A Lapse in Memory — Sources

This document provides the sourcing and analytical framework for the long-read piece A Lapse in Memory. Sources are organised by section and claim. Primary sources (company press releases, earnings call transcripts, regulatory filings, court opinions) are flagged ★. Secondary sources are flagged ☆. The closing section flags the piece’s original analytical contributions, which should be read as argument rather than reported fact.

Section 1 — The wafer that does not exist

SK Hynix Q1 calendar 2026 operating margin of 72 per cent, surpassing TSMC’s same-quarter 58.1 per cent.

Comparator quarters (per Section 1 paragraph 1).

The piece compares Hynix’s Q1 calendar 2026 figure against TSMC’s same-quarter result and against the most recently reported operating margins of Nvidia, Apple and Microsoft. The other comparators do not align on the same calendar quarter — fiscal calendars differ.

  • TSMC Q1 calendar 2026 (ended 31 March 2026): operating margin 58.1%. Same quarter as Hynix.
  • Nvidia Q4 fiscal 2026 (ended 26 January 2026, ≈ calendar Q4 2025): GAAP operating margin 65.0% ($44.3B operating income / $68.1B revenue). Source: Nvidia 8-K filing. Nvidia’s Q1 fiscal 2027 results (covering calendar Q1 2026) are scheduled for late May 2026.
  • Apple Q1 fiscal 2026 (ended late December 2025, ≈ calendar Q4 2025): operating margin 35.3%. Apple’s calendar Q1 2026 results report in early May 2026.
  • Micron Q1 fiscal 2026 (ended late November 2025): Q2 guidance pointed toward 67% gross margin for the quarter ending in March 2026 (covering calendar Q1 2026).

The piece accordingly frames the TSMC comparison as same-quarter and frames the others as “most recently reported.”

Convergence: Micron Q2 fiscal 2026 gross margin guidance ~67 per cent; Samsung Device Solutions division Q4 2025 operating profit ₩16.4 trillion on ₩44 trillion revenue, with DRAM operating margin estimated above 50 per cent.

OpenAI signed Letters of Intent with Samsung and Hynix in October 2025 for up to 900,000 DRAM wafers per month for its Stargate data-centre programme.

900,000 wafers per month represents around 40 per cent of global DRAM wafer-start capacity (estimated 2.2 to 2.3 million per month, 2025).

  • ☆ TechInsights estimate, cited via Tom’s Hardware: global 300mm fab capacity ~10M WSPM in 2025, with DRAM as a sub-segment.
  • ☆ TrendForce / TechInsights: global DRAM wafer production capacity ~2.2-2.3M WSPM in 2025.
  • ☆ Lambda Finance, April 2026: global 300mm wafer-fab capacity ~9M WPM by early 2026, “splits roughly in half between logic and memory” — implies DRAM capacity ~2.0-2.5M WSPM.
  • ☆ TrendForce / The Bell / Maeil Business: Samsung DRAM ~650-700K WSPM, Hynix ~mid-300Ks WSPM (as of 2025), ramping to ~600K by Q4 2026.
  • Calculation: 900,000 / ~2,250,000 ≈ 40%. Denominator is global DRAM-only capacity, not all 300mm fabs.

LOI is non-binding “up to” framework agreement.

  • Industry-standard LOI procurement architecture; non-binding framework agreements securing queue position. The “up to” structure is load-bearing — the figure is a ceiling, not a floor, with actual draw set by manufacturer execution and buyer take-up.

Samsung Q4 2025 corporate operating profit ₩20.1T (+208% YoY), highest quarterly profit on record.

SK Hynix full-year 2025 operating profit ₩47.2T; Q4 operating margin 58.4 per cent against TSMC’s same-period 54.0 per cent — first time in seven years that memory profitability has overtaken foundry; Hynix DRAM revenue +70.6% YoY, NAND +59% YoY.

Micron Q2 FY26 (calendar Q1 2026): GAAP operating margin 67%, gross margin 74%; Q1 FY26 revenue $13.6B (+57% YoY); DRAM revenue +69%, NAND +22%.

