Hard drive and SSD prices in 2026: why they doubled and when they ease

Hard drives and SSDs cost over twice their September 2025 prices because AI data centres booked years of output from makers that had stopped expanding, and no forecast sees relief before late 2027 for SSDs or 2028 for hard drives.

The makers say so in their own words. Seagate told investors on 28 July 2026 that “the vast majority of our nearline exabytes are now allocated into calendar 2028”. Western Digital said on 5 August that it was negotiating long-term agreements “for calendar year 2029, 2030, and 2031”. Sandisk’s finance chief, Luis Visoso, said the same day that “we therefore expect bits to remain on allocation beyond calendar year 2027”. The most optimistic dated forecasts put the turn in the second half of 2027, when TrendForce wrote on 21 July 2026 that the flash market’s supply-demand balance “is expected to turn positive”; Micron’s chief executive said on 30 September that memory and storage supply would be “much tighter” in 2027 and 2028 than in 2026, three months after pointing to a gradual improvement in 2028. None of the forecasts collected here dates a return to 2023 or 2025 prices, and the largest flash makers have written price floors into multi-year contracts.

“AI bought all the drives” is the usual explanation, and it is incomplete in three ways that matter to anyone trying to judge where prices go next. First, two different industries were caught in one loop: the hard drive shortage sent cloud buyers to high-capacity flash, which emptied the flash market, consumer SSD prices more than doubled, and buyers who could no longer afford an SSD went back to hard drives, until even a 1 TB desktop drive set a record wholesale price in Japan. Second, the shortage was built in the bust of 2022 and 2023, when both industries lost money and cut, and the flash makers cut again in 2025 while prices were already recovering. Third, “the price” is several prices. The hard drive makers’ own average price per terabyte rose by roughly a tenth to a fifth year on year in the June quarter, on their own statements, and the pace was quickening; German shop prices for the most popular hard drives rose 134 per cent in the year to September. The difference between those numbers is the subject of most of this article.

It follows the money through four layers. The factory: who makes the platters and the flash, the bust that left no slack, the second round of cuts, and a dated timeline of the turn. The demand: why AI inference and agents need storage, why the hard drive makers sold out rather than build, why new recording technology adds terabytes rather than drives, and why the flash makers gave their cleanrooms to DRAM. The contract: the quarterly flash series, the makers’ own statements and the long-term agreements. And the shelf: retail as the residual, three countries, used drives, cloud storage, tariffs and the euro. Then margins and collusion, new supply, every dated forecast beside the forecaster’s interest, what would shorten or lengthen the shortage, twenty-one short answers and what to do about it. Whether to buy now is the subject of a sibling guide, buy storage now or wait, and how much to buy of how much storage you really need.

One note on sources, because it changes how the numbers should be read. Where a figure is a maker’s, an analyst’s or a retailer’s, the document is named in the text with its date and linked where it is public; where a figure is worked out here, the arithmetic is shown so that it can be checked; where sources disagree, the text says so and says which one it uses. TrendForce publishes the only public quarterly series of flash contract prices, so the article leans on it and shows how often it revised its forecasts. Makers, and the companies that buy from them, are interested parties and are labelled so. Several “TrendForce” items are its news desk relaying Korean, Japanese or Taiwanese newspapers, and the original outlet is named where that is so. Prices on this site’s listing pages move every few hours, so the text names the page to read rather than freezing a site price into prose. The knowledge here is dated 2 October 2026: Micron’s results of 30 September are in it, Seagate’s and Western Digital’s September-quarter results, due in late October, are not, and anything dated after a source’s own date is called a forecast.

September 2026 brought a split, not a turn. 3DCenter’s index of the cheapest German retail offers, published on 14 September, had internal SSDs up 0.5 per cent on the month and internal hard drives up 7.3 per cent, and ComputerBase’s check of 17 September found hard drives 134 per cent and SSDs 129 per cent dearer than a year earlier, both new highs. On 30 September TrendForce forecast flash contract prices up another 15 to 20 per cent in the fourth quarter, faster than the 10 to 15 per cent it had forecast for the third, with enterprise SSDs “the only memory product category to see a larger increase than in the previous quarter”.

Hard drives and SSDs got expensive at once, through one loop

A hard drive and an SSD share almost nothing in manufacture: platters, heads and a motor from three companies on one side, NAND flash dies, a controller and often a DRAM cache from half a dozen on the other. Why one still costs several times less per terabyte is the subject of HDD or SSD. What they share is the buyer.

The hard drive side is the smaller cast. The Elec, relayed by TrendForce’s news desk on 22 July 2026, put Western Digital and Seagate each above 40 per cent of the hard drive market and Toshiba at about 17 per cent; the same item, citing Nikkei via MoneyDJ, said US hyperscalers, the largest cloud operators, take roughly 60 per cent of nearline drive shipments. Nearline means the high-capacity 3.5-inch drives that sit in data-centre racks, the family the enterprise drives guide describes. Western Digital said on 5 August that cloud customers were 89 per cent of its revenue.

The flash side has more makers and the same customers. TrendForce’s ranking of 18 August 2026 for the second quarter, which covers the five largest publicly listed brands:

TrendForce gives Samsung’s and Kioxia’s shares of the whole market, which put that market at about $78.7 billion (23.06 / 0.293); the other shares are worked from that total. The ranking covers publicly listed brands only, which is why China’s YMTC, which is not listed, is absent from it. By bits the order differs: Counterpoint’s figures for the same quarter, reported on 12 August, put Samsung at 25 per cent of NAND bits shipped, SK hynix with Solidigm at 22 per cent and YMTC third at 14 per cent, though only fifth by revenue in Counterpoint’s ranking, behind Micron and Kioxia, because it sells mostly into consumer and low-price channels. The same release carries the number that explains most of this article: enterprise SSDs took 48 per cent of all NAND bits shipped in the second quarter of 2026, against 26 per cent a year earlier.

The two industries met in the data centre, and the shortage ran round a loop. In order, each step from a dated source:

1. Hard drives ran out first. TrendForce reported on 15 September 2025 that nearline lead times “have ballooned from just a few weeks to over 52 weeks”, and that some North American cloud providers were evaluating SSDs for cold data, the bulk storage that hard drives had always held.
2. The overflow went to flash. On 25 September TrendForce forecast flash contract prices up 5 to 10 per cent for the fourth quarter, naming QLC enterprise SSDs bought in place of unavailable hard drives as the cause; by 29 October some SSD models faced delivery delays of more than a year and drives of 8 TB and more were booked through the second half of 2026.
3. Flash ran out next. TrendForce wrote on 3 March 2026 that severe hard drive shortages and extended lead times “accelerated a shift in orders toward NAND solutions”, and on 25 May that they had redirected orders to QLC enterprise SSDs.
4. Consumer SSDs more than doubled. German shop prices for the most popular SSDs rose 129 per cent in a year (ComputerBase), and Nikkei reported PC SSD prices up 3.5-fold.
5. Buyers went back to hard drives. The same Nikkei report, relayed on 9 September 2026, said expensive SSDs had pushed buyers to hard drives, and bulk 3.5-inch 1 TB drives reached a record $67.70 a unit, the sixth straight quarterly rise. Western Digital’s consumer revenue rose 38 per cent, helped by price, which it partly credited to “alternative products that are flash-based”.
6. Enterprise buyers moved back too. Western Digital reported “increased demand from enterprise OEM players in the storage space, especially as they are pivoting to more hybrid-based storage solutions”, disk and flash together.

Each industry’s shortage became the other’s demand. That is why the two price series rose together after a decade of moving apart, and why relief for one does not mean relief for the other: high-capacity flash replacing nearline disk would ease hard drives and tighten flash. Most accounts of the shortage cover one half; read separately, neither half makes sense.

How much hard drive and SSD prices rose depends on which price you read

Every figure quoted for this shortage belongs to one of five layers, and they moved at different times by very different amounts. Contract prices are what the makers charge PC makers, server makers and cloud providers, negotiated quarterly or under multi-year agreements, and they carry most of the volume. Component spot is broker trading in flash wafers and chips. Wholesale is distribution, bulk drives sold to system builders and shops. Retail is a new drive on a shelf. Used and recertified is the layer nobody publishes a series for.

The rows are not contradictory once their baselines and baskets are read. The US index in the last row began in March 2026, after most of the rise, and its publisher says it cannot measure September 2025 to March 2026. ComputerBase counts from mid-September 2025 and prices twelve popular models; 3DCenter, whose index is in the regional section, counts from July 2025 and prices the cheapest offer at each capacity, which is why one says 134 per cent and the other 157. Measured from September 2025 on its own index, 3DCenter’s figure is about 150 per cent (257 / 103 = 2.50), so most of the gap is method rather than baseline.

Two figures in wide circulation are misattributed. The first is “hard drive prices up 46 per cent”. It is ComputerBase’s basket of twelve popular drives in German shops, measured from 15 September 2025 to 14 January 2026, in euros; Tom’s Hardware, which spread it in English on 16 January, checked US listings and found the trend replicated “or perhaps worse”, but the 46 per cent itself is not a US figure, and several pages give it the wrong years. The second is “flash up 33 to 38 per cent in the fourth quarter of 2025”. TrendForce’s first forecast for that quarter, published on 25 September 2025, was 5 to 10 per cent, raised to 20 to 25 per cent on 3 December; 33 to 38 per cent was its first forecast for the first quarter of 2026, made on 5 January 2026 and revised upward twice within two months, as the section on contract prices shows.

