# The Memory Supercycle, From the Buyer's Side

A 105 percent quarterly price jump, three suppliers sold out through 2027, a wafer crowd-out that turns 22 percent of capacity into 9 percent of output, and one small bank's 2027 technology budget written around all of it. What the file says about who pays for the AI buildout.

- Region: 999
- Updated: 2026-09-06
- URL: https://xecon.dev/999/deep-dives/the-memory-supercycle-from-the-buyers-side
- Type: Deep dive · Market mechanics

## The line item

In May 2025 a 32-gigabyte kit of DDR5 memory, the kind that goes into an ordinary office computer, sold for about US$95. In August 2026 the same kit sold for no less than US$375, and typically for US$399 to US$479. Between those two prices sits the largest quarterly increase in the recorded history of the memory business, three manufacturers who say their entire 2027 output is already spoken for, and a manufacturing arithmetic that most buyers have never had to think about.

This account begins with a document rather than a headline: the fiscal 2027 information-technology budget working file of a small United States bank, prepared for its board and reviewed in early September 2026. The file asks for a capital budget of no more than US$500,000 across six programs. One of those programs is unusual for a bank of any size: a US$50,000 reserve of server memory and storage, bought in the fourth quarter of 2026 and held as inventory, on the stated reasoning that the same parts will cost materially more in the first half of 2027. A bank was proposing to hold DRAM the way a bakery holds flour before a harvest fails.

That line item is the entry point for this account. The memory shortage of 2025 to 2027 is usually reported from the supply side, as a story about Samsung, SK Hynix, Micron, and the hyperscalers buying high-bandwidth memory for AI accelerators. Seen from the buyer's side, from a small bank that needs 30 memory modules and a few disks, the same shortage reads as a tax on everyone who computes, levied by a capacity decision made in three fabs. This account follows the file's evidence outward: what the prices did, why the wafers went where they went, who has committed to buy the output, what supply is coming and when, and what a small buyer can do about it.

[Preliminary conclusions] First, the price increase is documented and unusual in scale: 105 to 110 percent in a single quarter for contract PC DRAM in the first quarter of 2026, and roughly 4.8 to 5.8 times the mid-2025 retail price by August 2026. Second, the cause is structural rather than speculative: high-bandwidth memory consumes about three times the wafer area of conventional DDR5 per bit, so each point of capacity moved to HBM removes roughly three points from the commodity supply. Third, relief has a date attached, and the date is not 2027. The earliest new fabs start producing in February 2027 and mid-2027, and the analysts the file relies on put the end of undersupply in the second quarter of 2028 at the earliest. Fourth, a minority view holds that the shortage peaked in the second quarter of 2026 and will flip to a glut by 2028. This account gives that view its own chapter.

## The price, in three numbers

Three numbers describe the price move. The first is the contract price for PC DRAM, which TrendForce reports rose 105 to 110 percent quarter over quarter in the first quarter of 2026, the steepest quarterly increase on record. Contract prices are what module makers and computer manufacturers pay under negotiated agreements. A doubling inside one quarter means those agreements were being rewritten faster than the products built on them could be repriced.

The second number is retail. Tom's Hardware, tracking United States listings in August 2026, found the cheapest 32-gigabyte DDR5 kit at US$375, with typical kits between US$399 and US$479, against roughly US$95 in mid-2025. The third number comes from Germany, where TrendForce reported a DDR5 price index up 486 percent year over year in August 2026, close to five times the prior-year level, while Chinese spot prices were reported rising 14 percent in a single week. The move is global, and it is not confined to one channel.

(Chart: DDR5 32GB kit, observed retail and base-case path, 2024 to 2029, US$ per kit)
  - Observed to Q3 2026, base-case forecast thereafter: 95, 100, 110, 460, 439, 452, 380, 320, 240, 180, 110
  Source: Tom's Hardware (August 2026); TrendForce quarterly forecasts; bank budget file (September 2026); xecon analysis.