Micron quarterly GAAP operating margin progression (used in Chart I):

  • Cal Q4 2024 (Micron Q1 FY25): $8.71B revenue / $2.17B op income = 25.0% (★ 8-K 18 Dec 2024)
  • Cal Q1 2025 (Micron Q2 FY25): $8.05B / $1.77B = 22.0% (★ 8-K 20 Mar 2025)
  • Cal Q2 2025 (Micron Q3 FY25): $9.30B / $2.17B = 23.3% (★ 8-K Jun 2025)
  • Cal Q3 2025 (Micron Q4 FY25): $11.32B / $3.65B = 32.3% (★ 8-K Sep 2025)
  • Cal Q4 2025 (Micron Q1 FY26): $13.64B / $6.14B = 45.0% (★ 8-K 17 Dec 2025)
  • Cal Q1 2026 (Micron Q2 FY26): ~$16.4B / ~$11B = 67% (★ 8-K Mar 2026)

The Q4 2024 → Q1 2026 trajectory: 25 → 22 → 23 → 32 → 45 → 67. The 42-percentage-point expansion in five quarters is the steepest sustained margin expansion in Micron’s history.

SK Hynix Q4 2025 DRAM bit shipments grew low single-digit per cent quarter-on-quarter; ASP rose mid-twenties per cent QoQ.

  • ★ SK hynix Q4 2025 earnings call transcript, 29 January 2026 (via Alpha Spread): https://www.alphaspread.com/security/krx/000660/investor-relations/earnings-call/q4-2025
  • ★ SK hynix Q4 2025 prepared remarks (verbatim): “DRAM shipments grew by low single-digit percent, driven by increased sales of HBM3E products and DDR5 for servers. Shipments of high-density DDR5 modules increased by approximately 50% Q-o-Q… ASP rose 20% Q-o-Q, reflecting a significant increase in conventional DRAM prices.”
  • ☆ Futurum, “SK Hynix Q4 FY 2025: Structural Shift to AI Memory Lifts Margins,” 2 February 2026

The supply-discipline argument turns on the gap between bit-shipment growth (low single digits) and ASP growth (mid-twenties), not on a literal bit-shipment decline.

Aggregate DRAM revenue +29.4% QoQ in Q4 2025 on shipments essentially flat.

  • ☆ TrendForce, multiple Q4 2025 reports.

Inventory at 2-4 weeks DRAM and 4-5 weeks NAND in October 2025; down from 13-17 weeks at end of 2024.

  • ☆ TrendForce inventory tracking, October 2025; multiple secondary sources citing TrendForce data.

Section 2 — From sand to socket

Production chain primer. Standard semiconductor manufacturing references; no specific claims requiring sourcing beyond general industry knowledge. Asianometry (YouTube channel) credited inline at section close as the best public-source education on the chain.

Five companies dominate silicon ingot/wafer production; Shin-Etsu and SUMCO together hold roughly half the global market.

  • Industry-standard knowledge; corroborated by USGS Mineral Commodity Summaries and SEMI industry reports.

ASML is sole EUV producer; machines $200-400M each, 18+ months lead time. SK Hynix recently placed a record $8 billion ASML order.

  • ASML public disclosures.
  • ☆ Industry analyst commentary, multiple sources, 2025.

HBM-bound wafer is committed during front-end processing (TSV drilling, thinning, HBM-specific I/O).

Section 3 — The disconnect

(Numbers cross-reference Section 1 sources.)

Section 4 — Phantom capacity

18-28 per cent of DRAM capacity allocated to HBM (rising); HBM4 ramping in 2026.

  • ☆ TrendForce HBM allocation tracking, 2025-26: HBM share of DRAM wafer output 13% (2024) → 19% (2025) → 23% (2026 baseline). With HBM4 ramping into year-end 2026, end-of-period figure may reach 28%.
  • ☆ The Bell / Maeil Business: Hynix ~30% HBM allocation (2025), projected ~40% by 2027.
  • ☆ TechInsights Q3 2025 wafer demand forecast: HBM accounts for 16% of total DRAM wafer allocation, projected 21% by 2030: https://www.techinsights.com/blog/wafer-demand-forecast-q3-2025-update

Wafer allocation by product class, 2024 / 2025 / 2026E (used in Chart II).