The retail rise per terabyte, worked from ComputerBase’s own table of 17 September 2026, shows a pattern the averages hide:

drive                          15 Sep 2025   16 Sep 2026   per TB, before -> after   multiple
Seagate IronWolf 4TB           EUR 94        EUR 195       23.5 -> 48.8              2.07
WD Red Plus 8TB                EUR 170       EUR 349       21.3 -> 43.6              2.05
Seagate BarraCuda 24TB         EUR 308       EUR 804       12.8 -> 33.5              2.61
Toshiba Cloud-Scale 18TB       EUR 288       EUR 858       16.0 -> 47.7              2.98
Kingston NV3 1TB (SSD)         EUR 50        EUR 144       50   -> 144               2.88
Samsung 990 EVO Plus 2TB (SSD) EUR 122       EUR 280       61   -> 140               2.30

In ComputerBase’s basket the biggest hard drives rose most, and the smallest SSDs rose most. The 18 TB to 24 TB drives in it are built on the same platforms as the nearline drives the cloud operators buy, and several are nearline models themselves, so they compete with the front of the queue. The BarraCuda 24TB cost 55 per cent of the IronWolf 4TB’s price per terabyte a year ago and costs 69 per cent of it now; the Toshiba 18TB went from 68 per cent to 98 per cent, so the discount for buying big has narrowed sharply, and for some enterprise models has all but gone. The cheapest offers tell a different story: 3DCenter’s index, which prices the cheapest drive at each capacity, shows no such pattern by size, and its hard drive rise in September came mostly from 1 TB drives, up 40 per cent in a month, while its larger hard drives rose by 6 per cent at most and its 8 TB SSD by 24. ComputerBase gives no reason for the SSD pattern, and its basket does not support the obvious one, that brands without their own flash rose most: Samsung’s 990 EVO Plus 1TB rose 148 per cent and Kioxia’s Exceria Plus G3 2TB 140 per cent, while Lexar’s NM790, built on bought-in flash, rose 138 per cent at 2 TB and 106 at 4 TB. What the basket does show is size: the two 1 TB drives rose most, at 189 and 148 per cent, so the cheapest way into an SSD lost the most ground. PC Games Hardware found the same in August, with 1 TB SSDs at EUR 145 to 180 per terabyte against EUR 111 to 149 for 4 TB models.

A percentage without its baseline month and its layer is not information. The listing pages on this site add a sixth vantage point, drives listed at the moment you look, ranked by price per terabyte: the hard drive listings and SSD listings correspond to the retail rows above, and how this site works out price per terabyte says what the number includes.

The shortage was built in the bust of 2022 and 2023

Accounts of this shortage start with AI. Its precondition was a glut.

Flash went first. Kioxia cut wafer starts by about 30 per cent at its Yokkaichi and Kitakami fabs from October 2022, citing weak demand, in a notice dated 30 September 2022. Micron reduced its wafer starts by nearly 30 per cent in both DRAM and NAND, it said on 28 June 2023, and its fiscal 2023 revenue halved, from $30.76 billion to $15.54 billion, with a net loss of $5.83 billion. Samsung held out longest. On 7 April 2023 it announced a “meaningful” cut to memory production as its first-quarter operating profit fell about 96 per cent, its lowest since 2009, and its memory business recorded its first loss in 14 years; in July it extended the cut. Even after the cuts, TrendForce expected flash prices to fall again in the third quarter of 2023, by 3 to 8 per cent, with client SSDs down 8 to 13 per cent.

The hard drive makers had the same year. Seagate’s fiscal 2023 revenue was $7.38 billion, down 37 per cent and its lowest in 17 years, and its chief executive Dave Mosley spoke of “a profound downturn in demand” (Blocks and Files, 27 July 2023). In the September 2023 quarter Seagate shipped 89.6 exabytes, 24 per cent fewer than a year earlier, and lost $184 million. Its nearline shipments fell 34 per cent, to 56 exabytes. The reason Mosley gave that October reads strangely now:

Spending priorities for CSPs have temporarily shifted toward AI-related infrastructure, which have further slowed the pace of demand recovery for mass capacity storage.

In 2023 AI took money away from storage. The cloud operators were buying accelerators instead of disks, and the drive makers sized their plans to that.

The size of what came next shows in Seagate’s own shipments:

Seagate nearline shipments
  Jul-Sep 2023 quarter    56 EB    (down 34% on a year earlier)
  Apr-Jun 2026 quarter   195 EB    (up 43% on a year earlier)
  195 / 56 = 3.5   -> three and a half times in eleven quarters,
                      with the number of drives "not really increasing"
                      (Seagate's chief executive, April 2026)

Retail shoppers saw the bottom in mid-2023: a 1 TB NVMe SSD cost about $50 around June, on Cloudzat’s price history, though trackers disagree on the depth of the trough. Backblaze’s hard drive purchase records, transactions rather than surveys, had fallen to $0.014 per gigabyte in November 2022, and Backblaze predicted a penny a gigabyte by mid-2025, “a 22TB drive for about $220”.

Two habits carried into the recovery. The flash makers had learned that cutting output restores price, and the hard drive makers had learned that the cloud operators could stop buying for a year. Neither industry added unit capacity for the next two years. The 2025 rally started from prices the makers had lost money at, with no spare capacity anywhere and every incentive not to build any.

A false dawn in 2024, and a second round of cuts in 2025

The first recovery was supplier-led. TrendForce forecast flash contract prices up 15 to 20 per cent for the first quarter of 2024, which it credited to supplier-led pricing: the makers’ cuts had worked, while demand had not returned. By June it expected the third-quarter rise to be curbed to a modest 5 to 10 per cent as suppliers ramped up production into lacklustre consumer demand, and it later dated the next decline from that quarter: flash prices “have been in decline since 3Q24”, it wrote in January 2025. On 31 December 2024 it forecast the first quarter of 2025 down 10 to 15 per cent, client SSDs down 13 to 18.

The makers answered with a second round of cuts. TrendForce reported on 22 January 2025 that Micron, Kioxia with its partner, Samsung and SK hynix with Solidigm would cut 2025 output by lowering utilisation and delaying node upgrades, and in July its news desk, relaying Commercial Times, reported that the top five makers had cut production by 10 to 15 per cent in the first half of 2025, Micron and Sandisk from the second half of 2024.

Prices turned within a quarter. For the second quarter of 2025 TrendForce forecast client SSDs up 3 to 8 per cent and wafers up 10 to 15 per cent, and named five causes on 26 March 2025: the cuts since the fourth quarter of 2024, three quarters of inventory depletion, US tariff pull-ins by consumer brands, the end of Windows 10 support, and what it called the “DeepSeek effect”, which it said was accelerating the adoption of edge AI. For the third quarter it forecast contract prices up 5 to 10 per cent and wafers up 8 to 13 per cent, and Commercial Times, in the same news-desk item, expected the shortage to stretch possibly into 2026. By 25 September TrendForce said the first-half cuts and inventory clearance had restored the market’s balance, and that capacity allocated to high-margin products was creating a “price support layer”.

The flash makers then cut again in the second half of 2025, while prices were already rising. Chosun Biz, relayed by TrendForce’s news desk on 13 November 2025, reported that SK hynix had cut its NAND output by about 10 per cent in the second half, that Micron was holding output at its main Singapore fab steady, and that Samsung and Kioxia had also scaled back, in a week when TrendForce measured spot prices for 512Gb TLC wafers up 17 per cent. Those cuts mattered more than their size, because they were made weeks before the demand arrived.

Two structural changes date from the same period. On 24 February 2025 Western Digital completed the spin-off of its flash business as Sandisk, leaving a hard-drive-only Western Digital. And Seagate’s first heat-assisted drives, Mozaic 3+ at 28 and 30 TB, launched in January 2024, began the shift of hard drive supply growth from more drives to more terabytes per drive. The hard drive makers, TrendForce said in September 2025, had not expanded production capacity in recent years.

The last flash price trough was the first quarter of 2025, eighteen months before this guide’s date. Contract prices have risen in every quarter since on TrendForce’s figures, six in a row to the third quarter of 2026, the last of them a forecast, with a seventh forecast for the fourth.

The turn began with hard drives, in September 2025

The first public sign was a lead time. TrendForce’s release of 15 September 2025 said that “lead times for nearline HDDs have ballooned from just a few weeks to over 52 weeks”, that the makers had not invested in production expansion and so could not answer the jump in cloud demand, and that some North American cloud providers were therefore evaluating SSDs for cold data. It added two numbers that would shape the next year: high-capacity QLC SSDs draw about 30 per cent less power than nearline hard drives, and their shipments could see “explosive growth” in 2026.

A month later TrendForce put a price on it. Its release of 14 October 2025 said the average hard drive price per gigabyte had risen from $0.012-0.013 to $0.015-0.016, as the makers passed the cost of the transition to heat-assisted recording on to customers, and that flash makers were fast-tracking nearline SSDs of 122 TB and 245 TB to take the business the drive makers could not serve.

average hard drive price per GB, TrendForce, October 2025
  before    0.012 to 0.013   midpoint 0.0125
  after     0.015 to 0.016   midpoint 0.0155
  0.0155 / 0.0125 = 1.24   -> about +24% at contract, at the very start

Lead times did not look the same from every seat. The distributor Fusion Worldwide put enterprise hard drive lead times at three to six months, as How-To Geek reported on 30 October 2025, and noted that the supply of recertified drives was thinning because fewer drives were coming back from data centres. TrendForce’s year counts hyperscale nearline orders; a distributor’s quarter counts the drives available to everyone else. Both are true at once, and the gap between them is the allocation this article keeps returning to.

