Irreplaceable data with no second copy. Buy now, and buy the backup drive before anything else. A single large hard drive or external drive is enough, and per terabyte the cheapest new drive in ComputerBase’s September check was the 24 TB BarraCuda at €33.50, a shingled drive that is fine for a backup copy and wrong for a NAS, as CMR vs SMR explains. What would change the answer: an off-site cloud copy you already pay for and trust.

A planned NAS build. Do not wait for 2028. Size the layout for three years, buy the first half of the drives now - a mirror of two large drives, recertified where you can vet them - and add the second half when you need it. The four-bay case above shows why recertified drives beat waiting. What would change the answer: a NAS whose drives would sit mostly empty for more than a year, in which case buy less.

An optional homelab expansion. This is the one hard drive case where waiting is a fair choice, because nothing is at risk. Set an end date and a target price per terabyte, watch the enterprise class on the listing pages, and buy recertified or used enterprise drives when the median crosses your number. What would change the answer: running out of room, which turns it into the NAS case.

A gamer who wants a bigger SSD. If the current drive works and has room, waiting to the end of 2027 is defensible and worth about €5 a month on a 2 TB drive. If it is full, buy a Gen4 drive rather than Gen5, or move the game library to a large SATA SSD or a hard drive and keep the fast drive for what is being played. What would change the answer: a console or laptop that takes only one kind of drive, in which case buy what fits when you need it.

A video editor or creator. Buy the working SSD capacity your projects need now, because a full scratch drive costs working hours, and put finished projects on hard drives, which are three to five times cheaper per terabyte even at 2026 prices (€33.50 to €47.67 per terabyte for drives of 18 TB and up against €140 to €170 for 2 TB NVMe drives in ComputerBase’s September check). What would change the answer: a project pipeline that is genuinely idle until next year.

A small business with growing data. Size the system for three years of growth and buy the drives for the next twelve to eighteen months, not five years, and lock any quote you get. Recertified enterprise drives with a multi-year warranty are a defensible business purchase, and the four-bay and 100 TB cases show they beat waiting on every path traced here. Tier the data: flash only for what people wait on. What would change the answer: a refresh you can defer without running unsupported or unprotected.

An enterprise buyer replacing or adding arrays. The constraint is allocation before price. “No meaningful open production remains for discretionary buyers except the hyperscalers,” Tekonyx’s Sid Nag told The Register in February, and Omdia’s Vlad Galabov expected “a very tough year for the standard enterprise general purpose server and for enterprise storage”. The server makers have passed the costs on: StorageSwiss reported on 6 May that Dell, Lenovo, HP and HPE raised server prices by about 15 per cent in late 2025 and early 2026, and that Dell raised hardware prices by a further 17 per cent on 30 March. TrendForce put nearline lead times at more than 52 weeks as early as September 2025. So order against the lead time rather than the budget cycle, take the longest quote validity offered, and before counting the recertified discount, check whether the array vendor’s support terms allow recertified or third-party drives at all. What would change the answer: allocation horizons shortening in the makers’ October results.

An archivist or data hoarder. Hard drives, recertified where possible, and never flash for cold data. Size the system for three years and buy drives for twelve to eighteen months at a time: the eRacks founder’s line from September, “the old habit was to buy extra capacity while it was cheap. It is not cheap, and it is not getting cheaper, so that habit now costs you”, is the right correction to a habit formed when prices only fell. At very large scale, LTO tape is the medium built for this, as a later section notes. What would change the answer: a turn in the hard drive data, which no forecast expects before 2028.

Buy fewer drives now and add more later, if the layout lets you

The habit formed over fifteen years of falling prices was to buy the biggest drive you could afford, because the price per terabyte fell with size and next year’s drive would be cheaper anyway. Less of that holds than it did. As the price section showed, the largest drives rose most in ComputerBase’s basket, and per terabyte only the shingled BarraCuda is still clearly cheaper new; the large CMR drives cost about what an 8 TB drive does, and the clear discount on large drives is the recertified one. The other half - that next year’s drive will be cheaper - is the half the forecasts have removed.

