Micron

Refresh delta (since 2026-07-07)

What changed — six material moves

1. Price: ~$984.75 → $870.00, −11.7%. The sector is now formally in a bear market. MU last traded $870.00 (Aug 10 2026, 11:35 EDT), prev close $877.57 (Aug 7). Market cap fell back below $1T to $982.6B. That is −30.7% from the $1,255 all-time high — and MU, Samsung, SK Hynix and the Roundhill Memory ETF are all more than 20% off recent closing highs. In July alone the KOSPI fell >20%, SK Hynix −33%, Samsung −16.5%. Forward P/E is now 6.11x, trailing 19.8x.

2. CXMT was handed $8.6B — the prior dossier's named falsifier is now funded and dated. ChangXin Memory listed on Shanghai's STAR Market and closed +466% on debut, raising ¥57.92B (~$8.6B) — Asia's largest IPO of 2026 — at a ~$487B market capitalisation. MU fell ~5.5% on Jul 27 alone; SanDisk −11.4%, SK Hynix ADRs −8.5% the same session. Its Shanghai fab completes in 2H CY2026 (up to ~400k wspm at full ramp), and analysts model CXMT's global DRAM share rising ~10% → ~18% by late 2028. The critical nuance almost every write-up buried: CXMT's IPO prospectus assigns the bulk of named project spending to DRAM wafer lines and process upgrades, with nothing earmarked for HBM. This is a war chest aimed squarely at the commodity book — precisely MU's un-contracted exposure — and not at the HBMHigh-bandwidth memoryMemory stacked in layers beside a processor so data reaches it faster. The usual bottleneck in AI chips. rent.

3. Pricing has NOT rolled over — the deceleration is demand-side. TrendForce's DDR4 8Gb spot benchmark rose $36.00 (Jun 30) → $42.08 (Jul 28), +16.9% in a month. Q3 CY26 DRAM contract prices are guided +13–18% QoQ and NAND +10–15%, down from roughly +60% in Q2 — and TrendForce attributes the slowdown explicitly to consumer electronics makers being unwilling and unable to absorb further increases, not to improved supply, with server DRAM still undersupplied. Prices are still rising. The second derivative turned; the first derivative did not. This is the single most important fact for the falsifier, and it is the one the tape is ignoring.

4. The SCA terms got materially more specific — and the floor is now quantified against a real benchmark. 16 agreements, 14 of them priced, totalling $100B of minimum revenue commitments, running calendar 2026 through end-2030, against $22B of customer deposits ($18B cash held). Coverage: ~20% of DRAM volume and ~1/3 of NAND volume, expected to eventually represent 50%+ of company revenue. Ceilings pinned to CQ2-2026 market prices. Floors lock gross margins above Micron's ~62% all-time-peak quarterly gross margin. ⚠ Conflict surfaced, not resolved: the prior dossier carried "~40% of revenue at fixed/ceiling pricing, un-contracted ~60%." The new figures are a volume measure (~20% of DRAM bits), not a revenue measure. They are not reconcilable from public data and they are not the same claim — ~20% of DRAM volume is a materially thinner shield than "~40% of revenue," because DRAM is ~76% of the book.

5. The incumbent supply response does not arrive until 2028-29 — later than the prior pre-mortem assumed. SK Hynix's $38B (~₩54tn) commitment breaks ground on the Y2 Yongin DRAM fab in July 2027, first cleanroom June 2029; the M17 Cheongju NAND fab breaks ground Feb 2027, first cleanroom Dec 2028. Micron's own Singapore HBM packaging capacity contributes from H1 CY2027, and it has begun a second, EUV-capable cleanroom. All three makers are now sold out of DRAM and HBM through all of CY2027, not just CY2026. The prior dossier's pre-mortem put the synchronised incumbent node ramp in H2 CY2027. On these construction schedules that is roughly a year and a half too early — for incumbents. CXMT is the only near-dated supply.

6. SK Hynix's US listing catalyst resolved — and it resolved badly, which cuts for the pair tilt. SK Hynix completed its $26.5B Nasdaq ADR listing on 2026-07-10 (largest US listing ever by a foreign company), debuting +13% at $168.01. Three sessions later its Korean line fell 15.4% — the largest single-day fall in its history — triggering a 20-minute market-wide halt as the KOSPI shed 9%. Cited causes: profit-taking, HBM4 volume disappointment (the Q2 shipment ramp did not come through at the expected magnitude), an operating profit tracking ~8% below consensus, and unresolved valuation of the US line against the Korean one.