Contract prices followed. DigiTimes reported on 15 December 2025 that hard drive prices had posted their largest rise in eight quarters on demand from Chinese and US cloud providers; TweakTown’s account of the same report gave a contract rise of about 4 per cent on the quarter, with 3.5-inch 1 TB desktop and surveillance drives at about $53 and 2.5-inch 1 TB drives at about $50.

Retail moved faster than contracts, because retail has no contract. ComputerBase tracks the twelve most-viewed hard drives on its price comparison, excluding eBay and Amazon Marketplace sellers so that scalpers do not distort the figure. Between 15 September 2025 and 14 January 2026 the basket rose 46 per cent on average, with individual drives up between 23 and 66 per cent; as Club386 relayed the figures, a Seagate BarraCuda 24TB went from EUR 307.90 to EUR 499.90 and a Toshiba MG10F 22TB from EUR 336.47 to EUR 558.99.

Tom’s Hardware checked the US on 16 January 2026 and found the trend replicated “or perhaps worse”: the BarraCuda 24TB, seen as low as $239 on sale, was $499 on Amazon. In Akihabara, Tom’s Hardware had reported on 8 November 2025, shops were already rationing drives after distributors suspended some deliveries.

Then the makers confirmed it on their calls. Seagate said on 27 January 2026 that “our nearline capacity is fully allocated through calendar year 2026, and we expect to begin accepting orders for the first half of calendar year 2027”, on fiscal second-quarter revenue of $2.83 billion at a 42.2 per cent non-GAAP gross margin; 2027 volumes had been agreed with some customers, the transcript records, but not yet the price. Western Digital’s chief executive Irving Tan was blunter two days later:

We’re pretty much sold out for calendar ’26. We have firm POs with our top seven customers.

Western Digital also had long-term agreements with two customers for 2027 and one for 2028, covering both exabytes and price, on revenue of $3.02 billion, up 25 per cent. The Register on 20 February quoted Sid Nag of the research firm Tekonyx: “no meaningful open production remains for discretionary buyers except the hyperscalers”.

By January 2026 the hard drive market had stopped clearing on price: the cloud operators had the year’s output on order, and everyone else bid for what was left. The flash market reached the same state about a quarter later.

SSD prices followed within weeks, and shops felt both by Christmas

The flash turn was faster, because flash is traded on spot markets as well as under contract, and the spot market moves daily.

New Daily reported on 29 October 2025, as relayed by TrendForce’s news desk, that Samsung, SK hynix and Kioxia were running their SSD lines at full capacity, with some models facing delivery delays of more than a year and drives of 8 TB and up booked through the second half of 2026. By November the controller maker Phison, whose chips sit inside many budget SSDs, was describing the flash it buys. Its chief executive Pua Khein-Seng said a 1 Tb TLC die had gone from $4.80 in July to $10.70 in November (DigiTimes, 10 November 2025), that the imbalance “will likely persist for several years”, and that new production lines would not run until late 2027. A month earlier he had told CommonWealth Magazine: “NAND will face severe shortages next year. I think supply will be tight for the next ten years.”

1 Tb TLC die, Phison's figures via DigiTimes
  July 2025       $4.80
  November 2025  $10.70
  10.70 / 4.80 = 2.23   -> more than doubled in four months

TrendForce’s measurement for November 2025 is the clearest single data point. On 1 December it reported flash wafer contract prices up between 20 per cent and more than 60 per cent in one month, with 512Gb TLC up more than 65 per cent, and its news desk had reported spot 512Gb TLC wafers up 17 per cent in a single week in mid-November. The fourth-quarter forecast of 5 to 10 per cent, made in September, was overtaken inside the quarter it described. TrendForce raised it to 20 to 25 per cent on 3 December, with a month of the quarter left.

On 3 December 2025 Micron announced that it would leave the Crucial consumer business, shipping through the consumer channel until the end of February 2026, “to improve supply and support for our larger, strategic customers in faster-growing segments”, in the words of its chief business officer Sumit Sadana. The Crucial P310 is one of the mainstream 2 TB drives in the US comparison below, so the exit took a familiar budget brand, and the flash that fed it, out of retail. Warranty service continues.

TrendForce’s first forecast for the first quarter of 2026, on 5 January, was a rise of 33 to 38 per cent in flash contract prices and more than 40 per cent for client SSDs, and the same release carried a sentence, about DRAM, that described the new order of things: “consumer buyers are prepared to pay more to secure priority access”. That forecast, as the section on contract prices shows, was far too low.

Shops showed it first in Europe. The Austrian site futurezone reported on 17 January 2026, from Geizhals data, that a Samsung 990 EVO Plus 1TB had gone from about EUR 80 to EUR 166, and that ComputerBase’s SSD basket had risen 74 per cent in four months. Gamers Nexus compiled US prices for the most popular 2 TB drives on 2 April 2026, from Newegg’s price histories:

Gamers Nexus gives the average for the four NVMe drives as +113.7 per cent and for 2 TB SATA drives as +76.3 per cent. The Fanxiang row matters later: a drive built on Chinese flash doubled like the rest, so Chinese supply did not act as a brake at retail.

SSDs rose faster than hard drives in the winter, and hard drives faster than SSDs from the spring. In the US comparison above the hard drive rose about a quarter while the SSDs doubled; in Germany by September 2026 the hard drive basket had overtaken the SSD basket. The reasons for the crossover are in the section on regional prices.

From glut to shortage in four years, in one table

One row per event, each from a named document. Read it as three phases: the bust and the cuts (2022 to 2024), the second round of cuts and the turn (2025), and record contract quarters, allocation and the first supply announcements (2026).

Two of those rows are often misreported. The 5 January forecast of 33 to 38 per cent is the one that turns up as “the fourth quarter of 2025”. And the Seagate row of 28 July is about nearline exabytes, the cloud product, not the desktop and NAS drives a shopper buys, though those come off the same lines.

AI eats storage at every step: training lakes, inference, agents and the KV cache

The first wave of AI spending bought accelerators and, as Seagate found in 2023, took money from storage. The second needed storage at every step, because AI moved from training models to serving them to millions of people, many of them through agents that act on their own.

Training is the familiar part: a corpus has to be kept, cleaned and versioned, and the cheapest place for bytes nobody waits on is nearline disk. Western Digital’s chief executive added a newer source on 5 August 2026: models for robots and vehicles, which he called physical AI, train on synthetic data sets that are generated and then have to be stored.

Inference is the part that changed the arithmetic. When a language model serves a long conversation it keeps intermediate results for every token so far, the key-value cache, rather than recomputing them; in Seagate’s words on its July call, the “key value, or KV cache, is used to retain and reuse that context efficiently”, and it grows as users multiply and conversations lengthen. Keeping it in accelerator memory is the most expensive option, so the industry built a storage tier for it. On 5 January 2026 Nvidia announced its BlueField-4-based Inference Context Memory Storage platform, claiming up to five times the tokens per second and five times the power efficiency, available in the second half of 2026, with Dell, HPE, Pure Storage, VAST Data and WEKA among the companies building on it. Citi then estimated the flash it needs: 1,152 TB of additional NAND for each Vera Rubin system, which, assuming 30,000 systems shipped in 2026 and 100,000 in 2027, would be about 2.8 per cent of global NAND demand in 2026 and 9.3 per cent in 2027, as Gamers Nexus reported Citi’s note on 2 April 2026. It is a bank’s estimate for a platform still shipping, so it is a forecast. In drives:

Citi's estimate: 1,152 TB of extra NAND per Vera Rubin system
  1,152 / 2  = 576    -> the flash in 576 consumer 2 TB SSDs, per system
  1,152 / 30 = 38.4   -> or about 38 enterprise SSDs of 30 TB
the whole fleet, on Citi's shipment assumptions
  2026:   30,000 systems x 1,152 TB = 34.6 million TB  -> 2.8% of NAND demand
  2027:  100,000 systems x 1,152 TB = 115.2 million TB -> 9.3% of NAND demand

Agents compound it. Western Digital’s chief executive put the claim on the record on 5 August: “Agents generate data at every step of a workflow, increasing both the volume of data created and the amount that must be stored over time.” TrendForce’s release of 30 September 2026 gave the flash version: the large-scale rollout of agentic AI “is driving exponential growth in data volumes for real-time retrieval and caching, while QLC penetration in vector databases continues to rise”, and it expected enterprise SSD bit demand to grow by more than 80 per cent in 2026.

The demand shows in published shares. Servers account for more than 40 per cent of all flash bit demand in 2026, TrendForce wrote on 21 July, with phones and notebooks together “nearly 40%”. The cloud operators’ capital spending will rise 98 per cent in 2026 and another 50 per cent in 2027, with DRAM and flash taking 47 per cent of it in 2026 and 68 per cent in 2027, on TrendForce’s estimate of 25 August. And Seagate’s chief executive pointed in April to his customers’ order books: the top three cloud providers “have nearly doubled their RPO to staggering $1.1 trillion”, remaining performance obligations being contracted revenue not yet delivered.

Inference made storage scale with users, not with models. A model is trained once and stored once; a cache of context is built for every conversation and every agent run, and the industry decided to keep it on flash. That is the change that turned an accelerator boom into a storage shortage.