  • DDR4 share collapse: 2024 ~30% → 2025 ~15% → 2026E ~5%. Sources: Samsung halted most DDR4 in 2025 with final shipments end-2025; SK Hynix reduced DDR4 share to ~20% (BISI/Wikipedia, citing manufacturer disclosures); Micron announced DDR4 phase-out then partially reversed late-2025 when DDR4 spot prices inverted above DDR5. NAND Research call notes (Nov 2025): “DDR4 production may fall to ~20% of 2025 levels in 2026.”
  • HBM share rise: 2024 13% → 2025 19% → 2026E ~28% (TrendForce baseline + HBM4 ramp).
  • DDR5 absorbs the residual: 2024 ~57% → 2025 ~66% → 2026E ~67%.
  • All splits aggregate across Samsung, SK Hynix and Micron; per-company allocation varies materially.

HBM-to-DDR5 trade ratio is 3:1 (and increasing in future generations).

  • ★ Micron Technology Q1 FY26 prepared remarks (December 2025), verbatim: “the 3-to-1 trade ratio with DDR5, and this trade ratio only increases with future generations of HBM”: https://investors.micron.com/static-files/088991c5-a249-4f66-a0a6-258d9b66f3f9
  • ☆ Tom’s Hardware: “HBM requires three times as much wafer space as DDR5”
  • ☆ Marvin Labs: “3:1 HBM wafer trade ratio is the structural foundation of the supply constraint thesis”

This is the piece’s clearest single primary-source anchor.

Raw 300mm substrate wafers approximately $200 each.

  • ☆ Tom’s Hardware, March 2022; current 2025-26 price range $200-280 per polished wafer for prime grade. Order-of-magnitude reference.

Samsung allocating roughly half of Pyeongtaek to HBM4 base-die production for 2026.

Long-term agreements refused Q4 2025 / Q1 2026, then offered March / April 2026.

  • ★ Micron Q2 FY26 earnings call (March 2026): Sanjay Mehrotra disclosed first 5-year Strategic Customer Agreement: https://investors.micron.com/static-files/e089f8c0-065d-47b8-9d02-bfa863cdb357
  • ☆ Korean financial press (Maeil Business, ETNews, ZDNet) Q1 2026 reporting on Samsung-Google/Microsoft 3-5 year deals.
  • ☆ Hynix-Microsoft multi-year DDR5 supply agreement: secondary press, Q1 2026.

Section 5 — The historical rhyme

2005 DOJ DRAM cartel pleas: $646M total fines, four guilty pleas (Samsung, Hynix, Infineon, Elpida); Micron leniency.

  • ★ US DOJ press releases 2004-2006; Samsung pleaded guilty October 2005, $300M fine.

2010 EU Commission €331M fine.

  • ★ European Commission press release IP/10/586, 19 May 2010. Available at the EC competition portal.

2018 California class action; China MOFCOM raid on the three manufacturers.

  • ☆ Reuters, Bloomberg coverage of China antitrust raids, May 2018.
  • ☆ California class action filings, public docket.

2022 Ninth Circuit ruling: parallel conduct in concentrated DRAM market does not state a Sherman Act conspiracy claim.

  • ★ In re DRAM Indirect Purchaser Antitrust Litigation, No. 21-15125 (9th Cir. 7 March 2022). Opinion by Judge Cathy Ann Bencivengo (S.D. Cal., sitting by designation), joined by Circuit Judges William A. Fletcher and Johnnie B. Rawlinson.
  • ★ Underlying district court ruling: Jones v. Micron Tech. Inc., 400 F. Supp. 3d 897 (N.D. Cal. 2019), Judge Jeffrey S. White presiding (case No. 4:18-cv-02518-JSW).
  • ★ Procedural posture: Rule 12(b)(6) affirmance — the court tested the complaint at its most favourable, assuming every factual allegation was true, and still held that the plaintiffs had not stated a claim. Not a finding on the merits.
  • ★ Key holdings: defendants’ contemporaneous capex cuts were “more suggestive of lawful conscious parallelism than conspiracy”; even the prior 2005 criminal convictions for price-fixing, which the court acknowledged “circumstantially supports Plaintiffs’ theory,” were insufficient when other plus factors fell short. The court applied the Twombly pleading standard (Bell Atl. Corp. v. Twombly, 550 U.S. 544 (2007)), under which a complaint must contain factual allegations sufficient to push a Sherman Act conspiracy claim from “conceivable to plausible.”
  • Opinion text: https://cdn.ca9.uscourts.gov/datastore/opinions/2022/03/07/21-15125.pdf
  • Justia summary: https://law.justia.com/cases/federal/appellate-courts/ca9/21-15125/21-15125-2022-03-07.html
  • Precedential weight: a published Ninth Circuit opinion — binding precedent in the Ninth Circuit and persuasive authority elsewhere.