Two corrections to the prior dossier — both found this pass

• ⚠ The name is NOT regulatory-clean. The prior dossier stated "Regulatory: clean (zero LR/AAER)." That is true of the SEC file only, and regulatory/regulatory-findings.md says so in its own body text — it covers SEC Litigation Releases and AAERs and explicitly instructs the reader to run a web search for FTC/DOJ/other actions. Nobody ran it. Samsung, SK Hynix and Micron were hit with a US price-fixing class action over the memory shortage on ~2026-06-30. This predates the boundary — it is a miss, not a delta, and it is reported as a miss.
• ⚠ FQ4'26 is a FOURTEEN-week quarter. From the 10-QForm 10-QThe quarterly version of the annual report. Lighter, and not audited.'s own MD&A: "Fiscal 2026 contains 53 weeks... The third quarter of 2026 contains 13 weeks and the fourth quarter of 2026 contains 14 weeks". The $50B revenue guide therefore embeds an extra week worth ~7.7%. On a 13-week-equivalent basis the midpoint is ~$46.4B, i.e. ~+12% QoQ rather than the ~+21% the headline implies. No prior dossier, and none of the secondary coverage read this pass, flags it.

What held (structural thesis intact)

• The raised-floor core is stronger, not weaker: $22B prepaid take-or-pay, $100B of minimum commitments to 2030, floors above the all-time-peak margin, sold out through all of CY2027. Every piece of new information on the contracts was favourable.
• Books are clean and the FQ3'26 re-extraction confirms it independently (Lens 10) — cash conversion, receivable quality and inventory discipline all check out against the primary statements for the first time on this shelf.
• Customer concentration is genuinely improving, and it is filed: one customer was 10% of 9M FY26 revenue vs 16% in 9M FY25.
• The bear vectors — commodity core, CXMT, peak margins, under-depreciation, capped upside, extreme AI-complex beta — all carry.

Phase A — Understand the business

• Lens 1 (Overview) — unchanged, see the previous dossier. IDM, four BUs, 16 SCAs. One addition: Tongluo (Miaoli County, Taiwan) wafer fab acquired from Powerchip in March 2026 for $1.8B cash, and construction-in-progress nearly doubled to $10.9B from $5.5B.
• Lens 2 (Supply chain) — unchanged, see the previous dossier Lens 2. One watch item added: reports of China domestically manufacturing DUV lithography tools, a technology-gap-closure vector behind part of the July de-rate — thinly sourced, flagged not carried.

Moats (REFRESHED — the anti-moat got funded, the moat got a term sheet)

The moat structure is unchanged in kind and sharpened in degree, and the two sides moved in opposite directions this pass.

The moat strengthened where it is contractual. The SCAs now have a known term (calendar 2026–2030), a known minimum ($100B), and a benchmarked floor (gross margins above the ~62% all-time peak). A five-year take-or-pay book with cash deposits and floors above any margin the industry has ever printed is not a memory-industry feature — it is closer to a utility's rate base, and it is the single most defensible thing about this business.

The anti-moat got $8.6B. CXMT's structural position is unchanged — state-backed, cost-advantaged, sub-scale in HBM — but its capital constraint is gone. The offsetting detail is the one worth holding onto: its own prospectus earmarks nothing for HBM. So the correct read is not "CXMT is coming for Micron." It is CXMT is coming for the commodity DRAM book, on a 2027-28 timeline, with the capital to do it — and is not, on its own disclosed plans, coming for HBM. That splits Micron cleanly in two: the contracted/HBM half is more defended than a quarter ago; the commodity half is materially more threatened.

Segments (REFRESHED — the BU table extracted from primary source for the first time)

The prior dossiers cited the DRAM/NAND product split ("DRAM $31.3B/76%, NAND $9.9B/24%"). The reportable-segment table was never extracted. It is now, and it contains the sharpest cycle-peak tell in the filing.

