Topicsmemory

Memory: How DRAM Works, Who Makes It, and Why It Costs What It Does

A sourced reference collection on DRAM — how the cell actually works and is built, how three companies came to make almost all of it, why the industry moves in cycles, and why AI repriced the memory in every device you own.

Assertions
47
Sources consulted
63
Read in full
25/63
Cited as evidence
25
Disputed
1

63 sources sit behind this page — including any that arrive with a concept this page shares with another collection. 25 were retrieved and read in full, and only those can back an assertion. 3 could not be retrieved, and 35 were surfaced and deliberately set aside. Every one of them is named in the register below, with the reason in view. How we source this.

Background

Our own synthesis, written to orient you — not evidence. Every factual statement here is asserted and sourced further down this page.

Almost every computer built since 1970 stores its working memory the same way: one bit as a smear of electric charge on a tiny capacitor, with one transistor deciding when that charge can be touched. Robert Dennard worked it out at IBM in 1966. The patent was granted in 1968. Everything below — the prices, the shortages, the fact that three companies make nearly all of it — comes out of that one structural choice.

The choice was economy. A DRAM bit needs one transistor and one capacitor where the faster SRAM alternative needs four or six, so DRAM could be made dense and made cheap. The cost is that charge leaks. Every row has to be read and rewritten within about 64 milliseconds or the data is simply gone, which on a device with 8,192 rows means refreshing one row roughly every 7.8 microseconds, forever, whether or not anyone is using it. That is what the word dynamic in the name is doing.

Making one is largely a problem of building a capacitor that holds enough charge in almost no area. Manufacturers build it vertically — etched down into the silicon, or stacked upward over the transistor — with depth-to-width ratios that had already passed 50:1 by the mid-2000s. Shrinking the cell shrinks the capacitor, and by the 1z and 1-alpha generations capacitance had fallen under 10 femtofarads per cell against a floor of about 6 or 7 that manufacturers try to hold. That wall is why progress has slowed: node transitions now buy only 10 to 15% more density each, against the 40% Micron reported moving to 1-alpha in 2021.

Who makes it has changed twice, and both times a downturn did the choosing. Japanese firms held around 80% of the market in the 1980s on yields reported 20 to 30% better than American rivals; currency moves and trade agreements helped push the United States out of commodity memory, and Intel went to microprocessors instead. South Korea entered in 1983, reached the top of the market by 1992, and has not been displaced. The 1997 Asian crisis produced SK hynix; Japan answered with Elpida in 1999; the 2008 crisis took Qimonda; and Micron’s purchase of Elpida in 2013 closed the process. Three companies now hold about 95% of DRAM revenue.

That concentration is why this market moves in cycles rather than curves. A fab is committed years before it ships anything, so supply always arrives on a decision made in a different world — downturns in 2011, 2019 and 2022. What changed recently is discipline: the three survivors now cut production in a downcycle instead of expanding into it.

Which brings the story to now. High Bandwidth Memory, the stacked DRAM that feeds an AI accelerator, is made on the same wafers in the same fabs as the memory in a laptop, but consumes far more wafer area per usable bit and sells for five to ten times as much per bit. TrendForce projects HBM taking about 18%, 22% and 30% of DRAM wafer input at the ends of 2025, 2026 and 2027 while returning only about 8%, 9% and 13% of the bits. The gap between those two series is conventional memory that never gets made — and with no fourth supplier of consequence, it reprices memory for everyone.

What is genuinely contested, and what is missing, is set out beside each claim rather than hidden: the retrieved sources describe contract-price forecasts rather than settled transaction prices; IBM and IEEE Spectrum disagree by a year on when Dennard had the idea; and Micron's figures now come from its SEC-filed results, while the remaining manufacturing gaps include imec's peripheral-transistor platform, an open-access HfO₂ trench-capacitor paper that was only surfaced in search, and an EUV-in-DRAM conference presentation that has since been read in full, closing that gap.

Figures

Every number below is asserted and sourced elsewhere on this page.

The crowding-out, measured

HBM takes a rising share of the wafers while returning a much smaller share of the bits. The distance between the two bars is conventional memory that never gets made.