Two features of that line matter for what follows. The rise was not gradual: the price roughly quadrupled between the second quarter of 2025 and the first quarter of 2026, then held rather than fell. And the descent, as forecast, is slow. The base case in the file does not return the kit to its 2024 price until the second quarter of 2029, three years after the peak. Even a buyer who believes the forecast is buying into a market that stays expensive for the whole of 2027 and most of 2028.

The memory business has always been cyclical. Prices rose sharply in 2017 and 2018 and collapsed in 2019; they rose again in 2021 and fell in 2022 and 2023. What distinguishes this cycle is the mechanism. Previous shortages came from demand outrunning a supply that was, in principle, fungible: a wafer of DRAM was a wafer of DRAM. This cycle comes from a change in what a wafer produces.

## The crowd-out

High-bandwidth memory, or HBM, is the memory stacked next to an AI accelerator such as an Nvidia H200. It is made from DRAM dies, on the same wafers and in the same fabs as the DDR5 in an office computer, but stacked vertically, bonded through the silicon, and sold at a large premium to the hyperscale buyers who need it. The premium is why the three manufacturers have moved capacity toward it. The physics is why that move starves everyone else.

TrendForce's June 2026 estimate is that HBM will consume 22 percent of total DRAM wafer capacity in 2026 while producing only 9 percent of total DRAM bits, and that in 2027 the split widens to 30 percent of wafers for 13 percent of bits. Micron, one of the three manufacturers, has said publicly that the silicon gap between HBM and DDR5 is widening with every generation: HBM3E already needs roughly three times the wafer area of DDR5 for the same capacity. The arithmetic follows directly. Every percentage point of wafer capacity reassigned to HBM removes about three percentage points from the supply of conventional memory.

(Chart: HBM share of DRAM wafers versus share of DRAM bits, 2025 to 2027, % of total DRAM)
  - 2025 wafers: 18
  - 2025 bits: 7
  - 2026 wafers: 22
  - 2026 bits: 9
  - 2027 wafers: 30
  - 2027 bits: 13
  Source: TrendForce (June 2, 2026); Micron technical statements (August 2026); xecon analysis.

[The crowd-out arithmetic] Take 100 wafers of DRAM capacity. In 2024, almost all of them made conventional memory. In 2026, 22 of them make HBM, and those 22 wafers yield the bit equivalent of about nine conventional wafers. The other 78 wafers make conventional memory. Total bit output has fallen by roughly 13 percent relative to the 2024 allocation, before any change in demand. In 2027 the figure is 30 wafers to HBM and roughly 17 percent fewer conventional bits. This is the mechanism the file calls a wafer crowd-out, and it is the reason a shortage caused by AI accelerators shows up in the price of an office computer.

The manufacturers are not being irrational. HBM sells at a multiple of the price of commodity DRAM per bit, and the customers for it, the hyperscalers building AI infrastructure, are the only buyers in the market who are increasing volume. A manufacturer that kept its wafers on DDR5 in 2026 would have forgone the most profitable product in the history of the industry. The rational allocation for three suppliers is, for everyone downstream of them, a shortage.

The next generation makes the penalty worse. HBM4, expected in volume in 2027 and 2028, is projected to carry a wafer penalty of up to five times conventional DRAM, according to the industry analyst Nomad Semi as cited in the file's upside case. If that projection holds, the same 30 percent of wafers would yield a still smaller share of bits, and the crowd-out would deepen as the accelerators improve.

## Sold out

By the middle of 2026 the three DRAM manufacturers, Samsung, SK Hynix, and Micron, were reported to have booked and sold their entire 2027 output, with no additional DRAM or HBM available to new customers. The claim rests on company statements to customers and on reporting from the supply chain, and it has not been contradicted by any of the three. It means that a buyer who was not under contract by mid-2026 is, for 2027, a buyer in the spot market or the secondary market.

Micron's chief executive has said the company is currently able to fill only 50 to 66 percent of customer demand for DRAM and NAND. The chief executive of Apacer, a Taiwanese module maker, has warned that supply to module makers could fall by more than 70 percent year over year in 2027. Module makers are the layer between the fabs and the retail shelf; a 70 percent cut at that layer is what a US$375 floor looks like from the other side.