FQ3'26 (quarter ended 2026-05-28), $M:

The finding: the highest-margin business at Micron is not the AI business. It is mobile and client. MCBU — commodity consumer memory — printed an 87.3% gross margin and 85.7% operating margin, above both data-center units. And the 10-Q says exactly how: MCBU revenue rose "primarily due to increases in average selling prices, partially offset by lower bit shipments". Pure price, negative volume, in the most commoditised segment on the books — and it is the segment TrendForce identifies as hitting the consumer affordability wall right now.

This reframes the peak-margin objection precisely. The bear case is not "AI margins will normalise." It is that ~28% of revenue at an 87% gross margin is commodity memory being sold at scarcity prices to buyers who are already balking, and that is the first block to go. Note also that AEBU — the weakest of the four at 79.0% GM — still runs roughly double any company-wide peak in Micron's history.

---

Phase B — Measure performance

Earnings Result (no new print — context and primary re-extraction)

FQ3'26 remains the latest print. FQ4'26 reports ~late Sep 2026. This pass re-derived the quarter from the 10-Q rather than the press release, and the primary figures differ slightly from what has been carried:

Two things worth naming. Interest expense went to zero in the quarter after $9.38B of 9M debt repayment took long-term debt from $14,017M to $5,140M — the balance sheet is now $100.7B of equity against $5.7B of gross debt. And the widely-quoted "$18.3B FCF" uses capex net of government incentives; on the 10-Q's gross capex basis it is $17,562M. Neither is wrong; they are different conventions, and the shelf now carries both with the distinction recorded.

FQ4'26 guide unchanged — and now correctly caveated. Revenue $50B ±$1B, GM ~86%, non-GAAP EPS $31 ±$1, FY26 capex ~$27B net of incentives. Read every one of those through the 14th week (see the corrections block). The revenue guide is ~$46.4B on a comparable-week basis. The margin and EPS guides are less distorted but not undistorted — a 14-week quarter spreads fixed costs over more revenue.

The next fundamental catalyst is unchanged and is now the whole ballgame: the FQ4'26 print plus any first CY2027 colour, ~late Sep. Everything the bull and bear disagree about resolves there or not at all.

Earnings Calls (REFRESHED — a new and uncomfortable tonal development)

The sentiment arc is as documented: "disciplined" → "structural shortage" → "transformed business model." Management still will not guide past FQ4 and will not commit CY2027 pricing — the honest tell, unchanged.

What is new is off-call and it is a tone shift worth flagging. On 2026-06-30, Mehrotra publicly blamed customers for the shortage: "Certain customers drove pricing significantly down in our industry. In 2023, our prices came down to one-third of what they were" — arguing years of price pressure left the industry underinvested going into the AI surge. CBO Sumit Sadana pointed implicitly at specific buyers, widely read as Apple.

The juxtaposition is the finding, and it is not comfortable. The same week that Micron's CEO publicly discussed industry-wide pricing dynamics, underinvestment and the causes of the shortage, Micron, Samsung and SK Hynix were named in a US price-fixing class action arising from that shortage. Nothing here suggests wrongdoing, and this is a defendant's ordinary commercial explanation. But a CEO narrating industry-wide pricing and capacity behaviour while a coordination claim is live is a genuine litigation-risk posture, and it is the kind of thing that reads differently in a deposition than on CNBC. Watch whether the FQ4 call is noticeably more lawyered — a sudden retreat from the "transformed business model" framing would be a tell about counsel, not about demand.

Comps & Valuation (REFRESHED — all marks co-dated 2026-08-10)

All marks. EV/EBITDA and P/B were not sourced co-dated this pass — recorded as n/a rather than carried stale from July.

The most useful thing in that table is what the forward multiple implies about consensus. At $870 and a 6.11x forward P/E, the street is underwriting roughly $142 of forward EPS. That is not the base case — it is almost exactly the prior dossier's bull case (FY27 EPS ~$140). So the framing has inverted since 2026-07-07, and this matters more than the price move:

The prior dossier said the market had stopped pricing the bull and started pricing the base. That is no longer the right description. At 6.11x the consensus estimate itself is a bull-case estimate, and the low multiple is the market's discount against that estimate — not a cheap price on a conservative number. Those are very different things, and conflating them is the standard way people lose money in memory. The honest statement is: the multiple is low because the market disbelieves the E, not because the E is conservative.