So the useful version of “buy smaller now” is buy fewer drives now, not smaller ones: the large drives you mean to keep, recertified where the discount is, but only as many as the next twelve to eighteen months need, in a layout that lets you add the rest later. The layout itself - bays, vdev width, parity - should be sized for the three years of measured growth that how much storage you actually need works out; only the drives are bought in stages. Where the layout cannot grow, buy its final width at once. The arithmetic on the four-bay case:

4 x HC580 24TB recertified, all now                      2,636 USD

2 now as a mirror, 2 a year later
  now                         2 x 659 = 1,318
  later, makers' path         1,318 x 1.10 = 1,450    1,318 x 1.20 = 1,582
  later, fast fall            1,318 x 0.781 = 1,029

  total, makers' path         2,768 to 2,900          +5% to +10% on all-now
  total, fast fall            2,347                    -11% on all-now

Splitting the purchase costs 5 to 10 per cent more on the makers’ path and saves about 11 per cent on the best case with a precedent, and it keeps half the money in your pocket for a year. Drives bought a year apart also come from different batches, which the NAS guide recommends as a cheap hedge against clustered failures, while noting that nobody has measured how much it buys.

Splitting the purchase works only if the layout allows it, and the decision is permanent in some layouts:

• Mirrors grow in pairs. A ZFS pool of mirrors takes another mirror vdev at any time; two drives now and two later is the natural shape.
• RAIDZ can now grow a drive at a time, since OpenZFS 2.3 added raidz expansion, with two caveats the NAS guide spells out: the parity level does not change, and data written before the expansion keeps its old ratio of data to parity until it is rewritten.
• Synology SHR and Unraid were built for adding drives, including drives of different sizes, which is the main reason to choose either for a growing home server.
• Traditional RAID5 and RAID6 can be reshaped, slowly, with a rebuild-like pass and the risk that goes with it.

What does not work is the other reading of the phrase: buying small drives now to replace them with big ones later. The NAS guide’s capacity-planning section prices that plan in risk rather than money - one full rebuild per drive replaced, eighteen days of degraded or resilvering operation to swap eight 20 TB drives - and in this market it also loses money, because small drives cost the most per terabyte (€48.75 to €50 for the 4 TB NAS drives in ComputerBase’s basket) and resell for the least. Bays are the scarce resource; fill them with the drives you will keep.

For flash the split is simpler. Most desktop boards have a second M.2 slot, so a 1 or 2 TB NVMe drive now and another later costs nothing in layout; most thin laptops have one slot, so buy the size the laptop will need for its life. Which slot takes which drive is in the drive interfaces guide.

Recertified enterprise hard drives are the discount that survived

A manufacturer-recertified drive has been returned to its maker or the maker’s agent, tested, restored to factory specification and resold with a warranty. ServerPartDeals, a large US seller of them, describes its stock as “returned to the manufacturer, restored to factory spec, and recertified by the OEM”, with warranties of two or three years that vary by model. On eBay a manufacturer-recertified drive is usually listed under one of the eBay Refurbished grades - Excellent, Very good or Good - refurbished, which only vetted sellers may use and which carry eBay’s one-year warranty - or simply as Used; “Certified refurbished” is reserved for the brand or its authorised sellers. The condition column on this site’s listing pages shows which, and buying used drives on eBay sets out what each code promises.

The discount, from dated figures:

The discount is a matter of level, not of trend, and the level is the same in both regions: recertified enterprise drives were the cheapest terabytes in the US comparisons traced here and in ComputerBase’s German basket. It is the one large discount traced for this guide that does not depend on a sale. Why it exists is less settled than that it does: eRacks’ US figures had the enterprise Exos SATA line rising far less than the consumer BarraCuda between March and August, but the German and the latest US data, set out in the price section, run the other way.