SK Hynix's chief executive, on the day the company listed on Nasdaq in July 2026, described 2027 as likely to be the worst year in the industry's history from a supply standpoint, and said the crunch could run to 2030. The statement is notable for its source. It was made by the executive who benefits most from the shortage, and it was made to investors, who could price it.

(Chart: DRAM demand growth versus supply growth, 2027 forecast, % year over year (UBS))
  - Demand: 36.2
  - Supply: 19.3
  - Gap: 16.9
  Source: UBS DRAM forecast (2026), as reported by WCCFtech; xecon analysis.

The bank's file leans on a UBS forecast for 2027: demand growth of 36.2 percent against supply growth of 19.3 percent, a gap of about 17 percentage points that UBS calls a record. TrendForce's own measure, a sufficiency ratio that nets supply against demand, sits at negative 1 to negative 2 percent for 2026 and worsens in 2027; UBS puts the end of undersupply no earlier than the second quarter of 2028. The two numbers describe the same market from different angles: the demand that is coming has already been counted, and the supply to meet it has not been built.

## Not only memory

The same buildout that drained the DRAM fabs has drained storage. The clearest single figure comes from an index maintained by the storage vendor VDURA: the price of a 30-terabyte enterprise solid-state drive rose from US$3,460 in the third quarter of 2025 to US$22,600 in August 2026, a 6.5-fold increase in 12 months. The drive is the same product. The buyers changed.

Consumer solid-state storage moved less but still moved: a price tracker maintained by CheapestSSD puts consumer NVMe Gen 4 storage at US$91 to US$138 per terabyte in September 2026, against US$38 per terabyte in mid-2023. Enterprise hard drives, the slow disks that hold the bulk of hyperscale data, are reported sold out at Western Digital for 2026, with 2027 and 2028 volumes already under contract, and prices up 46 to 50 percent between September 2025 and September 2026. TrendForce expects the NAND balance to turn positive in the second half of 2027; the hard-drive shortage runs longer.

(Chart: Component prices, August and September 2026, as a multiple of their baseline, × baseline price)
  - Ent. 30TB SSD: 6.5
  - DDR5 32GB kit: 4.8
  - Consumer NVMe /TB: 2.4
  - RTX 5090 street: 2.3
  - RTX Pro 6000 list: 1.87
  - Ent. HDD: 1.48
  Source: VDURA index via MLQ.ai; Tom's Hardware; CheapestSSD; BuySellRam; Dell product listing; DatacenterDisk; xecon analysis.

Graphics processors carry the shortage inside them. Memory is reported to be about 80 percent of the bill of materials of a high-end GPU, and the consumer RTX 5090, launched at a list price of US$1,999, was selling for US$4,381 to US$4,699 on the street in August and September 2026. Nvidia's professional card, the RTX Pro 6000 Blackwell with 96 gigabytes of GDDR7 memory, was listed at Dell for US$15,999.99 on September 4, 2026, against a launch list price of about US$8,565 in March 2025, an 87 percent increase that the file attributes to the GDDR7 supply constraint. The bank's file prices its inference accelerator, an Nvidia H200 NVL with 141 gigabytes of HBM3E, at US$33,500 from a reseller that listed it as out of stock with a two-week lead time from payment and no returns.

The pattern across all of these is the same. Any product whose cost is dominated by memory dies, whether the dies are DDR5, GDDR7, HBM, or NAND, has repriced upward by a multiple, and the multiple is largest where the hyperscale buyers compete most directly with everyone else.

## Who is buying, and who has stopped

The demand side has two halves, and they are moving in opposite directions. The half that is growing is machine demand. On OpenRouter, a marketplace that routes requests to language models, token consumption by autonomous agents grew 14-fold in the six months from February to August 2026 and now runs at 5.2 times the consumption of human users. Surveys of enterprise deployment put the share of companies running agents at production scale between 11 and 31 percent in the first half of 2026, depending on how strictly production scale is defined. Every one of those deployments is a memory buyer, directly or through a cloud provider.