18%
8%

10 pts

End 2025

22%
9%

13 pts

End 2026

30%
13%

17 pts

End 2027

The gap widens from 10 to 17 percentage points — that widening is the squeeze.

TrendForce, 2 June 2026. Both series are projections for the end of each year, not measured outcomes — the end-2026 and end-2027 points had not happened when they were published.

What the AI cycle did to one manufacturer

Micron's quarterly revenue, from its own filings. The final column is company guidance, not a result.

Revenue, billions of US dollars

$9.3bn

FQ3-25

$23.9bn

FQ2-26

$41.5bn

FQ3-26

$50.0bn

FQ4-26 guide

Micron Form 8-K Exhibit 99.1 filed with the SEC, 24 June 2026, for the quarter ended 28 May 2026. The first three columns are reported figures; FQ4-26 is Micron's own guidance of $50.0bn ± $1.0bn and had not happened when published.

Who actually buys the memory

Micron's revenue by business unit for the quarter ended 28 May 2026. Data centre in its two forms is roughly 61% of the total; everything a consumer holds is about 28%.

Total $41.4bn

  • Cloud Memory$13.8bn
  • Core Data Center$11.5bn
  • Mobile and Client$11.5bn
  • Automotive and Embedded$4.6bn

Micron Form 8-K Exhibit 99.1 filed with the SEC, 24 June 2026. Micron's own segment definitions, reported figures, one company of three — not an industry-wide breakdown.

Three companies, and the order changed

DRAM revenue share in the third quarter of 2025. SK hynix had passed Samsung, which had held 42.9% only five quarters earlier.

Total $41.4bn industry revenue

  • SK hynix33.2%
  • Samsung32.6%
  • Micron25.7%
  • Nanya, Winbond, PSMC and othersrest

TrendForce, 26 November 2025, for the completed third quarter of 2025 — measured revenue rather than a forecast, but one research firm's estimate of industry totals, not audited figures.

Concepts

The vocabulary this subject is built from, and what we can show about each.

DDR Generations

memory technology

DDR5 was published by JEDEC on 14 July 2020, splitting each DIMM into two independent 32-bit subchannels, moving voltage regulation onto the module via a PMIC, and making on-die ECC standard on every chip; SK hynix launched the first production DDR5 chip on 6 October 2020.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read
  • SupportsQualifiesRetrieved & read
    DDR5 SDRAM

    WikipediaOther

DDR generations re-standardise the interface, not the cell: DDR (1998, 200-400 MT/s, 2.5V), DDR2 (2003, 1.8V), DDR3 (2007, 1.5V), DDR4 (2014, 1.2V), DDR5 (2020, 3200-7200 MT/s, 1.1V), with prefetch doubling 2n to 16n each generation. None is compatible with another.

ReportedSupported by the sources below, not yet editor-reviewed.
3 sources3 retrieved & read

DRAM (Dynamic Random-Access Memory)

memory technology

DRAM production is highly concentrated: as of Q2 2025 Samsung, SK hynix and Micron together held 93.4% of the global DRAM market, with all other suppliers including CXMT accounting for 4.9%.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

J.P. Morgan estimated in August 2026 that DRAM prices would rise more than 400% from the start of 2024 to the end of 2026, alongside measured US index moves of +23% (CPI software and accessories, PPI storage devices) and +37% (import price index for computers and parts) since end-2024.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

TrendForce forecast conventional DRAM contract prices to rise 58-63% quarter-on-quarter in 2Q26 (published 31 March 2026) but only 13-18% in 3Q26 (published 3 July 2026), citing a higher base and weakening consumer demand — both projections, not measured outcomes.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

DRAM Cell (1T1C)

component

DRAM capacitors leak, so each row must be refreshed within 64 milliseconds — about one row every 7.8 microseconds on an 8,192-row device. That requirement is what 'dynamic' names, and it costs power and bandwidth permanently.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

A DRAM cell is one capacitor holding a bit as charge plus one access transistor, against four to six transistors per bit in SRAM — the structural choice that made DRAM cheap and dense, and from which its refresh requirement follows.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