Sherman Act doctrine of conscious parallelism.

  • Standard antitrust doctrine, established in Theatre Enterprises v. Paramount Film Distrib. Corp., 346 U.S. 537 (1954) and refined through Twombly (2007) and Iqbal (Ashcroft v. Iqbal, 556 U.S. 662 (2009)).
  • Definition (from the In re DRAM opinion): “Conscious parallelism occurs when two or more firms in a concentrated, interdependent market base their actions in part on the anticipated reactions of their competitors, and thus arrive at identical decisions independently, as they are cognizant of — and reacting to — similar market pressures.”

Manufacturer-articulated independent-rational-reason justification: oversupply risk.

The conscious-parallelism shield protects parallel restraint when each firm articulates an independent rational reason. The three manufacturers have, on the public record, advanced precisely such reasons — but the framing is specific. The defence is articulated against capex (not building new fabs) rather than against operational utilisation (running existing fabs below maximum). The distinction is substantive: building or not building a fab is a discrete, defensible capital-allocation decision; running an existing fab below capacity is a continuous operating choice that is harder to defend as parallel-but-independent.

The 2018–2019 oversupply collapse — the historical anchor the manufacturers reference.

The Samsung “minimize the risk of oversupply” formulation invokes a specific traumatic event: the 2018–2019 memory price collapse. Key benchmarks of that collapse:

  • Q4 2018 DRAM ASP fell approximately 30% quarter-on-quarter, the steepest single-quarter decline in DRAM history at that time.
  • SK Hynix operating margin fell from above 50% (Q3 2018) to single digits (Q3 2019) within four quarters.
  • Samsung memory operating profit fell roughly 56% year-on-year in 2019.
  • Micron took $1.4B in inventory writedowns across FY2019; share price fell from ~$60 to ~$30 between mid-2018 and the end of 2018.
  • The collapse ended with all three manufacturers cutting capex sharply through 2019–2020 — a cycle the manufacturers now publicly cite as the lesson informing 2026 discipline.
  • Specific 2018-2019 financial figures should be re-verified against company filings before publication.

Single-quarter incremental operating profit across the three manufacturers, Q4 2024 → Q4 2025: approximately $22 billion combined.

  • ★ SK Hynix Q4 2024 operating profit: ₩8.0828T (per company press release, 22 January 2025) ≈ $5.6B at ~₩1450/USD.
  • ★ SK Hynix Q4 2025 operating profit: ₩19.1696T (per FY25 release, 28 January 2026) ≈ $13.4B at ~₩1430/USD. Uplift: ~$8B.
  • ★ Micron Q1 FY25 GAAP operating income: $2.174B (per Q1 FY25 8-K, 18 December 2024).
  • ★ Micron Q1 FY26 GAAP operating income: $6.136B (per Q1 FY26 8-K, 17 December 2025). Uplift: ~$4B.
  • Samsung memory: Q4 2024 ~$2B, Q4 2025 ~$12B; uplift $10B. Samsung does not break out memory-only operating profit at the segment level; figure is analyst-attributed (Hankyung, KED Global, Korea Herald) cross-referenced against Q4 2025 Device Solutions disclosure of ₩16.4T ($11.5B).
  • Combined uplift: $4B + $8B + $10B ≈ $22B in a single quarter.
  • Used in Section 5 deterrent-math passage and Chart IV: “the marginal rent extracted by the configuration, in a single quarter, exceeds the entire historical enforcement record by more than twenty times.”

SK Hynix full-year 2025 operating profit doubled vs 2024.

  • 2024 Hynix operating profit: ₩23.4T (publicly disclosed).
  • 2025 Hynix operating profit: ₩47.2T (per FY2025 release).
  • Actual ratio: 2.02x.

Section 6 — The empty fab and the burning flare

Memory fab capital cost approximately $15 billion.

  • ☆ Industry standard estimate for greenfield 300mm DRAM fab; varies by node generation. CHIPS Act funded projects in the $10-20B range.

Iranian drone strikes against QatarEnergy’s Ras Laffan industrial complex disrupted operations beginning 2 March 2026; QatarEnergy declared force majeure on deliveries from the complex on 4 March; Strait of Hormuz closed to Western commercial shipping in early March; repair timelines estimated at three to five years.

South Korea sourced 64.7 per cent of helium imports from Qatar in 2025.