The half that is shrinking is the consumer. IDC forecasts personal-computer shipments down 11.3 percent in 2026, with a 20 percent decline in the fourth quarter, and smartphone shipments down 16.7 percent to about one billion units, the steepest contraction in the record. The cause IDC names is the memory shortage itself: manufacturers cannot get the parts at a price the consumer will pay. The market has begun rationing by price, and the consumer is the buyer being rationed out.

Between those halves sits a third buyer that the file makes visible: the small institution that wants to run a language model on its own premises. The reason it can want that is a change in what a small model can do. The bank's file cites Qwen 3.6-27B, an open-weights model that fits in 32 gigabytes of accelerator memory, scoring 94.1 percent on the AIME 2026 mathematics benchmark against 95 percent for Gemini 3 Pro, a frontier closed model. Epoch AI's index of model capability puts the best open models about four months behind the closed frontier as of May 2026, and finds open models improving at 2.73 index points a month against 1.83 for the frontier.

(Chart: Frontier closed model minus best 32GB-class open model, by benchmark, points (percentage points; ECI for the index))
  - AIME math: 0.9
  - GPQA: 6.4
  - ECI index: 8
  - SWE-bench: 10.4
  - Composite: 11
  Source: BenchLM (September 2, 2026); Epoch AI (May 29, 2026); Stanford AI Index 2026; xecon analysis.

A model that fits in 32 gigabytes and is within a point of the frontier on mathematics is a model a bank can run on one accelerator card, inside its own perimeter, on documents it cannot send to an outside provider. That is why a bank's technology budget contains an accelerator, a training workstation, and a memory reserve. The capability arrived at the moment the memory to run it became scarce, and the two facts are connected: the same buildout that made the open models possible is the one consuming the wafers.

## The supply response, with dates

New capacity is coming, and the record is specific about when. SK Hynix has moved the opening of the first phase of its Yongin Y1 fab in South Korea forward to February 2027, starting at 20,000 wafers a month and adding 60,000 wafers a month every six months to reach 360,000 by the first half of 2030. Micron's Idaho Fab 1 is scheduled to produce its first wafers in mid-2027, with revenue-grade output realistically late 2027 or 2028 after qualification. Samsung's P5 fab is in the file's base case for 2028.

The Chinese manufacturer CXMT is the largest addition outside the three majors. SemiAnalysis projects its capacity rising from 265,000 wafers a month at the end of 2025 to 420,000 by the end of 2027, lifting its global DRAM share from 13 percent to about 17 percent. The same reporting tempers the effect: analysts estimate CXMT captured 11.3 percent of global DRAM market growth in 2026 but expect that share to fall to 6.9 percent in 2027, in part because export restrictions on extreme-ultraviolet lithography equipment cap how far the company can migrate to newer nodes.

(Chart: Two capacity additions, thousand wafer starts per month, thousand wafers per month)
  - CXMT end 2025: 265
  - CXMT end 2027: 420
  - Yongin Y1 Feb 2027: 20
  - Yongin Y1 H1 2030: 360
  Source: SemiAnalysis (March 2026); TrendForce (July 14, 2026); xecon analysis.

Three things about these dates matter for a buyer. The earliest new wafers arrive in February 2027, at 20,000 a month, a rounding error against the three majors' output. Fab output is not revenue-grade output; qualification adds two to four quarters. And a wafer added in 2027 is subject to the same allocation decision as every other wafer: if the manufacturer assigns it to HBM, it relieves the accelerator market, not the office computer. The file's base case assumes HBM's wafer share stays at or above 30 percent through 2027. Under that assumption, the new capacity through 2027 roughly offsets the crowd-out and does not reverse it.

## The other side of the record

Every memory cycle has ended in a glut, and there is a documented case that this one will end the same way, sooner than the base case allows. Its most prominent proponent is Shuli Ren, a Bloomberg Opinion columnist, who argued in July 2026 that the shortage peaked in the second quarter of 2026, that the spot-to-contract price spread of about 130 percent in early 2026 was the signature of panic buying rather than end demand, and that when new capacity from all three majors and CXMT lands in 2027 and 2028 while consumer demand stays weak, the market could flip to oversupply by 2028.