It is not guaranteed to last. Recertified drives rose with everything else: ComputerBase’s recertified Exos 28TB went from €384 to €779 in a year, a rise of 103 per cent, and How-To Geek reported on 30 October 2025 that recertified drives had already become dearer, as the overall shortage, resellers holding stock and a better reputation for recertified drives all pushed up prices. DatacenterDisk’s US index has new enterprise drives rising faster than NAS drives over the last quarter, so the gap between the classes is narrowing from the enterprise side. And data centres keeping drives in service longer means fewer come back: Backblaze’s latest report retired drives at nearly eight and nearly nine years old. A supply that depends on what data centres throw away is thinner when data centres throw away less.

The risks are specific, and each has a check:

1. Used drives sold as new or as recertified. heise reported on 23 January 2025 that two German online retailers, one on Seagate’s dealer list, had sold used Exos drives (ST14000NM005G and ST16000NM001G) as new, with SMART values reset to near zero; Seagate’s FARM log showed 10,000 to 22,000 hours of use, and Seagate confirmed that SMART can be reset and that FARM was likely correct. On any Seagate drive, compare SMART power-on hours with the FARM log; buying used drives on eBay gives the commands and the second check, spindle hours against head flying hours, that needs no reference value.
2. Whose warranty it is. A recertified drive’s warranty usually comes from the seller, not the maker. Read the terms and the seller’s history before the price.
3. SATA or SAS. Many enterprise drives are SAS and need a SAS controller; a SAS drive does not plug into a SATA port. Drive interfaces covers the connectors.
4. The 3.3 V pin. Many enterprise SATA drives treat one power pin as a power-disable signal, and a power supply that still delivers 3.3 V on it keeps the drive from spinning up. The fix is an adapter or a different cable, and the enterprise-drives guide explains it.
5. Sector size and OEM firmware. SAS pulls from storage arrays can arrive formatted with 520-byte sectors and need reformatting; drives with a server maker’s firmware work but carry no maker warranty. Both are covered in the enterprise-drives guide.
6. Noise, heat and power. Enterprise drives are louder and warmer than NAS drives. In a living room that matters; in a cupboard it usually does not.
7. Prove it on arrival. Run the long self-test and a full write-and-verify pass inside the return window, before any data goes on it.

On this site’s pages the enterprise class and the SAS listings are where used and recertified enterprise drives concentrate, and the condition on each row tells you which you are looking at. Check the recording mode too: a few high-capacity families are shingled, and CMR vs SMR says which.

Other ways around the peak, and one that is closing

External drives, and shucking them. In the 2011 crisis Backblaze bought external drives from retailers and took them out of their cases, because externals were cheaper than the same drives sold bare. The same gap opened in 2026: 3DCenter’s March index found comparable internal drives “in most cases actually more expensive than their external counterparts”, with external drives up only 20 per cent since September 2025 while internal hard drives were up 59 per cent since July. That gap is closing. By September 3DCenter’s external index was up 97 per cent, with 7.9 per cent of that in September alone. The drive inside an external case can be a lower grade or a shingled model, opening the case usually ends the warranty, and some shucked drives need the same 3.3 V fix as enterprise drives; buying used drives on eBay covers what is inside. The simplest use of a cheap external drive needs no screwdriver at all: as the backup copy. External drives are on the USB listings.

An older flash generation. A Gen4 NVMe drive instead of Gen5, or a SATA SSD for bulk flash. Gamers Nexus found the average 2 TB SATA SSD up 76.3 per cent between November 2025 and March 2026 against 113.7 per cent for 2 TB NVMe drives, so the older interface rose less. For a game library, a photo archive on flash or a second drive, sequential speed is rarely what you notice.

QLC and DRAM-less drives. They put less silicon, or no cache chip, in each terabyte and cost less for it; for a game library or a second drive the endurance is ample, as SSD endurance works out from the ratings.

Used enterprise SSDs. XDA Developers found on 21 March 2026 used 4 TB SAS enterprise SSDs selling for $170 to $180, about half the $350 a new 4 TB consumer SATA SSD cost, with up to three times the rated writes of a consumer drive. They need a SAS port, as drive interfaces explains, and the vetting in buying used drives on eBay applies to them as much as to disks.