Peer / consensus frame: analyst consensus target $1,501.98 (Strong Buy, 46 analysts, S&P Global), +72.6% from $870; targets clustered $1,500–$1,700 elsewhere; BofA added MU to its US 1 List in July. ⚠ Unresolved conflict: two co-dated secondary reports have Citi's Atif Malik at $1,150 (cut from $1,400) and at $1,400 reiterated (Aug 4). Not reconcilable from the sources read — neither is cited as a number here. And note the direction of travel: consensus targets rose while the price fell ~30%. A 73% consensus gap widening into a falling tape is historically a statement about analyst revision lag, not a discovered opportunity.

Pair leg (SK Hynix): no reliable co-dated EV/EBITDA or forward P/E was obtained this pass. The July comparison in the prior dossier is stale and is not carried forward. What is known is qualitative and material — see Lens 8 and the pair block.

Stock-Price Catalysts (REFRESHED — the character of the drawdown changed)

The important observation of this entire pass: the prior dossier documented a repeating pattern — MU falls 13% on AI-complex sentiment with no company-specific news, then recovers. It explicitly used that pattern to classify the July dip as an add rather than a stop. This drawdown does not fit that pattern, and treating it as another sentiment dip would be pattern-matching against a changed cause.

Moves since 2026-07-07:

• ~Jul 13 — SK Hynix's Korean line −15.4% (its worst day ever), KOSPI −9%, market-wide halt; read-across to the whole complex. Cause: HBM4 volume disappointment + post-listing profit-taking.
• Jul 16 — SK Hynix and SanDisk −7%, MU −5% as CXMT's $8.6B IPO was readied.
• Jul 24 — CEO Mehrotra sells 31,285 shares at $906.48–$941.60 (10b5-1).
• Jul 27 — MU −5.5% on CXMT's +466% debut; SanDisk −11.4%, SK Hynix ADRs −8.5%.
• Aug 6 — KOSPI −4.37% on SanDisk's weak revenue guidance and a Western Digital selloff spilling into memory. (SanDisk guided 79–81% non-GAAP GM — the problem was volume, not margin.)
• Aug 7 — Morgan Stanley's Shawn Kim, a long-standing memory bear, turned bullish on SK Hynix and Samsung, calling the correction a tactical entry and raising SK Hynix FY26 EPS +13%.
• Net: −11.7% over the window; −30.7% from the high; the sector in a >20% bear market.

What moves it now (revised ordering): (1) CXMT capacity news — newly promoted to first, it is the only near-dated supply variable; (2) memory contract/spot ASP direction, where the second derivative is now the watched series; (3) broad AI-capex narrative; (4) HBM4 allocation news; (5) guidance > print.

Upcoming catalysts: FQ4'26 print + first CY2027 colour (~late Sep — the decisive one, and remember the 14th week); CXMT Shanghai fab completion (2H CY26) and any HBM capex announcement it did not make in its prospectus; TrendForce 4Q26 contract-price forecast; HBM4 12-high multi-customer qual; the price-fixing class action's early docket.

---

Phase C — Judge people & books

Management (REFRESHED — the 10b5-1 trimming continued)

Sanjay Mehrotra, SanDisk co-founder, unchanged. Capital allocation remains disciplined and is now visible in primary form: $9,380M of debt repaid over 9M FY26, interest expense to zero in FQ3, and only $650M of buyback across the whole nine months — i.e. management is conspicuously not repurchasing stock at these prices. Reading that fairly cuts both ways: it is either admirable discipline near a cycle peak, or a signal about how management views the shares. The prior dossiers read it as the former; it is worth holding both.

The mild red flag persists and grew. Mehrotra sold 31,285 shares on 2026-07-24 at weighted-average $906.48–$941.60, under a Rule 10b5-1 plan adopted 2026-01-30; he directly holds 387,064 shares after. Over 18 months: 20 transactions, net ~360,000 shares sold. The 10b5-1Rule 10b5-1 planA schedule an insider sets in advance saying when their shares will be sold, so the sales cannot be timed on private news. A sale made OUTSIDE such a plan is a live decision, which is why it carries more information. adoption date matters and is exculpatory — the plan predates this quarter's information. But the cumulative direction is one-way selling through the largest melt-up in the company's history, and it stays on the ledger as a yellow flag rather than a red one.