The precedent is real. In the 2016 to 2019 cycle, memory revenue fell about 30 percent peak to trough and the manufacturers' gross margins compressed from 59 percent to 27 percent. The mechanism in the downside case is also specific: original-equipment manufacturers and suppliers are carrying three to five weeks of inventory against a normal buffer of 10 to 15 weeks or more, and if spot prices fall those buyers will restock slowly, so demand does not rebound to absorb the new supply. In this reading, the consumer contraction IDC forecasts for 2026 is the condition that ends the shortage rather than a symptom of it.

The manufacturers' record cuts both ways too. The three majors and CXMT have between them announced or begun several hundred thousand wafers a month of new capacity, on published schedules, with equipment orders placed. The shortage is a shortage of built capacity, not of intent, and the intent is documented. What the downside case cannot show is a manufacturer moving wafers back from HBM to DDR5 while HBM still sells at a multiple; what the base case cannot show is consumer demand surviving three years of quintupled component prices intact. Both accounts rest on the record. The record does not yet choose between them.

(Chart: Three scenarios for DRAM supply through 2028, as weighted in the bank's file, % probability assigned)
  - Shortage to 2028: 60
  - Deeper shortage: 25
  - Early glut: 15
  Source: Bank budget file (September 2026); UBS; TrendForce; Goldman Sachs; Bloomberg Opinion; xecon analysis.

The file's own planners gave the early-glut case 15 percent, the base case 60 percent, and a deeper-shortage case 25 percent. The deeper case has its own named support: Goldman Sachs revised its 2027 DRAM shortfall estimate to 5.9 percent in July 2026, and the Apacer warning of a 70 percent cut to module makers belongs to it. Whatever the true weights, the asymmetry is the point. A buyer who prepays for memory and sees a glut in 2028 has overpaid by the difference between two prices. A buyer who waits and sees the base case pays the difference in the other direction, and may not be able to buy at all.

## The buyer's arithmetic

This is where the file becomes instructive, because it turns the macro record into a purchase decision with a date on it. The bank plans to buy 30 registered ECC memory modules of 64 gigabytes each, the DDR5-4800 parts that go into a server, for a virtualization host and a memory reserve. The reseller street price it documents is US$792 a module as of the second quarter of 2026, up from a range of US$520 to US$1,150 in March 2026. Its base case, drawn from the UBS forecast, puts the same module at US$1,100 in the first quarter of 2027, an increase of 38.9 percent across roughly six weeks of calendar; its upside case puts it at US$1,350, up 70.5 percent. A Citi projection the file cites goes further, to US$1,586 in the fourth quarter of 2026.

(Chart: 64GB DDR5 ECC registered server memory, unit price, observed and forecast, US$ per module)
  - Mar 2026 low: 520
  - Q2 2026 street: 792
  - Q1 2027 base: 1100
  - Q1 2027 upside: 1350
  - Q4 2026 Citi: 1586
  Source: Bank budget file, reseller pricing notes (September 2026); UBS; Citi; xecon analysis.

[The cost of waiting, as the file states it] Deferring the memory purchase from the fourth quarter of 2026 into the first quarter of 2027 costs the bank US$9,240 under its base case and US$16,740 under its upside case, on 30 modules.

The arithmetic is small and exact. Thirty modules at US$792 cost US$23,760. The same 30 modules at US$1,100 cost US$33,000, and at US$1,350 they cost US$40,500. The file states the cost of waiting one quarter as US$9,240 in the base case and US$16,740 in the upside case, and it adds a US$16,920 contingency for the first half of 2027 in case prices exceed the forecast ceiling. The whole program, memory and storage together, is US$50,000 out of a US$500,000 budget. Ten percent of a small bank's annual technology capital is being spent on holding inventory against a forecast, which is a thing banks are trained to avoid.

The file's reasoning for doing it anyway rests on three documented facts rather than on the forecast alone: 2027 capacity is sold out at all three suppliers, so the risk is availability as well as price; the parts in question are commodities with a long shelf life and a use in failure replacement regardless of the cycle; and the alternative, renting the same capacity from a cloud provider, carries its own escalation, since the same H100-class accelerators that rented for US$1.49 to US$6.88 an hour in September 2026 are built from the same scarce memory. The board is asked to approve the purchases in tranches, with a report on market conditions at each regular meeting, which is the mechanism by which the institution can stop buying if the downside case arrives.