The high-capacity flash that holds those caches also competes with hard drives for cold data. Kioxia’s management said on 31 July 2026: “The QLC NAND SSD for data center to replace nearline hard disk. That’s an opportunity.” Samsung said on 30 July that its QLC bit shipments in the second half would more than double those of the first and that it now offers 256 TB server SSDs, and SK hynix said that its Solidigm unit’s revenue from enterprise SSDs of 30 TB and above had more than tripled on the quarter. None of them says flash is cheaper than disk per terabyte; it is not, by about eighteen times at 30 TB on one index of August 2026, as HDD or SSD shows. They say that when disk cannot be had, flash is bought instead, at whatever it costs.

One caution belongs here. The demand narratives come from the makers, who sell drives and flash, and from Nvidia, which sells the platform; Seagate’s own line in July, “we believe storage demand will prove durable through investment cycles”, is a belief, not a measurement, and the 80 per cent growth in enterprise SSD bits is a forecast for a year not yet finished. What is measured is the shift already recorded: enterprise SSDs went from 26 to 48 per cent of all flash bits shipped in a year, on Counterpoint’s figures. Micron expects industry bit shipments to grow in the low 20s per cent in 2026, and on that assumption the arithmetic for everyone else is short:

share of flash bits going to anything but enterprise SSDs (Counterpoint)
  2Q25  100 - 26 = 74%        2Q26  100 - 48 = 52%
total bits up 21 to 24% in the year (Micron's "low 20s", taken as that range)
  0.52 x 1.21 / 0.74 = 0.85     0.52 x 1.24 / 0.74 = 0.87
  -> about 13 to 15% fewer bits for phones, PCs, retail SSDs and everything else
     (rough: second-quarter shares set against a full year's growth)

The rest of the market got less flash, not more, in a year when the industry made more of it than ever.

The hard drive makers are sold out by choice

A hard drive shortage is unusual. A flash shortage needs no explanation beyond a forecast that missed, because fabs take years to build. Hard drive plants are assembly and test lines fed by head wafers and platters, smaller investments, though Seagate’s chief executive Dave Mosley warned on 28 April 2026 that adding output is “not like just plugging in a few more machines”. What makes this shortage a choice is that both large makers have said, on the record and repeatedly, that they will not fund more units.

Western Digital’s chief executive, 5 August 2026:

This does not require spending CapEx to add unit capacity, but we are making the necessary investments in our heads and media operations, as well as in automation to increase our productivity.

Seagate’s chief executive Dave Mosley, on 28 April: “The total number of units is not really increasing.” And on 28 July, again Mosley: “we’re not really increasing the box count”, and “We remain disciplined in securing orders from these customers prior to initiating drive production”. The growth comes instead from putting more terabytes in each drive, the subject of the next section.

What the cloud operators get in exchange is certainty, written down. Seagate said in April that it was “finalizing build-to-order contracts with these customers through the end of fiscal 2027, which defines specific configuration and pricing”, and in July that its agreements set “both product configuration and pricing terms covering the entirety of calendar 2027”, with most nearline exabytes allocated into 2028. Western Digital said in August: “We are very much in the throes of discussions with customers to establish LTAs for calendar year 2029, 2030, and 2031 as well”.

The results, in Seagate’s and Western Digital’s releases, show what that discipline earns:

The volumes grew by less than the money. Seagate shipped 218 exabytes in the June quarter, 34 per cent more than a year earlier, of which 195 exabytes was nearline, up 43 per cent; the rest of its business, desktop, NAS, surveillance and edge drives, shipped 23 exabytes, 10 per cent fewer (Blocks and Files, 30 July 2026). Western Digital shipped 231 exabytes, up 22 per cent. Revenue and exabytes together give the price, and the result agrees with what each company said:

revenue growth / exabyte growth = growth in revenue per exabyte (rough: mix, non-drive revenue)
  Seagate          1.48 / 1.34 = 1.104   -> about +10%
                   (put to it on the call: "10% year-over-year price per exabyte growth in June")
  Western Digital  1.44 / 1.22 = 1.18    -> about +18%
                   (on the call: price per terabyte up "high teens" year on year)

The Seagate figure was put to the company by Morgan Stanley’s analyst Erik Woodring on the July call, who reminded it that it had previously spoken of price per exabyte growing in the mid to high single digits, and added that the September-quarter guidance “implies pricing growth closer to maybe 20% year-over-year or even above that”. The chief financial officer’s answer began “You are correct”, and went on: “the gap between supply and demand is now a little bit bigger than a few quarters ago”. Western Digital’s finance chief said its price per terabyte had gone from high single digits year on year in the March quarter to high teens in June, and an analyst on its call put the September quarter at “probably 20%”.

Woodring had set out the frame in a note of 15 June 2026: hard drive demand “growing 40% to 50% annually while supply growth runs closer to 30% to 35%”, shortages “through at least CY28”, nearline pricing “below $15 per terabyte” and vendors “targeting $25-30 per terabyte over the next two to three years”, as Yahoo Finance reported it. That last clause is an analyst’s account of the makers’ intentions, not a maker’s statement, and it is a forecast:

Morgan Stanley, 15 Jun 2026: nearline below $15/TB, makers aiming at $25-30/TB
  25 / 15 = 1.67     30 / 15 = 2.00    -> +67% to +100%
  spread over 3 years:  1.67^(1/3) = 1.19  -> about +19% a year
  spread over 2 years:  2.00^(1/2) = 1.41  -> about +41% a year

At contract, then, the hard drive makers’ prices were rising by a tenth to a fifth a year by the June quarter, and faster each quarter: from mid to high single digits in the March quarter to about a fifth implied for September, with Nikkei, via MoneyDJ and TrendForce’s news desk on 22 July, reporting nearline prices up about 10 per cent in the April-to-June quarter alone. That is still a fraction of the retail rise, and the path Morgan Stanley describes would keep them climbing at that pace or faster into 2028. The two published contract price levels do not sit easily together: TrendForce put the average hard drive at $15 to $16 per terabyte in October 2025, and Morgan Stanley put nearline below $15 in June 2026, after eight more months of rises. Nearline drives are the cheapest per terabyte, which may explain part of the gap; the comparison with retail below uses both. Seagate’s own tables give a third measure, revenue per terabyte shipped: $13.30 per nearline terabyte in the September 2025 quarter and $15.04 in the June 2026 quarter, closer to Morgan Stanley’s level than to TrendForce’s.

Somebody pays for the allocation, and the published figures name who. Seagate’s non-nearline exabytes fell while its prices rose: worked from its own tables, its revenue per terabyte outside the data centre rose by about a third in the year to the June quarter, against about a tenth inside it. It deferred lower-capacity heat-assisted drives for edge markets in April because “demand is so strong in the public cloud that we don’t have enough volume to also implement this lower capacity base” strategy, its finance chief said, and Nikkei reported in September that it had begun phasing out some low-capacity models. Toshiba, the third maker, is privately owned and says little in public. Backblaze, which buys drives by the thousand, added no new drive models in either of the first two quarters of 2026, “a curiosity, especially to have that happen two quarters in a row”, its Drive Stats report of 29 September noted, though that “doesn’t mean no new deployments”: it kept adding drives of models it already ran, mostly of 20 TB and up.

The makers chose margin over volume, said so on the record, and the buyers signed anyway. The reason they could is that nobody else makes hard drives: the three companies are the whole industry, and in 2023 they had all learned what happens when the cloud stops buying. Why that falls short of collusion, or does not, is a later section.

HAMR, ePMR and UltraSMR add terabytes, not drives

A hard drive maker grows supply without building a factory by making the bits smaller. The same platters, motor and actuator hold more data if the recording density rises, which is why HDD or SSD calls areal density the improvement that costs nothing in the bill of materials. Seagate’s July call put it as a slogan: “Advancing areal density is our North Star.”

The three makers are taking different routes to the same end:

The arithmetic of a platter shows why this is the lever:

Mozaic 4+:  44 TB / 10 platters        = 4.4 TB per platter
Mozaic 5:   5 TB or more x 10 platters = 50 TB or more, same drive count
Seagate's target: nearline exabytes up "in the mid-20% range" a year,
  with units flat  -> average capacity per drive up about a quarter a year

Moving the mix from 3 TB to 4 TB per platter “is, of course, giving us another boost in terms of profitability”, Seagate’s finance chief said in July (Blocks and Files), though TrendForce attributed part of the rise in average hard drive prices in October 2025 to the makers passing on the cost of the move to heat-assisted recording.

Three things keep this from reaching a shopper soon. The new platforms go to the hyperscalers first; Seagate’s own announcement of Mozaic 4+ describes broader availability as something planned. The low-capacity versions are deferred, which leaves the desktop and NAS drives most people buy on older platforms whose output is not growing. And much of Western Digital’s growth is shingled, which a hyperscaler’s software can manage and a home NAS often cannot, as CMR vs SMR explains.

The suppliers of the parts are expanding too, on slower clocks. The Elec, relayed by TrendForce’s news desk on 22 July 2026, reported that Resonac is expanding its Singapore plant from 160 million to 210 million platters a year, that HOYA is investing about JPY 50 billion in a glass-substrate plant in Vietnam due to open in 2028, that TDK is adding head capacity, and that Nitto Denko is investing JPY 43 billion through 2028.