DRAM Equipment and Materials Supply Chain

other

A 2021 Micron and imec presentation places EUV in DRAM at metal routing and dense pillars and contacts — the storage node landing pad and bit line pad — with Samsung shipping bit line pad patterning at D1x and D1z, one EUV level replacing five immersion masks, and productivity and defects named as the biggest obstacles.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

A single DRAM process step can rest on three companies most readers cannot name: for EUV dry resist, Lam Research supplies equipment, Entegris the precursor chemicals and materials with dual-sourced capacity, and Gelest the specialty precursors — a vapour-deposited resist Lam states uses five to ten times less raw material than spin-on.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

DRAM Market Segments

other

The same fabs serve unequal buyers: capacity moved to servers with cloud providers prioritised for high-capacity RDIMMs, while memory is 15-20% of a mid-range phone's bill of materials against 10-15% of a flagship's — and HBM sits above both at five to ten times DDR5's per-bit price.

ReportedSupported by the sources below, not yet editor-reviewed.
3 sources3 retrieved & read

DRAM Market Structure

other

Samsung Semiconductor was founded in 1978, Korea developed 64-kilobit DRAM by the end of 1983 as the third country to do so, reached the top of the DRAM market by 1992, and held 70.5% of it in 2022 — with SK hynix at 36% passing Samsung's 34% in Q1 2025.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

The three-supplier DRAM market is the residue of consolidation forced by successive downturns: Japan at ~80% share in the 1980s, the 1985 Plaza Accord and US exit, the 1997 Asian crisis creating SK hynix and Elpida, Qimonda's insolvency in the Great Financial Crisis, and the 2012-2013 acquisitions of Hynix and Elpida.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

DRAM Process Node

process

DRAM generations are named 1x/1y/1z then 1-alpha/1-beta/1-gamma rather than in nanometres; Micron reports 1-alpha from 2021 (+40% density over 1z) and 1-beta from 2022 (16Gb per die, ~35% better density), while independent analysis puts current node migration at only 10-15% density gain per generation.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

DRAM Storage Capacitor

component

A DRAM capacitor is built vertically to hold enough charge in almost no area — etched down as a trench capacitor or built upward as the stacked capacitor current manufacturers use — with depth-to-width ratios past 50:1 by the mid-2000s and cell areas of roughly 6 to 8 F squared.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Cell capacitance fell below 10 femtofarads by the 1z and 1-alpha generations against a 6-7 fF floor, and a TechInsights analyst assessed that 10 nm may be the last node for the conventional 6F2 cell — with capacitorless IGZO designs a post-2028 candidate.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

The peripheral logic and interconnect around the cell array must scale too or it eats the die; DRAM has used silane or TEOS silicon oxides as the interconnect dielectric for about 25 years, and thinning it further causes capacitive coupling — the wall logic hit twenty years earlier and answered with copper low-k.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

A DRAM capacitor is a metal-insulator-metal stack inside a deep hole — a titanium nitride bottom electrode, a high-k dielectric, a second electrode — with charge proportional to the hole's inner and outer surface area; the etch is limited by an amorphous-polysilicon hard mask that erodes as it cuts, and can be consumed before the hole is finished.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

DRAM Wafer Crowding-Out

process

HBM was projected to consume about 18%, 22% and 30% of total DRAM wafer input at the ends of 2025, 2026 and 2027 while contributing only about 8%, 9% and 13% of total DRAM bit supply — the widening gap between wafer share and bit share is the crowding-out of conventional DRAM, measured directly.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

High Bandwidth Memory (HBM)

memory technologyshared from another collection — see its own page for what it asserts

Memory Cost Pass-Through

other

IDC put memory at 15-20% of a mid-range smartphone's bill of materials and 10-15% of a flagship's, projecting 6-8% handset price rises and a 2026 market decline of 2.9% (base case) to 5.2% (pessimistic) — with entry-level devices, being the most price-elastic, hit hardest.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