One piece of the file can be tested against the shelf. On September 1, 2026, a large United States computer retailer listed the two 256-gigabyte four-module server-memory kits that fit the workstations in the file, a G.Skill T5 Neo DDR5-6400 ECC registered kit and a G.Skill Flare X5 DDR5-6000 kit, at zero units in stock, with pickup and shipping both unavailable and, for the registered kit, no price displayed while unavailable. A 192-gigabyte Corsair kit was listed at US$2,408.99 and out of stock. The shortage in the analyst reports is the shortage on the shelf.

What the file does not do is as telling as what it does. It does not attempt to buy HBM; no institution of its size can. It does not buy the accelerator ahead of need, because the accelerator is the one component whose price the file expects to compress as newer parts arrive. And it caps its own exposure: the reserve is sized to the bank's 2027 capacity plan and its failure-replacement history, not to a bet on the cycle. A buyer with more capital could take a larger position. This buyer took the smallest position that covers its operating need and documented why.

## Rent or hold

The obvious alternative to holding memory is renting it. The file prices that alternative too. In September 2026 an Nvidia H100 accelerator with 80 gigabytes of memory rented for US$1.99 an hour on RunPod's community tier, US$3.99 on Lambda, US$4.25 on CoreWeave, and US$6.88 on Amazon Web Services; an H200 with 141 gigabytes rented for US$3.89 to US$4.49 an hour. Against a purchase price of US$33,500 for the H200 NVL, the rental break-even is between about 7,500 and 8,600 hours, which is 10 to 12 months of continuous use, or two to three years at a working-day duty cycle.

On that arithmetic alone a small buyer should rent. The file chooses to buy for two reasons it states and one it implies. The stated reasons are that the workloads in question, compliance review and document processing on customer records, must stay inside the bank's own perimeter under its examiners' expectations, and that rental rates are themselves built on the same scarce memory and carry the same escalation risk with none of the residual value. The implied reason is availability: a rented accelerator can be withdrawn or repriced at the end of any term, and in a market where 2027 output is sold out, a small tenant is the first to be displaced.

The memory reserve follows the same logic at a smaller scale. Thirty modules of server memory held in a cabinet have no rental equivalent; they are either on hand when a host fails or a capacity expansion is scheduled, or they are bought at the spot price of that day. The file treats them as an insurance premium of US$23,760, paid once, with a documented expected saving of US$9,240 to US$16,740 if the forecast holds and a bounded loss if it does not. That is the whole of the buyer's position, and it is a modest one.

## Two scenarios

The account can end where the file ends, with the two cases a buyer must plan against. In the first, the base case that the file and most of its sources weight most heavily, the three manufacturers keep at least 30 percent of DRAM wafers on HBM through 2027, the new fabs at Yongin and Idaho ramp on schedule but into qualification, undersupply persists to the second quarter of 2028, and conventional memory prices ease slowly toward their 2024 level by 2029. In that case, every institution that computes pays a premium for three years, the consumer market shrinks to fit, and a small buyer's only leverage is timing: buy before the first quarter of 2027, in tranches, and hold.

In the second, the early-glut case, the shortage has already peaked, the spot-to-contract spread was panic rather than demand, the new capacity lands into a consumer market that has stopped buying, and by 2028 the manufacturers are dumping inventory into a market that undershoots the 2024 baseline. In that case, the buyer who bought in the fourth quarter of 2026 overpaid by roughly the difference between US$792 and the 2028 price of a module, on 30 modules, and holds parts it would have used anyway.

The third course, doing nothing and buying at need in the first half of 2027, is the one the file prices at US$9,240 to US$16,740 on memory alone, before the question of whether the parts can be had. That is the course most small institutions are on by default, because most of them have not written the file. The AI buildout's bill is being presented to every buyer of memory in the world. The only choice on the buyer's side is whether to read it before it arrives.