  • ☆ Korea International Trade Association (KITA) data, cited via Tom’s Hardware, CSIS, TrendForce, Hankyung, Forbes — multi-source corroboration.

Spot helium prices rose more than 40 per cent within weeks.

  • ☆ CSIS: “more than 40 percent.”
  • ☆ Multiple secondary sources: 40%+.

Airgas declared force majeure on US helium deliveries on 17 March 2026.

  • ☆ Industry press, March 2026.

Helium consumption per wafer is rising as nodes shrink, particularly for EUV-intensive processes.

  • ☆ Forbes; Korea Semiconductor Industry Association commentary.

Qatar accounts for approximately 30 to 34 per cent of global helium production.

Korean Fair Trade Commission opened DRAM/PC pricing investigation early 2026.

  • ☆ Korean financial press, Q1 2026.

Section 7 — The release valves that aren’t releasing

Generational divestment cascade as historical mechanism (SDRAM → DDR → DDR2 → DDR3 → DDR4 → DDR5; GDDR fork).

  • Standard industry history; corroborated across DRAM industry retrospectives, IEEE Solid State Circuits Society publications, and TrendForce historical reports.

CXMT market share moved from approximately 3 per cent to 5 per cent over the past year.

  • ☆ TrendForce DRAM market share tracking, 2024-2025.

SK Hynix Wuxi fab upgraded from 1z to 1a node, 2024-2025.

  • ☆ Korean financial press; secondary semiconductor industry coverage.

Samsung and Micron DDR4 EOL announced mid-2025, reversed late 2025; Samsung signed long-term non-cancellable contracts Q1 2026.

Samsung MLC NAND ending 2025; Kioxia phasing out 2D flash and BiCS3 by end 2028.

  • ☆ Industry press, 2025.

Micron Crucial brand wind-down announced 3 December 2025; last shipments February 2026.

Section 8 — The downstream tax

Contract DRAM pricing is private; retail aftermarket is the only public longitudinal price series.

  • TrendForce, DRAMeXchange and other contract-pricing trackers are subscription-only. PCPartPicker is the only public source for memory price-history data accessible to general readership.

PCPartPicker retail data — 32GB DDR5-6000 +242 per cent May to December 2025; 64GB +257 per cent; Crucial Pro 32GB $80 → $360.

Chip spot pricing — DDR4/DDR5 reached parity early June 2025; DDR4 traded above DDR5 through July-August 2025; DDR5 +298 per cent September-December 2025.

64GB DDR5 retail above PS5 Pro launch price; Texas Micro Center DDR5 prices removed from posted display.

  • ☆ Reddit/forum reports; Tom’s Hardware coverage November-December 2025.

Motherboard sales fell 40-50 per cent year-on-year.

  • ☆ Industry press, December 2025.

CyberPowerPC announced 25 November 2025 that DRAM input costs had surged 500 per cent since October; system price increases effective 7 December 2025.

Gartner February 2026 forecast: 130 per cent memory price surge by year-end; PC ASPs +17%, smartphone ASPs +13%; sub-$500 PC dead by 2028; PC lifetimes +15-20%; BOM 16% → 23%.

  • ☆ Gartner press release / report, February 2026.

IDC PC unit shipments 2026 forecast -11.3 per cent; smartphone Q1 2026 -6.8 per cent.

  • ☆ IDC Worldwide Quarterly PC Tracker, Q1 2026 release.
  • ☆ IDC Worldwide Smartphone Quarterly Tracker, Q1 2026.

Server DRAM contract prices Q1 2026 +60-70 per cent QoQ; Samsung Q2 +30 per cent.

  • ☆ TrendForce server DRAM contract pricing, Q1 2026 reports.

Samsung Galaxy S26 +16 per cent LPDDR5 cost.

  • ☆ Korean financial press, Q1 2026.

Hyperscaler 2026 capex around $602 billion aggregate (CreditSights); consensus range $602B–$700B.

EU Article 101 covers concerted practices alongside formal agreements; threshold lower than post-2022 Sherman Act bar.

  • ★ Treaty on the Functioning of the European Union, Article 101(1).
  • Standard EU competition law; the “concerted practices” doctrine catches conduct that is consciously parallel even absent formal agreement, where there is “knowing substitution of practical cooperation between [undertakings] for the risks of competition” (Suiker Unie, Case 40/73, 1975).
  • Established in Imperial Chemical Industries v. Commission (1972) — the “Dyestuffs” case.