Resonac, Singapore media:  210 / 160 = 1.31  -> about +31% platters a year

Areal density is how a hard drive maker grows supply without building a factory, and it grows only as fast as the newest platform ramps. Seagate’s target of mid-20s per cent exabyte growth sits below the 40 to 50 per cent demand growth Morgan Stanley estimates, and it depends on the HAMR ramp reaching 70 per cent of nearline exabytes by June 2027, as Seagate plans.

NAND flash contract prices compound to about five times

Contract prices are what PC makers, phone makers and cloud providers pay the flash makers. TrendForce’s quarterly forecasts are the only public series, and its revenue rankings and the makers’ own results are the only checks on them. The table sets each quarter’s first forecast beside its revisions and the evidence of what happened.

The first quarter of 2026 is the record: one quarter’s forecast raised twice in eight weeks, from 33 to 38 per cent to 85 to 90 per cent, as the makers moved capacity from flash to DRAM and North American cloud providers stockpiled enterprise SSDs, in TrendForce’s explanation of 2 February.

Multiplying the quarters out is the step nobody publishes:

flash contract prices, TrendForce's latest forecast for each quarter, as multipliers
quarters: 4Q25, 1Q26, 2Q26, 3Q26, 4Q26
  low ends   1.20 x 1.85 x 1.70 x 1.10 x 1.15 = 4.77
  high ends  1.25 x 1.90 x 1.75 x 1.15 x 1.20 = 5.74
calendar 2026 alone (1Q26 to 4Q26):
  low ends   1.85 x 1.70 x 1.10 x 1.15 = 3.98
  high ends  1.90 x 1.75 x 1.15 x 1.20 = 4.59

Every factor is a forecast, because TrendForce publishes no measured quarterly change for flash, and both quarters it revised were revised upward, so 4.8 to 5.7 times is a forecast-weighted estimate, not a measured index, and more likely an understatement than an overstatement. Three independent checks agree with it where they overlap.

The first is the makers’ own selling prices for the April-to-June quarter, which are measurements, not forecasts:

The second is spot. Gamers Nexus put the spot price of a 512Gb TLC wafer at over $23 in April 2026 against about $2.70, its earlier stable level, before October 2025, more than eight and a half times. Spot is thin and volatile, and it ran ahead of contract, as it usually does in a shortage.

The third is revenue, which rose almost entirely through price:

top-five flash revenue, TrendForce
  4Q25  $21.17 bn    1Q26  $38.90 bn    2Q26  $68.87 bn
  68.87 / 21.17 = 3.25   -> more than three times in two quarters
TrendForce's price forecasts for the same two quarters, 1Q26 and 2Q26:
  1.85 x 1.70 = 3.15    1.90 x 1.75 = 3.33   -> revenue sits inside the range
Kioxia, April to June: selling price +70%, bits +low single digits,
  revenue +76.2% on the quarter   -> nearly all of it price

For enterprise SSDs, the category that drove the rest, TrendForce estimated on 25 August a rise of about 35 per cent in the second half of 2025 and a cumulative 235 per cent in 2026, a forecast for a year not yet over.

The slowdown in the third quarter came from the demand side, not from new supply: consumer customers’ tolerance of higher prices had “reached its limit”, TrendForce said on 3 July, with PC makers sitting on high inventories. Then it re-accelerated, led by enterprise SSDs; the release of 30 September says “most incremental supply has already been committed, leaving little available on the open market”.

In every quarter where TrendForce changed its flash forecast in this cycle, it changed it upward, so read any upward forecast from this period as a floor rather than a central estimate. What that means for timing a purchase is the subject of buy storage now or wait.

Wafers, cleanrooms and DRAM: why the flash makers did not simply build more

Three of the five large flash makers, Samsung, SK hynix and Micron, also make DRAM, and in 2026 DRAM paid better, as the memory site’s guide to the 2026 DRAM shortage shows. The same cleanrooms, much of the same equipment and the same capital budgets serve both. Where a maker had a choice, flash lost.

The record is explicit. TrendForce’s revised forecast of 2 February 2026 said the makers were reallocating capacity from flash to DRAM. ETNews and The Bell, as relayed by TrendForce’s news desk on 30 March, reported that SK hynix’s Dalian plant in China is “the only available site for expanding NAND capacity” because its other facilities had been converted to DRAM, with equipment going in during the second half of 2026; China made 40 to 45 per cent of SK hynix’s flash in 2025. TrendForce’s revenue ranking of 18 August said suppliers were prioritising DRAM and HBM investment over flash expansion. And Micron’s new fab in Japan, where it held a groundbreaking ceremony in its June to August quarter, is a DRAM fab, with initial output expected in late 2028, it said on 30 September.

Cleanroom space is the constraint money cannot quickly relieve. The Korea JoongAng Daily reported on 12 March 2026 that cleanroom space would limit memory supply growth “this year and next”, in the words of an unnamed industry source, that the Korean makers were converting lines to newer processes rather than adding tools, and that new fabrication capacity in Korea was unlikely to come online before the second half of 2027. Samsung said on its July call that the lead time from new fab construction to actual wafer production “exceeds three years”.

So the flash makers grew bits the other way, by moving to finer processes with more layers. Samsung moved its Xi’an fab in China from 128-layer to 236-layer flash, with a second phase going to 286 layers in 2026 (ETNews via TrendForce’s news desk, 30 March), and scheduled its tenth-generation V-NAND for mass production in August. Kioxia’s newest fab, K2 at Kitakami, began operating in September 2025, and Kioxia began sampling its 332-layer BiCS10 in 2026, with mass production targeted for 2027 (Nikkei and EE Times Japan via TrendForce’s news desk, 3 July). Sandisk, Kioxia’s partner in those fabs, said in August that its BiCS8 generation was now the majority of its bit output, that capital spending would rise in dollars but fall to about 6 per cent of revenue in fiscal 2027, and on that call it announced no new fab; three weeks later Kioxia announced, with Sandisk, site preparation for a third Kitakami fab, due to operate in fiscal 2029. A node transition adds bits without adding wafers, but slowly, and it costs output while lines are converted.

The result is an industry that grew its bit shipments by about a fifth in 2026 and still fell short:

industry flash bit shipments, Micron's estimate (30 Sep 2026)
  2026:            up in the low 20s per cent
  2027 and 2028:   up in the mid 20s per cent, "supply constrained" in both
enterprise SSD bit demand, 2026: up more than 80% (TrendForce, 30 Sep 2026)
enterprise SSD share of bits a year earlier, 2Q25: 26% (Counterpoint)

Enterprise SSD demand growing by 80 per cent from about a quarter of all bits asks for roughly all of that growth on its own, which is why, as the AI section worked out, everything else got less flash.

There is a second, smaller link to DRAM inside the drive itself. Many SSDs carry a DRAM chip as a cache for the controller’s mapping tables, and TrendForce wrote on 3 July 2026 that shortages of internally sourced DRAM were constraining the supply of small-capacity, high-performance enterprise SSDs.

A flash maker could add bits in 2026 in two ways, a new fab or a finer node, and only the second was available before 2027, while the cleanrooms that might have taken new flash tools were going to DRAM. That is why a flash market growing its output by a fifth could still be in deficit, by 4 to 5 per cent of bits in 2026 on TrendForce’s estimate of 21 July.

The makers in their own words: allocated, contracted and in no hurry

The makers’ calls are the best evidence on supply, and the least neutral: every statement below was made to investors by a company that profits from a tight market. The table collects the ones that bear on timing, with the date.

The Sandisk line about the TAM, its total addressable market, is the consumer half of the story: its consumer revenue fell 32 per cent on the quarter while data-centre revenue rose 103 per cent, to $2.977 billion, and its revenue of $8.97 billion grew 51 per cent, “approximately 1/3 from higher volumes and 2/3 from higher pricing”.

Behind the words are contracts. In January 2026 the makers were still selling quarter by quarter; by the middle of the year the largest buyers had signed multi-year agreements, with volumes committed, deposits paid and prices bounded.

The cloud providers asked for these. TrendForce’s release of 31 March 2026 said they were “willing to accept higher prices and sign LTAs to secure a stable supply”. For the makers they turn a price spike into an income floor; Sandisk said in August that it expects attractive margins even at floor pricing. For the buyers they cap the next increase: some agreements signed from the second quarter of 2026 include price ceilings, TrendForce noted on 25 August. For everyone without one, they shrink the supply left to negotiate over.

A floor set at the top of a cycle limits how far contracted prices can fall, and a ceiling keeps the largest buyers’ prices below everyone else’s. That combination is the clearest reason the forecasters who expect supply to catch up in 2027 or 2028 still do not expect 2023 prices.

Retail is the residual, which is why shoppers were hit hardest

Allocation, not price, decided who got drives first, and the queue has a shape. Each row is what was reported for that kind of buyer.

The difference between the top of the queue and the bottom is the most useful number in this article, and nobody publishes it directly. It can be bracketed from the figures above:

hard drives, change on a year, by layer
  Seagate revenue per exabyte, June quarter       about +10%
  Western Digital price per terabyte, June qtr    "high teens"
  ComputerBase, 12 popular drives in German shops +134%   (Sep 2025 to Sep 2026)

price per terabyte, by layer
  hard drives at contract, average (TrendForce, Oct 2025)    $15 to $16
  nearline at contract (Morgan Stanley, Jun 2026)            below $15
  German retail hard drives (PC Games Hardware, Aug 2026)    EUR 33 to 52, with 19% VAT
    without VAT:  33 / 1.19 = 27.7    52 / 1.19 = 43.7
    in dollars at 1.1593 (ECB, August 2026 average):  about $32 to $51
  -> retail about two to three and a half times the contract price per terabyte
     (32 / 16 = 2.0     51 / 15 = 3.4)

The comparison is rough: the cloud buys 24 TB to 44 TB nearline drives and the shop sells desktop and NAS drives of every size, and the contract figures are a research firm’s and an analyst’s, eight months apart. But it is consistent with everything else in the record. The shopper is not paying the cloud’s price plus a margin; the shopper is bidding for the part of the output nobody contracted, and that part shrank. Seagate’s non-nearline exabytes fell 10 per cent in a year when it shipped a third more in total.