## Sources

1. [Tom's Hardware, "32GB of DDR5 now costs $375 minimum: AI shortage continues to squeeze PC building," citing TrendForce contract-price data, August 2026.](https://www.tomshardware.com/pc-components/ddr5/32gb-of-ddr5-now-costs-usd375-minimum-ai-shortage-continues-to-squeeze-pc-building)
2. Fiscal 2027 information-technology budget working file of a small United States bank (board proposal, unpublished), with evidence register, price schedules, and three-scenario forecast; reviewed September 4, 2026. The institution is not named.
3. [Tom's Hardware, "Micron says the silicon gap between HBM and DDR5 is widening with every generation," August 2026.](https://www.tomshardware.com/tech-industry/semiconductors/micron-says-the-silicon-gap-between-hbm-and-ddr5-is-widening-with-every-generation)
4. [TrendForce press release, DRAM sufficiency ratio and 2026 to 2028 outlook, July 30, 2026; UBS undersupply timeline as cited therein and in source 10.](https://www.trendforce.com/presscenter/news/20260730-13158.html)
5. [TrendForce, "Germany DDR5 prices near 5x YoY in August; China reportedly sees 14% WoW jump as global rally continues," August 17, 2026.](https://www.trendforce.com/news/2026/08/17/news-germany-ddr5-prices-near-5x-yoy-in-august-china-reportedly-sees-14-wow-jump-as-global-rally-continues/)
6. [TrendForce press release, HBM share of DRAM wafer capacity and bit output, 2025 to 2027, June 2, 2026.](https://www.trendforce.com/presscenter/news/20260602-13074.html)
7. [TweakTown, "Memory capacity for all of 2027 has reportedly been booked and sold, with no more DRAM or HBM available," June 2026, citing supplier statements.](https://www.tweaktown.com/news/113004/memory-capacity-for-all-of-2027-has-reportedly-been-booked-and-sold-with-no-more-dram-or-hbm-available/index.html)
8. [Tom's Hardware, "Memory price surge begins to cool as consumers hit affordability limit; AI demand still keeps DRAM and NAND prices climbing through Q3 2026," including Micron fulfillment rate and Apacer chief executive's warning, 2026.](https://www.tomshardware.com/pc-components/ram/memory-price-surge-begins-to-cool-as-consumers-hit-affordability-limit-ai-demand-still-keeps-dram-and-nand-prices-climbing-through-q3-2026)
9. [Tom's Hardware, "SK Hynix says 2027 will be the worst year for memory shortage, forecasts crunch to last until 2030," remarks by the SK Hynix chief executive on the company's Nasdaq listing day, July 10, 2026.](https://www.tomshardware.com/pc-components/dram/sk-hynix-says-2027-will-be-the-worst-year-for-memory-shortage-forecasts-crunch-to-last-until-2030-ceo-shares-grim-outlook-on-the-day-sk-hynix-gets-listed-on-nasdaq)
10. [WCCFtech, "UBS doubles its DRAM price forecast as memory demand outpaces supply by 17 points through 2027," 2026.](https://www.wccftech.com/ubs-doubles-its-dram-price-forecast-as-memory-demand-outpaces-supply-by-17-points-through-2027/)
11. [MLQ.ai, "AI storage squeeze lifts 30TB enterprise SSD to $22,600 while HDD capacity is booked years ahead," citing the VDURA pricing index, August 1, 2026.](https://mlq.ai/news/ai-storage-squeeze-lifts-30tb-enterprise-ssd-to-22600-while-hdd-capacity-is-booked-years-ahead/)
12. [CheapestSSD, SSD price tracker, consumer NVMe Gen 4 price per terabyte, September 2026.](https://cheapestssd.com/ssd-price-tracker/)
13. [DatacenterDisk, "Hard drive prices up 50 percent in 2026," including Western Digital and Seagate contract status, September 3, 2026.](https://datacenterdisk.com/news/hard-drive-prices-up-50-percent-2026)