The Anthropic Series-H stake (Jun 22) remains an undisclosed-size strategic investment in a key customer — consistent with buying demand-side optionality, terms still unknown.

Forensic red flags (REFRESHED — first primary balance-sheet audit on this shelf)

The FQ3'26 balance sheet and cash-flow statement were extracted from the 10-Q this pass. Three of four checks come back clean, and cleanly so — but the regulatory line was wrong and one new caveat emerges from the cash flow.

✅ Receivables — the scary-looking number is fine. Receivables went $9,265M → $31,025M (trade: $7,163M → $26,894M), and the 9M cash-flow statement shows a −$19,953M working-capital drag. That looks alarming until you compute days:

DSO ≈ 59.0 days at FQ3'26 versus ≈ 57.6 days at FY25 year-end.

Receivables tripled because revenue tripled. Collection quality is flat. This is the check that would have caught channel-stuffing or a quality-of-revenue problem, and it passes.

✅ Inventory — genuinely, unusually lean. Inventories were flat at $8,355M → $8,567M while revenue quadrupled, and finished goods FELL 43%, $1,094M → $621M. Work-in-process rose modestly ($6,401M → $6,960M). A memory company at a cycle peak with falling finished goods is selling everything it makes.

✅ Customer concentration improving, and it is filed — one customer 10% of 9M FY26 revenue vs 16% prior year [Note 17].

⚠ Cash conversion is good but flattered by deposits. 9M OCF $45,702M on 9M net income $47,268M = 96.7% conversion — excellent. But $5,203M (other noncurrent liabilities) + $2,139M (other current liabilities) = $7,342M of that inflow is customer deposit / unearned money, not earnings. Ex-deposits, conversion is ~81.2%. Still healthy, and the deposits are real cash — but it is cash Micron owes product against, and it should not be double-counted as both a moat and free cash flowFree cash flowCash left after paying to run and maintain the business. Unlike profit, it is hard to flatter with accounting choices..

⚠ Under-depreciation, now quantified from primary source. FQ3 D&A $2,364M against FQ3 gross capex $7,826M — 3.3x. Accumulated depreciation is $68,565M against $124,991M of gross PP&E, with $10,935M of construction-in-progress not yet depreciating at all and $4.20B of equipment not placed in service. This is a mechanical, dated FY28-29 cost headwind, not an opinion. Unchanged as a bear vector, now with a number on it.

⚠ The $22B is mostly not on this balance sheet yet. Other current + other noncurrent liabilities rose ~$7.8B between 2025-08-28 and 2026-05-28 — well short of the $18B cash deposits claimed, because most SCAs were signed after quarter-end. FQ4'26 is the first filing where the full deposit base should be visible. That is the single highest-value audit outstanding on this name.

❌ REGULATORY — the prior dossier's "clean" was wrong. regulatory/regulatory-findings.md (last fetched 2026-06-29) reports zero SEC LR and zero AAER — accurate, and it explicitly states that FTC/DOJ/other agencies are not indexed in EDGAR and require a web search in Lens 10. That search was not run in prior passes. It was run this pass: Samsung, SK Hynix and Micron were named in a US price-fixing class action over the memory shortage, ~2026-06-30. Also live: a separate report of a price-fixing suit against Micron. Plus the pre-existing May-2023 China CAC ban (geopolitical). This is a real overhang and it is aimed precisely at the thesis's load-bearing claim — that pricing power is contractual and structural rather than coordinated. It does not require merit to cost money or management attention. Contra the prior three dossiers, this name carries a live legal overhang.

---

Phase D — Project & stress-test

Forward Projection (RE-FORECAST — bear floor raised on hard contract terms)

⚠ Model integrity — read before any number below. A model workbook now exists at companies/micron/model.xlsx, built this pass. Its status, honestly:

• Fixed: our figures had period keys (Q3-2026) that fail the repo's schema (^\d{4}-Q[1-4]$). our model was reading 0 quarters and seeding a conventional opening balance sheet. Repaired to three fully-sourced quarters (2025-Q4, 2026-Q1, 2026-Q2) extracted from 10-Qs already on this shelf. Opening balance sheet now reads sourced: yes.
• Still broken for citation purposes: our model reports "No computed values — 51 formulas and no cached results." The workbook must be opened in Excel and saved before it produces anything. It also carries two live seed gaps: only 3 quarters of history (LTM partial) and revenue growth seeded at 8% because there are fewer than 8 quarters.
• Therefore: no valuation output from the workbook is cited anywhere in this dossier. Every figure below is hand arithmetic with the work shown. Citing a DCF built on 3 quarters and an 8% seed would be exactly the false precision this battery exists to prevent.