J.P. Morgan estimated the memory price shock added 0.2-0.4 percentage points to inflation, with every 10% hardware cost increase lifting core CPI and PCE by roughly 0.1 points, and noted Meta raising capital expenditure guidance by $10 billion citing higher AI hardware and memory costs.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Memory Industry Revenue

other

Micron's quarter ended 28 May 2026 split as Cloud Memory $13,769 million, Core Data Center $11,524 million, Mobile and Client $11,521 million and Automotive and Embedded $4,634 million — data centre roughly 61% of revenue against about 28% consumer and 11% automotive and embedded.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Micron reported revenue of $41.456bn for the quarter ended 28 May 2026 against $9.301bn a year earlier, with GAAP gross margin rising from 37.7% to 84.6%, operating cash flow of $25.39bn and net capital expenditure of $7.1bn, guiding the next quarter to $50.0bn ± $1.0bn.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

DRAM industry revenue went from $22.9bn in 2Q24 (+24.8% on the quarter) to $41.4bn in 3Q25 (+30.9%), and the lead changed hands: SK hynix 33.2% share on $13.75bn, Samsung 32.6% on $13.5bn and Micron 25.7% on $10.65bn, up 3.7 points — Samsung did not shrink, the market grew faster around it.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

Memory Long-Term Agreement (LTA)

other

Multi-year long-term agreements have displaced quarterly contract negotiation in memory: SK hynix stated in July 2026 it had finalised LTAs with around ten customers, and J.P. Morgan reports some hyperscaler agreements run five years or longer — committing capacity years ahead and away from buyers who do not sign them.

ReportedSupported by the sources below, not yet editor-reviewed.
3 sources3 retrieved & read

NAND Flash

memory technology

TrendForce forecast NAND flash contract prices to rise 70-75% quarter-on-quarter in 2Q26 and 10-15% in 3Q26, driven by enterprise SSD demand for AI data centres rather than by HBM substitution.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

SK hynix stated in July 2026 that its 321-layer NAND products already represented the largest share of its total production, with roughly 50% of domestic capacity targeted on that node by year end.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

The Memory Cycle

other

Independent analysis puts HBM at five to ten times the per-bit price of DDR5 — the premium that pulls capacity toward HBM faster than a conventional-memory shortage can bid it back.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Memory is cyclical because fabs are committed years before output ships, so supply arrives on decisions made in a different demand environment; downturns are identified in 2011, 2019 and 2022, and the three remaining makers now cut production in downcycles rather than expanding reflexively.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

Timeline

What actually happened, in order, with sources.

Coverage
  • 6 United States
  • 5 South Korea
  • 3 Japan
  • 1 China
  • 1 Germany
  • 3 Unattributed

Where this topic’s events took place, as far as our sources establish it. Events with no single location — a standards publication, say — and events we have not yet attributed are both counted as unattributed rather than omitted.

  1. 2026

    SK hynix Begins Mass Shipments of HBM4

    shipment milestoneSouth KoreaAsia

    SK hynix stated in its 29 July 2026 results announcement that it began mass shipments of HBM4 during the second quarter of 2026, with a full production ramp planned for the second half of the year.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  2. 2026

    SK hynix Finalises Long-Term Supply Agreements with Around Ten Customers

    partnershipSouth KoreaAsia

    SK hynix stated on 29 July 2026 that it had finalised long-term supply agreements with around ten customers, without naming them or disclosing durations or volumes.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  3. Nov 23, 2025

    CXMT Launches China's First Domestically Developed DDR5 Memory

    technology generation milestoneChinaAsia

    CXMT launched China's first domestically developed DDR5 DRAM on 23 November 2025 at the China International Semiconductor Expo — up to 8,000 Mbps, 24Gb maximum single-die capacity, seven module variants, with a 96GB module targeted for mass production in 2026.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  4. 2021

    Micron Ships Its First 1-Alpha Node DRAM

    technology generation milestoneUnited StatesNorth America

    Micron released 1-alpha node DRAM in 2021, reporting a 40% density improvement over 1z and about 15% power savings on LPDDR5 — a larger step than the 10-15% per generation independent analysis reports for current nodes.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  5. Jul 14, 2020