Hard drives for bulk. Even after doubling, a large hard drive is three to five times cheaper per terabyte than an NVMe drive, and anything that is stored rather than worked on belongs there. HDD or SSD sets out where each still wins.

A small SSD in front of a large hard drive. Keep the working set on flash and the bulk on disk; NVMe caching covers the ways to do it and the ones that lose data.

Tape, at archive scale. LTO-10 cartridges hold up to 100 TB compressed and read at up to 1,200 MB/s, according to the LTO Program’s roadmap. Tape drives are expensive and slow to access, so this is for archives of tens of terabytes and up that are written once and rarely read, not for a home server.

Pre-populated NAS units and prebuilt systems. Retailers have argued that a NAS sold with drives can undercut buying the parts separately while old stock lasts (Newegg’s own guide, July 2026, a seller’s view). The same logic applied to prebuilt PCs on the memory side, as the memory site’s buy-now-or-wait guide documents. Check what drives are inside before counting the saving.

Cloud as a bridge got dearer too

Renting storage while waiting for drives to get cheaper is the obvious bridge, and the cloud providers who buy the drives have passed the shortage on.

Wasabi gave its reason in plain words: “increased costs for storage hardware, energy, and data center operations driven by growing demand and supply constraints”. Backblaze announced its rise on 17 March, and its finance chief told investors in May that its equipment cost 30 per cent more per unit than a year earlier. The object-storage prices that a home server or a small business would use for a bridge have risen. The consumer plans that phones back up to still charge $9.99 or €9.99 a month for 2 TB, and for phone photos they remain the sensible answer.

As a bridge for a NAS’s worth of data, renting loses. Eight terabytes in B2 for a year costs $667, more than two new 8 TB drives at September’s Amazon price; 100 TB costs $8,340, more than a new array. Cloud makes sense for the copy you needed anyway - an off-site backup - and for a few hundred gigabytes of overflow for a few months. It does not make sense as a way to wait out a shortage measured in years.

Timing a purchase: the reporting calendar, Black Friday, and a number you write down

If you are buying in the next year anyway, the question is not whether but when, and three tools help.

The reporting calendar. The figures that move this market arrive on a rhythm a buyer can follow:

A buyer does not need to read all of it. The one release that would change this guide’s answer is the first TrendForce quarterly forecast with a flat or negative NAND range - every one published since September 2025 has been a rise

• and the one statement that would change it for hard drives is Seagate or Western Digital saying that allocation horizons are shortening rather than lengthening.

The sales. Black Friday 2026 falls on 27 November, the day after US Thanksgiving. Buying guides routinely promise 20-35 per cent off SSDs on Black Friday; none of the ones surveyed for this guide cites a dataset for it. The record from last year points the other way: Black Friday 2025 fell in the middle of the surge, and the 2 TB NVMe drives Gamers Nexus priced in November 2025 at $145 to $190 cost $300 to $400 by March 2026. A sale in a rising market is a discount from a price that is about to be higher, which is worth taking if you were buying anyway and worth nothing as a reason to wait. Amazon’s October Prime event, which deal sites were already trailing at the start of the month, comes first, and the same rule applies to it.

A number you write down. The most useful thing to wait for is a price, not a date. Pick the class you would actually buy - 8 TB NAS drives, recertified 18 to 24 TB enterprise drives, 2 TB Gen4 NVMe - note today’s median price per terabyte for that class on this site’s listing pages, and decide in advance the price at which you buy. Then check it weekly. For SSDs one third-party anchor helps: MonsterDealz’s deal guide of 17 August 2026 calls anything under €100 per terabyte a good price this year and anything under €80 a very good one, against a median of €104.50 across the deals it had posted in 2026. A dip is a chance to make a purchase you were going to make anyway, not the start of a trend. 3DCenter’s internal hard drive index fell 1.9 per cent in March 2026 and then climbed to +157 per cent by September, and DatacenterDisk warns that its own SSD floor, which fell 14.3 per cent in the thirty days to 2 October, is a single listing, on that day a used drive. One month of falling is not a turn.