FY26 is effectively locked: revenue ≈ $129B, non-GAAP EPS ≈ $73. Caveat the FQ4 component for the 14th week.

Two revisions, in opposite directions, and they roughly cancel — which is why the base is unchanged.

(1) The bear's timing got worse. CXMT is funded, its fab completes 2H CY26, and its DRAM share is modelled 10%→18% by late 2028. The prior pre-mortem's H2 CY2027 date is now the right order of magnitude for CXMT specifically — where before it was a guess.

(2) The bear's floor got materially better supported — and this is the most valuable new arithmetic in the pass:

$100B of minimum revenue commitments across calendar 2026–2030 = ~$20B/year of contractually-minimum revenue, at floor prices management states clear gross margins above its ~62% all-time-peak quarterly GM. ~$20B × ~62% ≈ $12.4B of contracted gross profit per year, through 2030, before one dollar of spot. Against annual opex running ~$6.4B (FQ3 R&D $1,316M + SG&A $407M, annualised ≈ $6.9B), the contracted book alone roughly covers the entire operating cost base.

That is the structural change, and it is what "contracted like a utility" actually means in numbers. It makes a repeat of FY2023's −$5.83B loss (the same company, same assets, three years ago — preserved in notes.md) arithmetically very hard while the SCAs run. The bear is no longer "Micron loses money"; it is "Micron earns $15–20 instead of $105." That is a different and much better bear than memory investors are used to underwriting — and it is the single strongest argument for owning this name at all.

The honest counterweight, which must be stated in the same breath: that shield covers ~20% of DRAM volume. ~80% of DRAM bits are un-contracted and are exactly what CXMT's $8.6B is aimed at. And the ~$20B/yr floor is a minimum — it is the disaster case, not the expectation; in any decent scenario the SCAs price well above their floors and the number is far larger.

Brier forecast (carried, live, not re-logged): MU FY27 non-GAAP EPS ≥ $95, p≈0.58, resolves 2027-10-15. No re-log — no fiscal year has closed. Note for the record that the street's implied ~$142 forward EPS sits well above this bar, so consensus currently disagrees with a 58% probability in the optimistic direction.

Bull vs Bear (RE-JUDGED at $870)

Bull. Co-#2 in a three-player DRAM oligopoly that has converted the un-forecastable trough into $100B of take-or-pay minimums running to 2030 at floors above its best-ever margin, with $22B of customer cash on deposit, sold out of DRAM and HBM through all of CY2027, and no incumbent supply relief until 2028-29 on published construction schedules. The books survived their first genuine primary audit on this shelf: DSO flat at ~59 days through a tripling of revenue, finished goods down 43%, 96.7% cash conversion, customer concentration falling 16%→10%, interest expense at zero against $100.7B of equity. Two frontier-lab anchors (NVIDIA HBM + Anthropic primary-supplier/co-design). And at $870 you are 30.7% below the high with the fiscal year now closing at ~11.9x.

Bear. (1) ~80% of DRAM volume is un-contracted, and CXMT now has $8.6B and a fab completing this half aimed at exactly that. (2) The peak-margin objection is sharper than previously framed: the highest gross margin in the company (87.3%) is the commodity mobile/client segment, on rising price and falling bits, in the exact demand pocket TrendForce says is now balking. (3) Contract price growth has decelerated from ~+60% to +13–18% QoQ in two quarters — the direction of the second derivative is unambiguous. (4) Under-depreciation is a mechanical 3.3x gap that becomes an FY28-29 cost headwind. (5) Upside is capped — the largest SCA ceilings are pinned to CQ2-2026 prices, so the contracts that floor the downside also cap the upside, and at 6.11x the market is already using a bull-case E. (6) A live US price-fixing class action attacks the thesis's central claim. (7) The CEO has sold net ~360k shares across 18 months and the company bought back only $650M in nine months.

