The only US-based maker of both DRAM and NAND, and the number-two force in high-bandwidth memory — the stacked DRAM that gates every AI accelerator. Multi-year HBM contracts are quietly turning a notorious memory-cyclical into a contracted-capacity supplier.
Our Position
Research
The verdict
"LONG — MEDIUM-HIGH. The −13% pullback to ~$985 (from the $1,133 that anchored the prior NEUTRAL) is the change that flips the entry: the quality is unchanged and confirmed (clean books, $22B prepaid take-or-pay raising the floor, data-center run-rate >$100B, floor-pricing margins management says beat any prior peak), but the price now pays a fair-to-cheap ~9x forward on a raised trough instead of a full ~8x on the rip. The bull case is no longer fully in the price. Falsifier is unchanged (DRAM/HBM ASPs roll over OR CXMT floods the un-contracted ~60% OR HBM4 multi-sourcing compresses the premium). Prior dossier said 'own the dips, not the rip' — this IS the dip it was waiting for."
Primary sources
SEC filings
Research Trail
Covered in the Knowledge Base
Hardware & Computing
The balance sheet is fixed but the business is not — at ~4x EV/sales with a -27% GAAP gross margin, a commoditizing substrate core, and a ~35% burned-strategic overhang (Renesas), WOLF is priced for an AI-datacenter / 200mm-utilization inflection the P&L will not confirm before FY27; WATCHING until gross margin crosses zero.
The purest listed play on the Big-3's DDR4 exit — a real supply-shock oligopoly windfall (GM 30%→53% in six quarters, sold out through 2027), but +890% in a year prices a transient legacy-DRAM squeeze as if it were a structural moat; WATCHING, not chasing, at 52× trailing peak-cycle earnings.
Source documents — open to read in full
FQ3'26 (quarter ended 2026-05-28) remains the latest and is fully carried from deep-dive-2026-06-29.md: revenue $41,456M (+346% YoY), GM 84.9% [transcript] / 84.6% [10-Q], operating income $33,318M (80.4%), net income $28,243M, GAAP diluted EPS $24.67 / non-GAAP $25.11, FCF $18.3B, net cash $24.4B. FQ4'26 guide unchanged: revenue $50B ±$1B, GM ~86%, non-GAAP EPS $31 ±$1, capex ~$10B → FY26 ~$27B.
What's new is only the tape context: the stock has since given back ~13% into a sector-wide semiconductor reassessment. Nothing in the fundamental picture changed — this is a multiple/sentiment move on unchanged (and still-guided-higher) earnings, which is precisely the setup a raised-floor long wants. The next fundamental catalyst is the FQ4'26 print (~late Sep 2026), which tests the $50B/86%/$31 guide AND (critically) whether management gives any first CY2027 colour — the single most view-changing data point outstanding.
Sentiment arc unchanged: "disciplined" (cyclical) → "structural shortage" → "transformed business model." The honest tell also unchanged — management won't guide past FQ4 and won't commit CY2027 pricing; the CFO did not deny a "normalized mid-70s" GM, only asserted floor pricing beats prior peaks. See prior dossier Lens 6 + the FQ3'26 transcript excerpt.
| Metric | MU (Jul 6 2026) | MU (prior, Jun 28) | Read | Source |
|---|---|---|---|---|
| Price | ~$984.75 (AH ~$960) | ~$1,133.50 | −13% in 8 days, AI-complex-wide | |
| Market cap | ~$1.11T | ~$1.27T | — | |
| Trailing P/E (GAAP) | ~22x | 25.6x | falling as E catches up | |
| Forward P/E | ~6.8x (Yahoo) / ~9.4x (date-lagged) | ~7.9x | cheap only if E is durable | |
| EV/EBITDA | ~15.8x | 18.4x | de-rated with the price | |
| Price/Book | ~2.2–2.5x | 2.52x | middling, not extreme |
The de-rate matters for the entry. At $1,133 the prior dossier's fair-value read was "~8x peak EPS, upside capped, bull priced." At ~$985 the same forward EPS band is ~9x trailing-style / mid-single-digits forward on the ~$73 FY26 EPS estimate, EV/EBITDA has fallen to ~15.8x, and the "premium to the HBM leader" flag has closed (MU ~15.8x vs Hynix ~17.85x EV/EBITDA — MU now the cheaper of the two on that measure). The valuation objection that drove the prior NEUTRAL is materially weaker at this price.