What would change the outlook, and what is not known

Every forecast in this guide could be wrong, and the record says the likelier error is that prices stay higher for longer. But there are specific events that would move the answer in either direction, and most of them would be public before they reached a shop.

Two of those deserve a line each. Tariffs: the Section 232 proclamation of 14 January 2026 put a 25 per cent duty on a narrow class of advanced computing chips from 15 January, with exemptions that include consumer uses in the US, and left room for what EY’s summary calls “significant broader tariffs”. A second phase extending duties to laptops, consoles and data-centre servers was reported by Tech Times on 31 August with no rates, exemptions or dates finalised, and nothing traced for this guide says whether drives or SSDs would be covered. It is a reason to watch, not a reason to buy early. A bust is the scenario that would bring prices down fast, and no source attaches a date or probability to it.

What is not known, and would change this guide:

• There is no public hard drive contract price index. The hard drive half of this guide rests on maker statements, one bank’s model and a Japanese wholesale reference.
• No realised quarterly NAND change from TrendForce was found in its press releases, so the contract compounding here multiplies forecasts, the most recent of them published on 30 September.
• The retail indices disagree by region: German baskets still rising in September, the US matched-model index -0.4 per cent over thirty days, with NAS drives -4.1 per cent and enterprise drives +0.7.
• Gartner’s 2026 NAND figure appears as +202% and +234% in different coverage, and its documents were not available to check.
• The makers’ forecasts are self-interested and the research firms’ have lagged, so neither camp’s date is a central estimate.
• Seagate’s and Western Digital’s September-quarter results, due on past cadence in late October, will update the allocation horizons that carry the hard drive half of this guide.

The short answers, with where each is argued

What to do with this on a listing page

1. Decide your horizon and the medium first. Need it within a year: buy. An optional SSD upgrade with no deadline: wait, with an end date. An optional hard drive purchase: buy recertified or buy fewer, rather than waiting for a fall nobody forecasts.
2. Confirm the capacity before the price, with how much storage you actually need: size the layout for three years of measured growth, and buy the drives the next twelve to eighteen months need.
3. Start from the ranking, not from a brand. The hard drive listings and SSD listings rank every row by price per terabyte, item plus shipping; the methodology page says how it is worked out and where it stops being reliable.
4. Compare fixed prices. Bid-only rows are already left out, because a bid is not a price; ?buy=now drops every auction for a clean comparison.
5. Look at the enterprise class before the consumer one. The enterprise hard drives and SAS listings are where the scarcity premium is smallest; read the condition on each row and vet a used or recertified drive with the checks above.
6. Pick one capacity point and track its median. The cap filter selects one exact capacity, such as 8 TB drives or 2 TB SSDs. Note the median on a date, write down the price at which you buy, and check it weekly.
7. Price the alternatives on the same measure. External drives are on the USB listings, NVMe drives on the NVMe listings, and every drive sits on one ranking; a large hard drive for bulk and a smaller SSD for speed is usually cheaper than flash for both.
8. Buy on a dip only if you were buying anyway. One month of a falling floor is not a turn, and in this cycle the dips have reversed within months.
9. Re-read this after late October, when Seagate’s and Western Digital’s results update the allocation horizons, and again in January, when TrendForce’s first-quarter 2027 forecast should land.

Everything above is as of 2 October 2026. This cycle’s forecasts have mostly run low and its relief dates have mostly moved outwards, so the guide will be revised as new figures arrive rather than defended. Until they do, the short version stands: if you need storage, buy it no later than you need it, in drives you will keep and a layout you can grow, recertified where you can vet them; if what you have is enough, the only wait with a forecast behind it is a wait for flash, it runs to the end of 2027 at least, and it pays about five euros a month.