Pre-mortem (H2 CY2027 – CY2028). CXMT's Shanghai fab ramps through CY2027 into a consumer market that has already demonstrated it will not absorb further increases; MCBU's falling bit shipments turn into falling ASPs; the un-contracted ~80% of DRAM volume re-prices −35%; HBM4's premium compresses as all three makers ship into Vera Rubin with Micron holding the smallest allocation; the $27B/yr capex finally lands in D&A. GM goes 86% → ~45% over four quarters. FY28 EPS ~$15-20, "6x forward" reveals itself as ~45x collapsing earnings, and the class action settles for a number that matters. The SCA floors hold — and they are only 20% of DRAM volume, so they cushion rather than prevent.

Contrarian view. The market has taken 30.7% out of this stock while the contracts got longer (to 2030), the sold-out window extended (through all of CY2027), the floor got quantified (above a 62% all-time peak), and the competitor's own prospectus declined to fund HBM. Every operating fact that arrived this pass was neutral-to-good; the price fell on a competitor's share price. The catch — and it is a real one — is that 6.11x is not a cheap multiple on a conservative number; it is a sceptical multiple on a bull number. You are not being handed a mispricing. You are being asked whether $142 of forward EPS is real. The contracts say the floor under it is much higher than any prior cycle. They do not say the $142 is right.

Devil's Advocate (REFRESHED — the short case got two real gifts)

The short case is stronger than it was on 2026-07-07, and intellectual honesty requires saying so plainly.

What the short gained this pass:

1. The falsifier got funded. "CXMT floods the un-contracted book" was a hypothesis with no balance sheet behind it. It now has $8.6B, a ~$487B market cap, a fab completing this half, and a credible path from ~10% to ~18% of global DRAM. The single most-cited bear vector on this name stopped being speculative.
2. The "clean books, clean record" claim lost a leg. Three consecutive dossiers asserted regulatory-clean. A US price-fixing class action against all three DRAM makers was already on file. The bull case has been carrying an overhang it did not know about — and the claim under attack is precisely "our pricing is contractual, not coordinated."

And the short's own numbers are better than the bulls admit: the highest margin in the company is in its most commoditised segment, on falling volume; contract price growth has fallen from ~60% to 13–18% in two quarters; the CEO keeps selling; the buyback is conspicuously absent; the $50B guide is inflated by a 14th week that no coverage read this pass mentioned; and at 6.11x forward the market is already using an EPS number that equals the bull case, so "cheap" is doing far less work than it appears.

Where the short is nonetheless wrong today — and this is the crux:

• The falsifier requires prices to break, and prices are rising. DDR4 spot +16.9% in July; Q3 contracts +13–18% QoQ; server DRAM still undersupplied. A short thesis whose central mechanism has not begun is a bet on timing, not on facts in evidence.
• The deceleration is demand-side, not supply-side. TrendForce is explicit: buyers won't absorb more, not that supply improved. That is a ceiling on price, not a floor break — a very different thing for a company whose downside is contractually floored.
• CXMT's own prospectus earmarks nothing for HBM. The short's strongest capital event is aimed at commodity DRAM, not the rent.
• The incumbent supply the pre-mortem needed does not exist until 2028-29. SK Hynix's first new DRAM cleanroom is June 2029.

Net: the short is early and better-armed than a month ago. Its thesis is dated to CY2027-28 and its mechanism has a start date. The correct expression of the bear case is not a short here — it is a smaller long. Which is exactly the conviction move this dossier makes.

Management Questions (carried — see the previous dossier + the Anthropic addition in the previous dossier)

The 15 questions stand. Add three from this pass, all high information value: 16. "The $100B of minimum commitments covers ~20% of DRAM volume. What percentage of DRAM revenue does it cover — and how do you reconcile that with the ~40%-of-revenue figure in circulation?" (the surfaced conflict) 17. "FQ4 is a 14-week quarter. What is the $50B guide on a 13-week-equivalent basis?" (the question no analyst appears to have asked) 18. "CXMT's IPO prospectus earmarks no spending for HBM. Do you underwrite them as a commodity-DRAM competitor only, and what would change that?"

---

The MU-vs-SK-Hynix relative read (pair leg — catalyst RESOLVED)

The pair's dated catalyst has now happened, and the tilt survived it.