Flash shows the same thing from the seller’s side. Sandisk’s consumer revenue fell 32 per cent in a quarter while its data-centre revenue doubled, and its finance chief’s description, that consumer prices “have also come up, and there has been some impact on the TAM itself”, is a polite way of saying that fewer people bought. TrendForce wrote on 31 March 2026 that PC and phone makers were being forced to cut storage capacities, and that suppliers meant to hold prices through the second quarter “by continuing to limit supply to client SSDs”; on 30 September it wrote that PC brands were cutting SSD capacities in mainstream models and running on inventories.

The devices followed. Gartner, as CRN Australia reported on 13 March 2026, expected combined DRAM and SSD prices to rise about 130 per cent by the end of 2026, PC prices to rise 17 per cent, and businesses to keep their PCs about 15 per cent longer. IDC forecast on 2 June a fall of 11.3 per cent in PC shipments and on 26 August a record fall of 16.7 per cent in smartphones, with memory prices to “continue increasing until at least 2028”. HP said in February that memory had risen to about 35 per cent of a PC’s materials cost, the detail of which is in the memory site’s DRAM shortage guide. A laptop at last year’s price now tends to carry a smaller SSD, and one with the same SSD costs more.

Cloud storage repriced too, which says what the hardware costs

Cloud storage companies buy drives in bulk and publish their prices, which makes them a useful second opinion on what the hardware now costs.

Wasabi raised its pay-as-you-go price from $6.99 to $7.99 per terabyte-month from 1 July 2026, citing “increased costs for storage hardware, energy, and data center operations driven by growing demand and supply constraints” in its notice. It had been $6.99 since October 2023, up from $5.99, as Storj noted at the time. Backblaze raised B2 from $6 to $6.95 per terabyte-month from 1 May 2026, removing standard API transaction fees at the same time, after an increase from $5 to $6 in 2023. Backblaze’s chief financial officer Marc Suidan explained the hardware side on its call of 4 May 2026:

The general equipment cost is 30% higher than it was on a per-unit basis a year ago.

Backblaze also said it was accelerating capital spending from 2027 into 2026, a professional buyer with full visibility of the market choosing to buy early.

Wasabi           7.99 / 6.99 = 1.143   -> +14.3%
Backblaze B2     6.95 / 6.00 = 1.158   -> +15.8%  (with API fees removed)
equipment cost   +30% per unit, year on year (Backblaze, May 2026)

The storage price rose about half as much as the equipment, because drives are only part of what a stored terabyte costs: power, space, redundancy and staff are the rest, and they did not rise by a third. Consumer cloud plans have not moved at all so far: Apple’s iCloud+ price list of 16 September 2026 still has 2 TB at $9.99 a month, and Google One’s European page on 2 October 2026 has 2 TB at EUR 9.99. Cloud prices rose by about a sixth while the hardware under them rose by about a third, which is the clearest public evidence that the retail doubling is a shelf price, not a cost of storing data.

Drive prices in Germany, the US and Japan: one shock at three sizes

Germany has the best public record, because two publications check the same products every month and both exclude eBay and Amazon Marketplace sellers. ComputerBase’s series, each figure against mid-September 2025:

SSDs led until June, and hard drives have led since. 3DCenter’s index, which takes the cheapest offer from regular retailers on Geizhals each month, crossed a month sooner: against July 2025 SSDs stood at 200 and hard drives at 196 in May, and in June hard drives led, 212 to 203. On that July baseline hard drives had also led from September to November 2025, before the flash surge. The sharpest monthly rise for internal SSDs came in December 2025, at 28.2 per cent, followed by 22.3 in January, and in September they rose 0.5 per cent; the sharpest for internal hard drives came in January 2026, at 17.8 per cent, April added 15.1 against 6.3 for SSDs, and September still added 7.3. Against July 2025 it has internal hard drives up 157 per cent and SSDs up 125.

Three things explain the crossover. Flash contract increases slowed in the summer because PC makers were holding stock and resisting further rises, as TrendForce’s July release said, and retail flash prices followed. Hard drive allocation did not ease at all, because the drive makers’ contracts run by the year. And buyers who had been priced out of SSDs moved to hard drives, the substitution Nikkei described in September.

External drives lagged throughout; 3DCenter calls them a latecomer to the crisis. Against September 2025 they were 23 per cent dearer in March 2026, when internal drives were about 54 per cent dearer on the same baseline, and 97 per cent dearer in September, against about 150 per cent (both worked from its table). The gap that makes shucking, buying an external drive to take out the drive inside, worth checking has not closed in the index; whether a given external drive is cheaper per terabyte than its internal twin is a model-by-model question.

In the popular basket the largest drives rose most: ComputerBase’s Toshiba Cloud-Scale 18TB went from EUR 288 to EUR 858, up 198 per cent. 3DCenter’s cheapest-offer index found its largest rise at the other end, 1 TB drives up 264 per cent since July 2025, from EUR 33 to EUR 120, and PC Games Hardware put hard drives at EUR 33 to 52 per terabyte in August.

The US has no equivalent monthly basket, so the evidence is a set of snapshots, each named and dated:

The eRacks row, a server seller’s own tracking of the lowest price it found each month, holds an inversion: between March and August the consumer BarraCuda overtook the enterprise Exos per terabyte, because the consumer line rose six times as fast, which is one reason the enterprise drives guide matters more this year than last. The DatacenterDisk index began after most of the rise, so it measures the plateau, not the climb.

Japan’s evidence is wholesale: Nikkei’s record $67.70 for a bulk 3.5-inch 1 TB drive in July to September, up 15 per cent, and $61.20 for a 2.5-inch one, up 10 per cent, the sixth straight quarterly rise, which it put down to the SSD price, Chinese surveillance demand and Seagate phasing out some low-capacity models. No comparable public index for the United Kingdom or the rest of Europe was found in this research.

The same shock produced different numbers in each country because the baselines, the baskets and the stock in each channel differed, not because the cause did. Every series that covers September 2025 to September 2026 shows hard drives and SSDs roughly doubling or more.

Used and recertified drives got dearer for a reason of their own

Recertified drives are drives returned to the maker or pulled from data centres, tested, wiped and resold, usually with a shorter warranty. Their supply is not the factories’ output but the data centres’ retirements, and it fell for a reason of its own before the shortage of new drives sent more buyers to it.

The mechanism was described early. How-To Geek reported on 30 October 2025, citing the distributor Fusion Worldwide, that fewer drives were cycling back from data centres and that some resellers were buying and holding. The logic is simple: an operator that cannot get new drives for a year keeps the old ones spinning. Backblaze’s Drive Stats for the second quarter of 2026 fit it: two HGST models retired only at around eight and nine years of service, and no new drive models arrived for two quarters running, though Backblaze kept deploying models it already ran.

Third-party measurements put the discount and the rise in the same range:

The recertified Exos in the first row roughly doubled, like the new drives around it, so a recertified drive is cheaper than new but not cheap. The enterprise SSD in the last row is the more interesting substitute, because endurance favours it: XDA compared about 1,200 terabytes written for a consumer 2 TB drive with up to 3,600 for an enterprise equivalent, and what those figures mean in years of use is worked out in SSD endurance.

Fraud predates the shortage; a higher price only raises its reward. heise reported on 23 January 2025 that used Seagate Exos drives had reached at least two German online retailers and been sold as new, with unremarkable SMART values while the drives’ fuller logs, Seagate’s FARM data among them, showed 10,000 to 22,000 hours of use; Seagate confirmed that SMART values can be reset. How to read the FARM log, prove a drive’s capacity and judge its hours is the subject of buying used drives on eBay. Age matters too: a peer-reviewed study of 443,156 Backblaze drives, reported by Blocks and Files on 7 August 2026, found it a major predictor of failure.

No neutral index of used drive prices exists. What can be said in general terms is that the second-hand market followed the new one up, at a discount that third-party measurements put at roughly a third, and that its supply depends on retirement decisions the shortage has delayed. On this site the enterprise listings and SAS drives are where most ex-data-centre drives appear, ranked by price per terabyte at the moment you look. Recertified drives got dearer because their supply is the data centres’ retirements, and the data centres stopped retiring.

Tariffs and the euro explain a few per cent, not a doubling

Tariffs come up in every discussion of US hardware prices in 2026, and for drives they are a small and poorly measured part of the story.

What the US actually did, as of late September 2026: it opened a Section 232 investigation into semiconductor imports in April 2025 and, by a proclamation of 14 January 2026, imposed a 25 per cent tariff from 15 January on a narrow set of advanced computing chips defined by performance thresholds, Nvidia’s H200 and AMD’s MI325X among them, with exemptions for chips imported for US data centres, research, consumer and industrial uses and others (EY Global Tax News, 15 January 2026). A second phase that would extend duties to products built from covered chips had no final rates or date as of 31 August (Tech Times), and whether memory, SSDs or hard drives would be covered was not stated. The Supreme Court ruled 6-3 on 20 February 2026 that tariffs imposed under emergency powers were unlawful; a temporary 10 per cent tariff that followed lapsed on 24 July, replaced by Section 301 duties of 10 or 12.5 per cent on imports from 60 trading partners (Wikipedia’s summary, read on 2 October 2026). Whether a given drive is exempt depends on its classification and origin, which this article has not checked.