    JEDEC Publishes DDR5

    standards milestone

    JEDEC published DDR5 on 14 July 2020 — two independent 32-bit subchannels per DIMM, on-module voltage regulation via a PMIC, and on-die ECC standard on every chip; SK hynix shipped the first production part on 6 October 2020.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
    • SupportsRetrieved & read
      DDR5 SDRAM

      WikipediaOther

  6. 2014

    DDR4 Arrives at 1.2 Volts

    standards milestone

    DDR4 arrived in 2014 at 1600-3200 MT/s and 1.2 volts, down from DDR3's 1.5V, and held the mainstream for six years — the longest run of any generation so far.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
    • SupportsRetrieved & read
      DDR SDRAM

      WikipediaOther

  7. 2013

    Micron Acquires Elpida, Completing the Three-Supplier Market

    otherUnited StatesNorth America

    Micron acquired Elpida in 2013, a year after SK Telecom acquired Hynix for a reported $3 billion — the two transactions that closed four decades of consolidation into three suppliers holding roughly 95% of global DRAM revenue.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  8. 2009

    Qimonda Files for Insolvency

    otherGermanyEurope

    Qimonda filed for insolvency protection in the year after Micron's 2008 purchase of a 35.5% stake in Inotera, removing Europe's last major DRAM maker and leaving four players — Samsung, Hynix, Micron and Elpida.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  9. Jun 2000

    JEDEC Finalises the First DDR SDRAM Standard (JESD79)

    standards milestone

    JEDEC finalised the first DDR SDRAM specification, JESD79, in June 2000, formalising an interface that transfers on both clock edges — the basis every generation through DDR5 revises.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
    • SupportsRetrieved & read
      DDR SDRAM

      WikipediaOther

  10. 1999

    NEC and Hitachi Fold Their DRAM Operations into Elpida

    otherJapanAsia

    NEC and Hitachi folded their DRAM operations into Elpida in 1999, after the 1997 Asian financial crisis that had already forced the LG Semicon and Hyundai Electronics merger in Korea which became SK hynix.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  11. 1992

    South Korea Reaches the Top of the DRAM Market

    otherSouth KoreaAsia

    Korean industry reached the top of the global DRAM market in 1992, about nine years after its first 64-kilobit part, and had not been displaced by 2022, when it held 70.5% of the market.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  12. 1986

    The US-Japan Semiconductor Agreement

    otherJapanAsia

    The 1986 US-Japan Semiconductor Agreement had Japan maintain minimum chip prices and raise the foreign share of its domestic market from 10% to 20%; the US imposed 100% punitive tariffs on certain Japanese memory chips in 1987.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  13. 1983

    South Korea Develops 64-Kilobit DRAM

    technology generation milestoneSouth KoreaAsia

    By the end of 1983 South Korea had developed 64-kilobit DRAM, the third country to do so after the United States and Japan, five years after Samsung Semiconductor was founded in 1978.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  14. 1976

    Japan Launches the Super LSI Project

    otherJapanAsia

    Japan established the Super LSI project in 1976, uniting Fujitsu, Hitachi, NEC, Mitsubishi Electric, NTT and Toshiba; within four years it had developed advanced electron-beam lithography, and Japan went from about 15% of the global semiconductor market in the early 1970s to roughly 50% in the 1980s.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  15. 1973

    Mostek's MK4096 Introduces Multiplexed Addressing

    technology generation milestoneUnited StatesNorth America

    Mostek introduced the 4-kilobit MK4096 with multiplexed addressing in 1973, keeping pin counts manageable as density rose; its 16-kilobit successor the MK4116 reached about 75% market share by 1976.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read
  16. Oct 1970

    Intel Releases the 1103, the First Commercial DRAM

    production milestoneUnited StatesNorth America

    Intel released the 1103 in October 1970 — the first commercially available DRAM, holding 1,024 bits, and a three-transistor design rather than Dennard's one-transistor cell. It displaced magnetic core memory within about two years.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  17. Jun 4, 1968

    The DRAM Patent Is Granted

    standards milestoneUnited StatesNorth America

    The patent on Dennard's single-transistor DRAM cell was granted on 4 June 1968 — a date given by Micron's account, where IBM's own page states only the year. No source retrieved gives the patent number.