Peer frame (co-dated, Jul 2026):
The pattern is exactly as Lens 8 documented and the Jul-6 pullback confirms it: MU trades on AI-complex sentiment as much as its own fundamentals, with extreme two-way beta.
The EPS scaffold is unchanged from 2026-06-29 (no new quarter): FY26 ~75% actual → FY26 rev ≈ $129B / non-GAAP EPS ≈ $73. Scenarios carried:
| Scenario | FY26 (locked) | FY27 | FY28 | Drivers |
|---|---|---|---|---|
| Base | ~$129B / EPS ~$73 | ~$165B / EPS ~$75 | tightness holds to ~mid-CY27; GM eases 86%→high-70s as nodes + CXMT loosen commodity DRAM late CY27; SCA floors cushion | |
| Bull | ~$129B / ~$73 | ~$250B / ~$150 | tightness persists beyond CY27 (mgmt claim); SCA floors hold; HBM4 share/yield edge; CXMT sub-scale in HBM | |
| Bear | ~$129B / ~$73 | ~$90B / ~$18 | synchronized CY27 node ramp + CXMT flood + capex digestion; un-contracted ~60% re-prices −30-40%; depreciation steps up; SCAs defend ~40%, spot 60% craters |
What the −13% did to the read: at $1,133 the market priced ~$140 NTM EPS (~8x) — i.e. the bull. At ~$985 the market prices closer to my base (~$105 FY27) at ~9x forward — i.e. the price now demands only the base case, not the bull, to work. That is the asymmetry shift: you are no longer paying up for the best outcome; you're paying a fair multiple on the raised-floor base, with the bull as upside and the contracts (bear FY27 ~$60, not the old ~$30 no-contract trough) cushioning the downside. The un-contracted 60% still drives a brutal FY28 bear ($18) — the falsifier hasn't changed, only the price you enter at.
Brier forecast (carried, still live): MU FY27 non-GAAP EPS >= $95, p≈0.58, resolves 2027-10-15 [tracked — not re-logged; no fiscal year closed].
Bull (strengthened by the entry). Co-#2 (21%, deliberately capped near DRAM share) in a 3-player HBM oligopoly that has put $22B of prepaid customer cash + take-or-pay minimums on the books — converting the un-forecastable trough into a contracted floor management says beats any prior peak margin even at floor pricing. Books clean (97% conversion, lean inventory); disciplined peak capital allocation; now anchored by two frontier-lab demand relationships (NVIDIA-HBM + Anthropic primary-supplier/co-design); data-center run-rate >$100B; HBM4 shipping with a claimed yield-ramp edge. And after the −13% pullback the market prices only the base case (~9x forward), not the bull — the "it's all in the price" objection that drove the prior NEUTRAL has weakened with the tape.
Bear (impairment vectors — unchanged). (1) Still a commodity core; ~60% of revenue un-contracted / spot-exposed. (2) CXMT floods exactly that spot book (DDR5 −25-30% cost, state-subsidized; named in the 10-K). (3) ~85% GM is a cycle peak by definition — all four segments at 79–87% simultaneously is the textbook top; CFO didn't deny mid-70s normalized. (4) Under-depreciation becomes a real 2027-28 GM headwind. (5) Upside capped — largest SCA ceilings pinned to CQ2-2026 peak. (6) Extreme AI-complex beta cuts both ways — the −13% this week is the same mechanism that can deliver −35% on a genuine cycle-turn signal.
Pre-mortem (H2 CY2027): nodes ramp on schedule across all three incumbents + CXMT adds commodity DDR5; AI-capex growth decelerates as hyperscalers digest 2025-26 buildouts; un-contracted 60% re-prices −35%; HBM4 multi-sources and its premium compresses; the ~$27B+ capex now depreciates. GM 86%→~50% in three quarters, FY28 EPS ~$18, "9x forward" becomes 55x collapsing earnings.
Contrarian view. The market still half-prices a roll-over (~9x NTM) the contracts are built to prevent — and after this week's sentiment-driven de-rate, it's pricing the base, not the bull, on a business whose hardest-ever evidence ($22B prepaid) says the trough is structurally higher. Customers don't prepay $18B cash for capacity they expect to be cheap. The catch remains: at $985 you still own a commodity cyclical near a margin peak — but you're now paid to take that risk, not paying up to.