Tariffs did move demand in time. TrendForce named “anticipated U.S. tariff increases spurring early production by consumer brands” as one cause of the flash recovery in the second quarter of 2025, a pull-forward of buying, not a change in supply. No primary source found in this research quantifies a tariff effect on drive prices, and one US tracker, DatacenterDisk, published a correction on 28 September 2026 withdrawing a claimed tariff effect of 17 to 145 per cent and an enterprise hard drive rise of 46 to 50 per cent, because neither traced to a primary source.

The decisive evidence is geographic. The doubling happened in Germany, where no US tariff applies. ComputerBase’s 134 per cent and 129 per cent are euro prices in German shops, and the only currency effect on them runs the other way from what people assume:

ECB reference rate, 15 September 2025    1.1766
ECB reference rate, 16 September 2026    1.1537
1.1766 / 1.1537 = 1.020   -> over ComputerBase's own window, a dollar-priced
                             drive costs about 2% more in euros from currency alone
against +134% for hard drives and +129% for SSDs in German shops

The euro had strengthened by 13 per cent during 2025, from a January average of 1.0354 to September’s 1.1732 on the ECB’s reference rates, which flattered European prices just before the shortage began; since September 2025 it has given a little back. Currency explains about two points of a rise of more than a hundred and twenty, and tariffs, whatever they add in the US, cannot explain a rise that happened at the same size in a market they do not reach.

Record margins, and the price-fixing question

The makers’ own accounts show where the money went:

The flash makers’ margins are DRAM-inflated for the three that also make memory, but Kioxia and Sandisk make only flash, and their 80 per cent and 83 to 85 per cent belong to flash alone.

Margins like these invite the question whether the shortage is engineered. The only lawsuit so far concerns DRAM: a proposed class action filed in June 2026 in the Northern District of California alleges that Samsung, SK hynix and Micron used the shift to HBM as a pretext to restrict commodity DRAM supply, untested allegations covered in the memory site’s DRAM shortage guide. No comparable case over flash or hard drives was found in this research.

The case for suspicion rests on the record of behaviour. The flash makers cut output in the second half of 2025 while prices were already rising, a move Chosun Biz, as TrendForce’s news desk relayed it, described as part of “a coordinated push to lift NAND prices”; a newspaper’s word for simultaneous cuts is not proof of an agreement, but it is the word that was used. In March 2026 TrendForce described suppliers holding prices “by continuing to limit supply to client SSDs”. The hard drive makers say openly that they will not add units, and three companies are the whole industry; that concentration itself came from the last shock, when the top two makers’ share rose from 62 to 85 per cent between the third quarter of 2011 and the first of 2012 after Seagate bought Samsung’s drive business and Western Digital bought HGST, an “oligarchy where the top players can control pricing”, in the words of an IHS analyst at the time (Computerworld, 6 June 2012). And the vocabulary is the vocabulary of restraint: Seagate is “disciplined in securing orders”, Western Digital will not spend to add unit capacity, Sandisk’s capital spending falls to 6 per cent of revenue.

The case against rests on everything else. Both industries lost money in 2023 and were told by the cloud operators, in effect, that demand could vanish for a year. Flash capacity takes more than three years to build, on Samsung’s own statement. The makers are building: Kioxia began site preparation for a third Kitakami fab in August 2026, targeting operation in fiscal 2029, and with Sandisk plans about JPY 5 trillion of investment over six years to 2032 (Blocks and Files, 28 August 2026); Micron’s Singapore flash fab is due to produce in the second half of 2028 and Micron said on 30 September that it would raise capital spending in fiscal 2027. And the hard drive makers’ own price rises at contract, a tenth to a fifth a year by mid-2026, are far below the retail rise, which is not what a cartel extracting maximum price from its largest customers would look like. What separates disciplined supply from coordinated supply is a question for a court, and nothing public settles it.

New flash fabs and drive platforms have dates, and most say 2027 and 2028

Each announcement moves through stages. For flash, a cleanroom is built, tools go in, the first wafers come out, and output ramps over several quarters; a node transition skips the building but costs output while lines convert. For hard drives, a new platform is qualified by the large customers and then takes a growing share of output. The dates below are the stage each source names.

Samsung’s P5 is a mixed site, described by the Korea JoongAng Daily as able to make HBM, DRAM, flash and potentially foundry chips; how much of it will make flash has not been published.

Read down the flash rows and a pattern appears. Before 2028, new flash bits come from node transitions in existing buildings: V10, BiCS10, Dalian’s equipment. The new buildings that add wafers for flash, Micron’s Singapore fab, SK hynix’s M17 and Kioxia’s Fab3, are dated late 2028, December 2028 and fiscal 2029. That is consistent with Micron’s industry outlook of 30 September, bit growth in the mid-20s per cent in 2027 and 2028 with supply still constrained, and with Phison’s warning of November 2025 that new lines would not run until late 2027.

when the new flash buildings add output (dates as given by each source)
  Micron Singapore      2H 2028
  SK hynix M17          first cleanroom Dec 2028  -> output after that
  Kioxia Fab3           fiscal 2029, i.e. April 2029 to March 2030
  YMTC Wuhan phase 3    50,000 wafers a month by 2027, against about
                        200,000 a month at its existing fabs (Evertiq)
                        200 + 50 = 250   -> about +25% for YMTC alone

China is the one source of new wafers before 2028. TrendForce expects Chinese suppliers to produce nearly 19 per cent of flash bits in 2027, and YMTC already shipped 14 per cent in the second quarter of 2026 on Counterpoint’s count. But YMTC has been on the US Entity List since December 2022, it sells mostly into consumer and low-price channels, and, as the Gamers Nexus table showed, a drive on its flash doubled in price like everything else.

The hard drive rows are closer in time, but they add terabytes per drive, not drives, and their first customers are the hyperscalers. The first date on any announcement is the start of a ramp, not the arrival of supply: new flash buildings land in late 2028 and 2029, and everything before that is node transitions and China.

When will hard drive and SSD prices come down? Every dated forecast, with each forecaster’s interest

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

The Citi figures are the most explicit model in the table, and they reconcile from its own inputs:

Citi, note of 14 Sep 2026: flash demand growth against supply growth
  2027  demand +29%, supply +21%:   1.21 / 1.29 = 0.938  -> about 6% short
  2028  demand +33%, supply +25%:   1.25 / 1.33 = 0.940  -> about 6% short
  (Citi gives -6.1% and -5.5%; its exact method is not published)

Count what the table says about flash, keeping to the statements about flash itself: TrendForce, Kioxia, Sandisk, Phison, Jefferies and Citi’s NAND ratios. None expects relief in 2026, and none expects falling contract prices before the second half of 2027; TrendForce’s own forecast for the current quarter is a further rise. After that they split by who is speaking. TrendForce places the turn in the second half of 2027; Kioxia and Sandisk say 2027 will still be short; Citi has the flash deficit widening in 2027 and barely narrowing in 2028; Jefferies sees no meaningful relief until 2028. The memory-wide forecasts lean the same way: Gartner’s analyst places a modest decline in the second half of 2027, IDC saw no relief before the end of 2027 and memory prices rising until at least 2028, Micron and Samsung place 2027 tighter than 2026, and the SK group says 2030. The makers have an interest in describing supply as tight while they negotiate multi-year contracts, and TrendForce in being right about bits; neither interest settles who is correct. Micron’s own view moved within one quarter, from gradual improvement in 2028 to “much tighter” in 2027 and 2028, the strongest reason to treat any single date as provisional.

For hard drives the table is thinner and more uniform. The two large makers say their output is allocated into 2028 and are negotiating beyond it, and Morgan Stanley, the only independent forecaster with a dated view found in this research, expects shortages “through at least CY28”. No dated forecast puts a hard drive turn before 2028.

So, narrowly: for SSDs, the first relief is forecast between the second half of 2027 and 2028, with TrendForce earliest and the makers later; for hard drives, nobody with a dated forecast expects relief before 2028; and for a return to 2023 or 2025 prices, nobody has a date at all. No drive forecaster has published a level at which prices might settle. The only such band in circulation, 30 to 50 per cent above 2024 averages, is TrendForce’s and Bernstein’s for RAM, as PC Games Hardware summarised them in August; the long-term agreements’ floors point the same way for flash, without saying where.

The yardstick that remains is the one set out in the section on which price you read: ComputerBase’s September 2025 prices per terabyte, as a stand-in for the pre-shortage level. The distance between a listing’s price per terabyte on the day you look and those figures is the shortage premium still in it; how much of it comes off, and when, is the question no published forecast answers for drives.