Section 9 — The lapse

The forward-scenario analysis and the closing image are analytical claims drawn from the configuration described in earlier sections. No specific sourcing required.

Analytical framework

The piece makes several analytical moves that are original contributions and should be read as argument rather than reported fact:

The unstarted-versus-undiced wafer distinction. The contract reserves undiced wafers — back-end optionality stays with the seller. The phantom capacity is the unstarted wafer — front-end fab-time hasn’t been scheduled. These are two distinct concepts doing different analytical work. The piece introduces the distinction in Section 4 as named and returns to it in the Section 9 close.

The 3:1 displacement ratio plus HBM front-end commitment as combined mechanism. The Micron primary source establishes the trade ratio. The piece’s analytical contribution is connecting the ratio to the front-end-committed nature of HBM production (TSV drilling, thinning, HBM-specific I/O) to explain why HBM allocation reduces commodity bit output by significantly more than its share of capacity.

Deterrent math reframed against incremental profit, not total profit. Section 5 close. The regulatory question is reframed from “would prosecution succeed” to “would prosecution matter even if it succeeded.” A single quarter’s incremental profit at one manufacturer alone exceeds the cumulative historical fines paid by all three over twenty years.

The generational divestment cascade and why HBM breaks it. Section 6. HBM is additive to the DRAM portfolio, not replacement; the legacy fab is therefore not made obsolete; the divestment mechanism that historically ended supply squeezes is disrupted because there is no reason to divest legacy capacity to a second-tier player when both products coexist profitably.

Ras Laffan as legitimisation, not cause. Section 8. The supply-discipline configuration was in place by late 2024, before Ras Laffan was a known risk and before the Stargate Letters of Intent were signed. Helium provides ratified cover for behaviour that predates the strike. “The configuration was already in place. Ras Laffan made it speakable.”

Warm sockets and the closed-circle reframing. Section 4 close. The most novel structural contribution. AI capacity at scale is constrained not by chips but by the operational data-centre footprint required to run them — substations, switchgear, water rights, fibre, cooling, staffing, grid interconnect — which the piece terms warm sockets. The chip-side LOI is therefore an option contract on warm-socket build-out, not a procurement document. Both manufacturer and hyperscaler share an incentive to minimise physical realisation. The configuration extracts rent from outside the closed circle and distributes the gains within it.

Three-way convergence as structural fingerprint. Section 1 paragraph 2. Three independently operated companies in three different countries, with different boards, cost structures, capex profiles and AI exposure, printing the same margin anomaly in the same quarter is the empirical signature that distinguishes structural configuration from coincidence.

The hyperscaler tax as alignment, not extraction. Section 7 close. Hyperscalers and frontier-AI buyers built pricing power against the rest of the technology supply chain through the LOI architecture in commercial alignment with the manufacturers, not by accident or as victims of squeeze. The tax funds both manufacturer supply discipline and hyperscaler competitive moat.

Methodological notes

Why retail data is the lead indicator for end-customer impact. Contract DRAM pricing is private; trackers like TrendForce DRAMeXchange are subscription-only. PCPartPicker is the only public longitudinal price series accessible to general readership, which is why Section 7 leads with retail data despite contract pricing being the more analytically pure measure.

Capacity denominator. The 40 per cent figure for the OpenAI LOI is calculated against global DRAM-only wafer-start capacity (estimated 2.2-2.3M WSPM for 2025), not against all 300mm fab capacity. The denominator distinction matters because logic and other non-memory production runs on the same 300mm equipment.

Comparator-quarter alignment. The Section 1 opener compares Hynix Q1 calendar 2026 (the most extreme reading) against TSMC Q1 calendar 2026 (same quarter, directly comparable) and against the most recently reported figures from Nvidia, Apple and Microsoft. Fiscal calendars do not align — Nvidia’s Q1 calendar 2026 results report in late May 2026, Apple’s in early May. The piece accordingly frames the TSMC comparison as same-quarter and the others as “most recently reported.”

Samsung memory operating profit attribution. Samsung does not disclose memory-only operating profit at the segment level. The piece uses Device Solutions division figures (which combine memory and foundry but with memory dominant) and notes the analyst attribution explicitly. The “memory operating margin estimated above 50 per cent” framing is sourced to Korean financial press analyst commentary.