14. [Club386, "SSD pricing chaos could end in 2027," citing TrendForce NAND balance forecast, 2026.](https://www.club386.com/ssd-pricing-chaos-could-end-2027/)
15. [BuySellRam, "Nvidia consumer GPU price report, August 2026," street prices and memory share of bill of materials.](https://www.buysellram.com/blog/nvidia-consumer-gpu-price-report-august-2026/)
16. [Dell, product listing for the Nvidia RTX Pro 6000 Blackwell Max-Q Workstation Edition, US$15,999.99, captured September 4, 2026; March 2025 launch list price per Nvidia announcement.](https://www.dell.com/en-us/shop/nvidia-rtx-pro-6000-blackwell-max-q-workstation-edition-300w/apd/490-bldl/graphic-video-cards)
17. [ViperaTech, product listing for the Nvidia H200 NVL 141GB PCIe accelerator, US$33,500, out of stock, captured September 4, 2026.](https://viperatech.com/product/nvidia-h200-nvl-graphic-card-141-gb-passive-pcie-900-21010-0040-000)
18. [Winzheng Analytics, "AI agents surpass humans in token usage: 14x growth on OpenRouter in 2026," August 10, 2026.](https://www.winzheng.com/en/article/ai-agents-surpass-humans-token-usage-14x-openrouter-2026)
19. [Enterprise AI agent statistics 2026, compiling McKinsey (January 2026) and LumiChats (March 2026) survey data on production-scale deployment.](https://paul-okhrem.com/enterprise-ai-agents-statistics-2026/)
20. [IDC, "PC market enters volatile territory as memory shortage persists through 2027," and IDC smartphone forecast, June and July 2026.](https://www.idc.com/resource-center/blog/pc-market-enters-volatile-territory-as-memory-shortage-persists-through-2027/)
21. [BenchLM, model comparison, Qwen 3.6-27B versus Gemma 4 31B, with frontier reference scores, September 2, 2026.](https://benchlm.ai/compare/gemma-4-31b-vs-qwen3-6-27b)
22. [Epoch AI, "The gap between open and closed models on the Epoch Capabilities Index," May 29, 2026.](https://epoch.ai/data-insights/open-closed-eci-gap)
23. [TrendForce, "SK Hynix reportedly starts Yongin Y1 equipment orders, moves up phase 1 opening to February 2027," July 14, 2026.](https://www.trendforce.com/news/2026/07/14/news-sk-hynix-reportedly-starts-yongin-y1-equipment-orders-moves-up-phase-1-opening-to-feb-2027/)
24. [Tom's Hardware, "Micron details new U.S. fab projects: Idaho Fab 1 comes online in 2H 2027, New York fabs come later," 2026.](https://www.tomshardware.com/pc-components/dram/micron-details-new-u-s-fab-projects-idaho-fab-1-comes-online-in-2h-2027-new-york-fabs-come-later-hbm-assembly-in-the-u-s)
25. [SemiAnalysis, "China's CXMT is set to challenge DRAM," capacity and share projections, March 2026.](https://newsletter.semianalysis.com/p/chinas-cxmt-is-set-to-challenge-dram)
26. [WCCFtech, "CXMT grabbed 11.3% of DRAM growth in 2026, but analysts say China's momentum collapses to 6.9% by 2027," July 2026.](https://wccftech.com/cxmt-grabbed-11-3-of-dram-growth-in-2026-but-analysts-say-chinas-momentum-collapses-to-6-9-by-2027/)
27. [Shuli Ren, Bloomberg Opinion, on the memory boom and the case for an early peak, July 2026, as syndicated by Yahoo Finance.](https://finance.yahoo.com/technology/ai/articles/wait-michael-burry-memory-boom-151825822.html)
28. Citi Research DRAM price projection for the fourth quarter of 2026, as cited in the reseller pricing notes of source 2 (82 percent year-over-year increase for 64GB DDR5-4800 ECC RDIMM).
29. Retail inventory check, United States computer retailer listings for G.Skill T5 Neo 256GB DDR5-6400 ECC RDIMM kit, G.Skill Flare X5 256GB DDR5-6000 kit, and Corsair Vengeance 192GB DDR5-5200 kit, September 1, 2026.