The short case is structurally unchanged and worth keeping honest: the "$100B contracted" is ~$5B audited RPO; ~60% of revenue is spot into a CXMT flood; management won't guide past Q4 or commit CY2027; the CEO is trimming into the rip; ~85% GM is the most mean-reverting number on the tape; under-depreciation guarantees a 2027-28 cost headwind. The one update that helps the short: the −13% pullback shows how fast this equity de-rates on sentiment alone — a genuine cycle-turn signal (not just sentiment) would take it far more than 13%. The counter: the drawdown is sentiment-driven with no fundamental deterioration, and it improves the long's entry rather than validating the short. See prior dossier Lens 13 for the full teardown.
The 15 questions stand (ordered by information value: the SCA floor-margin claim #1, FY27 %-protected #2, RPO reconciliation #3…). Add: "The Anthropic partnership — is it exclusive on any product line, what volume does it represent, and how does it interact with your NVIDIA-HBM allocation given HBM's trade-ratio pressure on non-HBM supply?"
The pair expresses "the memory oligopoly wins" without market beta: long the better risk/reward leg, underweight the other. Brief lean = long MU / underweight Hynix. The battery's job is to confirm or reject that lean on the refreshed marks.
The two legs, side by side:
| Dimension | Micron (MU) | SK Hynix (000660.KS) | Edge |
|---|---|---|---|
| HBM share | ~21% (deliberately capped near DRAM share) | ~50–62% (leader); ~mid-50s–70% of NVIDIA HBM4 | Hynix (rent-capture) |
| Op margin (latest) | 80.4% (FQ3'26) | 72% (Q1'26) | MU (but both peak) |
| Fwd P/E (co-dated) | ~6.8x | ~6.06x | ~parity (Hynix slightly cheaper) |
| EV/EBITDA (co-dated) | ~15.8x | ~17.85x | MU cheaper (reverses prior dossier) |
| Contracted floor | $22B prepaid, ~40% rev at fixed/ceiling, "beats prior peak at floor" | Sold-out 2026 + prepaid, "short into 2027/2030" | ~tie (both contracted) |
| Customer concentration | NVIDIA-HBM + now Anthropic primary-supplier (diversifying) | ~2/3 HBM to NVIDIA's roadmap | MU (less single-customer) |
| HBM4 momentum | >$1B HBM4 shipped; claims mature yields faster than HBM3E | Leader, but Samsung out-qualified it on HBM4 timing; keeps allocation on yield-at-volume | contested — MU the catching-up momentum story |
| Re-rating catalyst | US-listed already; SK-Hynix Nasdaq listing is a read-across tailwind | $14B Nasdaq listing (H2 26) explicitly to force a Micron-level re-rate | Hynix (owns its own catalyst) |
| Liquidity / access | US large-cap, deep options, index | Korea-listed (ADR-limited until the Nasdaq float); chaebol governance discount | MU (US liquidity) |
The relative call — long MU / underweight Hynix — survives, but the rationale is refined:
Pair falsifier (from the brief, confirmed): the spread driver reverses — Hynix extends its HBM4 lead (Samsung fails to converge and Hynix widens allocation), OR MU's HBM4 qual/yield slips, OR the gap closes for the wrong reason (MU re-rates on hype not delivery). Watch: HBM4 multi-customer qual results (H2 CY26), the SK-Hynix Nasdaq-listing pricing (does it re-rate Hynix past MU?), and relative DRAM/HBM ASP data.
Pair level logic: beta-neutral sizing (dollar-matched or beta-adjusted long-MU / short-or-underweight-Hynix); entry on the current ~parity relative-value spread; target = MU convergence/outperformance as its HBM4 ramps and the Anthropic/US-liquidity premium is recognized; stop = if Hynix's HBM4 lead visibly widens (the pair's thesis-break). Because the legs are at valuation parity, this is a catalyst/momentum spread, not a mean-reversion spread — size it accordingly (smaller; it's a tilt, not a screaming value gap).
A textbook semiconductor cyclical wearing a secular-growth costume — the 2024–25 CIS recovery is real and automotive is a genuine share-taking engine, but 2026 consensus EPS has already been cut BELOW 2025's actual and Q1-2026 net profit fell ~42% YoY while the stock still holds ~35x trailing; the automotive ramp must out-run mobile normalization and domestic price competition to justify the multiple. WATCHING — constructive only on a reset toward ~25x or hard evidence auto mix is structurally l