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

Two earlier storage shocks show what a turn looks like. The first is the largest previous hard drive price shock. Floods in Thailand in late 2011 shut factories behind about a quarter of world hard drive production. Backblaze’s purchase records show what followed:

Backblaze's hard drive costs, 2011 Thai floods (Backblaze, 26 Nov 2013)
  Hitachi 3TB, the model it bought:  Sep 2011 $0.044/GB   Nov 2011 $0.083/GB
    0.083 / 0.044 = 1.89   -> +89% in two months (Backblaze puts it at 88%)
  its average cost per GB peaked at $0.064 and was back to $0.044 in Sep 2013
    -> 24 months to return to the pre-flood price

IHS iSuppli forecast in June 2012 that prices would not return to pre-flood levels until 2014, partly because PC makers had signed long-term agreements locking in prices about 20 per cent above pre-flood levels, while drive production would recover completely by the third quarter of 2012. The factories recovered within a year; the prices took two years, held up in between by contracts signed near the top. That shock was a supply loss that ended when the plants dried out. This one is a demand shift with no end date, which argues for a slower turn, not a faster one.

The second is the flash bust of 2018 and 2019, when 3D NAND yields finally rose. TrendForce forecast in October 2018 contract prices down 10 to 15 per cent in the fourth quarter of 2018 and 25 to 30 per cent in 2019. As a thought experiment, not a forecast, here is how long falls at that pace would take to undo the retail rise recorded in Germany:

undo a 2.3x rise (ComputerBase SSD basket, +129%) at the 2019 pace
  at -30% a year:  ln(2.3) / ln(1 / 0.70) = 0.833 / 0.357 = 2.3 years
  at -25% a year:  ln(2.3) / ln(1 / 0.75) = 0.833 / 0.288 = 2.9 years

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

A slump in AI spending is the main route to an early end, and it would reach the price layers in order of how loosely they are contracted. Spot and retail would move first, as SSD retail prices showed by flattening in the summer of 2026 on inventory alone. Contracted volume would move last: Sandisk’s agreements carry floors, Seagate’s prices for 2027 are fixed, and Micron’s agreements run to 2030. A slump would lower retail prices before contract prices, and contracted prices no further than their floors.

The QLC row cuts both ways. If high-capacity flash keeps replacing nearline disk, the hard drive shortage eases and the flash shortage deepens; if the new hard drive platforms ramp faster than planned, cloud buyers return to disk and flash loosens. The loop that started the shortage could run in reverse.

Apart from Chinese flash, the least certain brake, every route to an early end runs through demand falling, not supply rising. How to act on that is in buy storage now or wait.

Five things that would settle much of this are not published: a hard drive contract-price series (only TrendForce’s October 2025 average, Morgan Stanley’s “below $15” and Nikkei’s one-quarter rise exist), Toshiba’s own outlook, the share of output each tier of buyer receives, any measure of tariff pass-through on drives, and a neutral index of used and recertified prices. Every conclusion above is drawn around those gaps, not through them.

Twenty-one questions people ask, each answered in brief

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

Why are hard drives so expensive in 2026? The three makers did not add capacity after the 2023 downturn, AI data centres booked their output into 2028, and shops sell what is left.

Why are SSDs so expensive? Enterprise SSDs went from about a quarter to nearly half of all flash shipped in a year, the makers gave cleanroom space to DRAM, and TrendForce’s contract forecasts compound to about five times.

Why did hard drives go up if AI uses SSDs? AI uses both. Cloud operators keep training data and cold copies on disk, and when disk ran out they bought flash, which then ran out too and sent consumers back to disk.

Why did RAM rise more than drives? ComputerBase’s September check had RAM up 370 per cent in a year against 134 for hard drives and 129 for SSDs. DRAM makers trade ordinary wafers for HBM, which takes about three times the wafer per gigabyte on the memory site’s figures, and hard drives need no memory wafers at all. The memory side is in the 2026 DRAM shortage.

Will SSD prices go down in 2027? Perhaps late in the year. TrendForce places the first relief in the second half of 2027, and Gartner’s analyst a modest decline then; the makers say 2027 will be tighter than 2026, and no forecast sees a fall in 2026.

Will hard drive prices go down in 2027? No dated forecast expects relief before 2028. Seagate’s output is allocated into 2028 and Western Digital is negotiating to 2031.

Will prices return to 2023 levels? No forecaster has a date for it. The 2023 prices were loss-making for the makers, and long-term agreements carry floors.

Is this a cartel? The only lawsuit concerns DRAM. The makers’ restraint is public and deliberate; whether it is coordinated is for a court.

Did tariffs cause it? No. German prices rose 134 per cent for hard drives without any US tariff, and the euro moved by about 2 per cent over the same year.

Is “46 per cent” a US figure? No. It is ComputerBase’s German basket from September 2025 to January 2026; Tom’s Hardware found US prices similar or worse.

Why did big drives rise more than small ones? In ComputerBase’s basket of popular models they share platforms with the nearline drives the cloud buys. In Germany the discount narrowed: a 24 TB BarraCuda still costs about a third less per terabyte than a 4 TB IronWolf, but an 18 TB Toshiba enterprise drive now costs nearly as much. Among the cheapest offers, 1 TB drives rose most of all.

Did HAMR not make drives cheaper? It adds terabytes per drive, and the first ones go to hyperscalers. Low-capacity versions were deferred.

What happened to Crucial? Micron stopped Crucial retail shipments in February 2026 to supply larger customers; warranty service continues.

Are recertified or used drives a way round it? Partly. They rose too, a recertified Exos 28TB doubled in German shops, but trackers put used and recertified drives about a third below the same model new. See buying used drives on eBay.

Is shucking external drives still cheaper? Sometimes. External drives rose more slowly: 3DCenter had them 97 per cent dearer in September 2026 than a year earlier, against about 150 per cent for internal drives. Compare price per terabyte model by model.

Will Chinese SSDs bring prices down? Possibly from 2027, as Chinese makers approach a fifth of flash bits. So far a drive on Chinese flash doubled like the rest.

Is cloud storage a cheaper place to wait? Business cloud storage rose about 15 per cent in 2026; consumer plans have not moved. Whether renting beats buying is worked through in buy storage now or wait.

Why did my new laptop come with a smaller SSD? PC makers cut storage in mainstream models to hold prices, TrendForce reported in March and September 2026.

Would an AI bubble bursting fix it? It is the main route to an early end. Retail would fall first; contracted prices last, and no further than their floors.

Should I buy now or wait? If the need is real, no forecast rewards waiting through 2026; the full decision is in buy storage now or wait.

How much should I buy while it is expensive? What your data needs plus honest growth, not a round number: how much storage you really need.

Coping is a question of timing, size, medium and condition, at home and in a business

Timing. If the storage is needed within a year, no dated forecast rewards waiting: TrendForce’s forecast for the current quarter is another rise, and nobody with a dated view puts a hard drive turn before 2028. If the drives you have work and your data has a second copy, waiting for an SSD upgrade is defensible. The full decision is in buy storage now or wait.

Size. Every terabyte is priced at shortage rates, so buy what the data and its copies need for the years you can see; how much storage you really need works it through.

Medium. At the smallest sizes a hard drive no longer saves anything: 3DCenter’s cheapest 1 TB hard drive cost EUR 120 in September, against EUR 109 for the cheapest 1 TB SATA SSD and EUR 140 for the cheapest 1 TB PCIe 4.0 NVMe drive. At large sizes compare price per terabyte capacity by capacity, and HDD or SSD says which medium suits which job.

Condition. Recertified and ex-data-centre drives rose too, but sit about a third below the same model new on DatacenterDisk’s comparison; check each one as buying used drives on eBay describes.

Businesses. Backblaze pulled capital spending from 2027 into 2026; server makers raised prices about 15 per cent, with Dell adding 17 per cent in March; and Gartner expects businesses to keep their PCs about 15 per cent longer. Quotes that hold to delivery, and capacity sized to measured growth rather than to a round number, are the levers a buyer without a long-term agreement has.

What to do with this on a listing page

1. Read price per terabyte, not price per drive. The hard drive listings, SSD listings and NVMe listings rank by it, and every drive puts both media on one scale.
2. Check each capacity rather than assuming big is cheaper. The large-drive discount narrowed in 2026; compare a filter such as 4 TB drives with 3.5-inch drives sorted by price per terabyte.
3. Price enterprise drives beside consumer ones. In the US the consumer lines rose faster in 2026; enterprise listings and SAS drives are where ex-data-centre stock appears, and enterprise drives and drive interfaces say what they need to run.
4. For a NAS, exclude shingled drives with no SMR; the reasons are in CMR vs SMR and drives for a NAS.
5. Remember that bid-only auctions are hidden by default, because a bid is not a price, and Buy It Now only removes auctions altogether.
6. Check every used drive on arrival, the FARM log and the capacity, as buying used drives on eBay describes, and check a used SSD’s wear against SSD endurance.
7. Know what the number includes: how this site works out price per terabyte.
8. Re-read the page on the day you buy. Nothing in this article freezes a site price.

The drives went to AI data centres first, as nearline disk and then as high-capacity flash; the makers had stopped adding capacity after the bust of 2023 and chose to sell out rather than build; and the dated forecasts put the first relief for SSDs between late 2027 and 2028 and for hard drives no earlier than 2028, with nobody on the record dating a return to old prices. That is the state of the evidence on 2 October 2026. The next pieces are Seagate’s, Western Digital’s, Samsung’s and SK hynix’s results in late October on past cadence, Kioxia’s in November, Micron’s in December, TrendForce’s first forecast for 2027 around the turn of the year, and ComputerBase’s monthly check in mid-October. Read each of them, like every figure above, by its layer and its date. The memory side of the same shortage is in the memory site’s guide to the 2026 DRAM shortage, and whether to buy memory now in buy RAM now or wait.