The prior dossier flagged the SK Hynix Nasdaq listing as "Hynix owns its own catalyst" — the one dimension where Hynix had the edge, an explicit dated event intended to force a Micron-level re-rate. It executed on 2026-07-10 at $26.5B, debuted +13% at $168.01 — and then the Korean line fell 15.4% three sessions later, its worst day on record, on HBM4 volume disappointment and operating profit tracking ~8% below consensus. Over July, SK Hynix fell ~33% against Micron's ~12%.

Scorecard against the four asymmetries the prior dossier named:

The tilt survives — but two things argue for holding it smaller, not larger. First, no co-dated relative valuation marks were obtained this pass, so I cannot say where the spread now sits; after a ~33% Hynix fall against ~12% for MU, the relative value has moved materially toward Hynix and the July parity read is stale. Second, Morgan Stanley's Shawn Kim — a long-standing memory bear — turned bullish on SK Hynix and Samsung on 2026-08-07, raising Hynix FY26 EPS +13% and calling the level a tactical entry. That is a direct, dated argument against the underweight leg from a source whose prior was the other way, and it deserves to be recorded rather than argued away.

Pair status: the thesis was validated by the catalyst, and the trade is more crowded-out by the price action than it was. A spread that has already moved ~20 points in your favour on the leg you were underweighting is a spread with less left in it. Falsifier unchanged: Hynix extends its HBM4 lead / MU's qual slips / the gap closes for the wrong reason. Level logic unchanged (beta-neutral, small, catalyst-not-value spread) — and re-mark both legs before acting; the valuation inputs here are stale and are labelled as such.

---

MarketCall review finding (assessment only — no row touched)

ACTIVE call — MEDIUM · "Micron: priced like a cyclical at the top, contracted like a utility."

Verdict: SUPPORTS — both halves of the premise strengthened, with one material new risk the call did not contemplate.

• "Contracted like a utility" — STRENGTHENED, materially. Every contract fact that arrived this pass improved: the term is now known (calendar 2026–2030), the minimum is known ($100B), the floor is benchmarked (above the ~62% all-time-peak GM), the sold-out window extended to all of CY2027, and management targets the book reaching 50%+ of revenue. The Lens 11 arithmetic — ~$12.4B/yr of contracted gross profit against a ~$6.9B opex base — is the most utility-like fact yet established about this company.
• "Priced like a cyclical at the top" — STRENGTHENED. The stock is −30.7% from its high in a sector bear market at 6.11x forward while guidance went up. The market is pricing it more like a topping cyclical now than when the call was written.
• ⚠ The new risk the call did not contemplate: the US price-fixing class action (2026-06-30) attacks the "utility" leg at its root — the claim that pricing power is contractual rather than coordinated. And the shield is thinner than the framing implies: ~20% of DRAM volume, not ~40% of revenue (conflict surfaced in the delta block, unresolved).

Net: the call's premise is intact and better-evidenced than when written. It is not a candidate for closure. The honest amendment is that the utility is a minority of the book and now carries a legal overhang. Any conviction or level change routes to our position log for Connor. No MarketCall row was read for edit, created, or modified.

---

Gaps

1. Model computes nothing — model.xlsx needs opening in Excel and saving; then it needs ≥8 quarters (5 more) before its revenue-growth driver is anything but an 8% seed. This is the highest-value follow-up on the shelf, and it is the missing piece of the Damodaran case study.
2. SCA coverage as a share of revenue — the ~20%-of-DRAM-volume figure and the ~40%-of-revenue figure cannot be reconciled from public disclosure. Management question #16.
3. The $22B deposit base is not yet visible on a balance sheet. FQ4'26 is the audit.
4. Micron-sourced HBM revenue / bit share — still third-party estimates only. Reports put Vera Rubin HBM4 at Hynix ~60-70% / Samsung ~25-30% / Micron the remainder, which would make MU the smallest NVIDIA supplier — materially thinner than the ~21% overall HBM share implies. Not filed anywhere; treated as unconfirmed.
5. Anthropic deal terms — exclusivity, volume, pricing, Series-H stake size: all undisclosed.
6. SK Hynix co-dated relative multiples — not obtained this pass; the pair's valuation leg is stale.
7. Price-fixing class action — no docket detail, claim size, or procedural posture obtained. Newly opened.