    ReportedSupported by the sources below, not yet editor-reviewed.
    2 sources2 retrieved & read
  18. 1966

    Robert Dennard Invents the One-Transistor DRAM Cell

    technology generation milestoneUnited StatesNorth America

    Robert Dennard conceived the one-transistor DRAM cell at IBM in 1966, storing a bit as charge on a capacitor with a single access transistor against six per bit in competing designs — though IEEE Spectrum dates the conception to 1967.

    DisputedSources disagree. Both accounts are shown below.
    2 sources2 retrieved & read
  19. Date unknownno source establishes an occurrence date

    Samsung Accelerates Pyeongtaek P5 Mega-Fab Expansion

    capacity announcementSouth KoreaAsia

    Samsung advanced its Pyeongtaek P5 mega-fab expansion by at least six months; P5 Fab 1 and Fab 2 combined are reported to reach roughly 600,000 twelve-inch wafers per month, against Samsung's current DRAM output of about 650,000 per month.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read

Source register

All 63 sources behind this page — what we read, what we tried to read and could not, and what we looked at and set aside, with the reason in view for each. A concept shared with another collection brings its own references with it, so some entries here were surfaced for a neighbouring topic rather than this one.

Cited as evidence
25
Tried, could not read
3
Surfaced, set aside
35
Cited sources 25 distinct links

Original publisher links. Files open on the publisher’s site; we do not host copies. A linked document is not an additional source or an independent verification.

Tried, could not read3

We attempted these and were refused or served nothing. Nothing on this page rests on them; they are published so the gaps are checkable rather than invisible.

Surfaced, set aside35

These came up while researching and were deliberately not used. We do not claim to have read them — each is listed with why it was passed over, so the shape of the survey is visible and not just its conclusions.

Coverage & limits

What this page does and does not claim.

Third packet through the generic ingestion pipeline, substantially expanded in Sprint 4C from a two-year news window into the subject itself. Twenty-five sources were retrieved and read; another thirty-eight are disclosed as non-evidentiary references, split into three tried-and-blocked items and thirty-five surfaced-and-set-aside items. The EUV-in-DRAM conference presentation that earlier revisions named as an unparseable gap has since been read in full, by decompressing the PDF's own content streams when no text extractor was available, and it now supplies the layer-by-layer account of where EUV is actually used in a DRAM. Of the sources still unread, only three were ever refused: Gartner sits behind a human-verification challenge that this project will not attempt to defeat, one Tom's Hardware article is paywalled, and a DatacenterDynamics piece has since been withdrawn and now returns 404. The remaining thirty-five were never fetched at all — they were surfaced and judged redundant, paywalled, or unreconcilable, and each says which. SEC EDGAR and Applied Materials both refused the automated fetcher but were later read through ordinary browser access to their public pages; the Micron filing closes the collection's former financial-attribution gap, and the Applied Materials account closes its former process-materials gap. The remaining named manufacturing gaps are imec's peripheral-transistor platform, an open-access HfO₂ ALD trench-capacitor paper that was only surfaced in search, and an EUV-in-DRAM conference presentation that was retrieved but initially unparseable, and has since been read in full. Two sources disagree by a year on when Dennard conceived the DRAM cell, and that disagreement is published rather than resolved. Market-structure history rests on an independent analyst newsletter, which is not a primary record of market share and is labelled as such. Geographically this collection reaches the United States, Japan, South Korea, Germany and China across sixty years. Every price figure is a forecast issued on a stated date by a named forecaster, never a measured outcome. Not yet editor-reviewed; every assertion reads as reported.

Source check, 2026-09-17. Numeric-presence checks passed for 47 assertions using available source text, which may be cached. This is not verification of their meaning. What this check does and does not prove →

  • Not editor-reviewed unless labelled. Assertions marked Reported are assembled from the sources shown and have not yet been checked by an editor. Only Primary source and Corroborated mean a human verified them.
  • Disagreements are preserved, not resolved. Where sources conflict, both accounts appear and the assertion is marked Disputed.
  • Retrieval status is disclosed per source. A source we could not open is never counted as evidence for an assertion.

This page is also available as structured data: /api/v1/topics/memory