MU › analyze
MU · Analyze
Price: $892.67 · Market cap: $1.01T · Sector: Semiconductors (DRAM/NAND/HBM) · Beta 2.21 Verdict: 🔴 TRIM · Conviction [4.0] (as a long-term holding — this says nothing about near-term direction, which could easily be higher)
A portfolio-specific passage was removed from the public build.
The one-paragraph version
Micron is printing the best numbers in its 48-year history — an 84.9% gross margin and an 80.4% operating margin on a commodity product, revenue up 346% year over year, and debt cut from $16.1B to $6.4B in four quarters. It has also just signed something genuinely new: 16 take-or-pay customer agreements running to 2030 with ~$100B of contracted minimum revenue and $22B of customer collateral posted. None of that is in dispute, and it is why this report credits Micron's normalized earnings power at five times its pre-HBM decade average rather than dismissing the cycle as a repeat.
The problem is the price. At $1.01T the market is capitalizing earnings that require an ~86% gross margin to persist, in an industry whose prior all-time peak was ~62% and which posted a negative gross profit three years ago. The deeper problem is that even in the best years of the best cycle Micron has ever had, this business does not convert profit into owner cash: cumulative free cash flow across FY2021–FY2025 was $1.22B on $136.5B of revenue — a 0.90% five-year FCF margin. And the decisive test is that the bull case does not survive its own best argument: grant three more years at the current run-rate with 100% of cash returned, and you still end up holding the residual at 21–34x through-cycle earnings.
1. General Health (§1)
Revenue and margins — the most violent operating leverage in semiconductors
| Quarter (FY ends Aug) | Revenue | Gross margin | Op margin | Diluted EPS |
|---|---|---|---|---|
| FQ3 FY25 (May-25) | $9.30B | 37.7% | 23.3% | $1.68 |
| FQ4 FY25 (Aug-25) | $11.31B | 44.7% | 32.6% | $2.83 |
| FQ1 FY26 (Nov-25) | $13.64B | 56.1% | 45.0% | $4.60 |
| FQ2 FY26 (Feb-26) | $23.86B | 74.4% | 67.6% | $12.07 |
| FQ3 FY26 (May-26) | $41.46B | 84.6% | 80.4% | $24.67 |
| FQ4 FY26 guide | $50B ± $1B | ~86% | — | $31.00 ± $1.00 |
Revenue +346% YoY; the sequential increase of +$17.6B is the largest in company history. DRAM was $31.3B (76%) with pricing +low-60s%; NAND $9.9B (24%) with pricing +mid-80s%. Bit volumes grew only low-to-mid single digits — this is essentially all price.
Segment revenue FQ3: Cloud Memory $13.77B (83% GM) · Core Data Center $11.52B · Mobile & Client $11.52B · Auto & Embedded $4.63B. Datacenter-facing is 61% of revenue.
🚩 The single most important table in this report
Micron's ability to turn accounting profit into owner cash, across a full cycle:
| FY | OCF | Capex | FCF |
|---|---|---|---|
| 2021 | $12.47B | −$10.03B | +$2.44B |
| 2022 | $15.18B | −$12.07B | +$3.11B |
| 2023 | $1.56B | −$7.68B | −$6.12B |
| 2024 | $8.51B | −$8.39B | +$0.12B |
| 2025 | $17.53B | −$15.86B | +$1.67B |
| Cumulative | $55.25B | −$54.03B | +$1.22B |
$1.22B of cumulative free cash flow on $136.5B of cumulative revenue — a 0.90% FCF margin over five years. FY2025 was a record profit year ($8.54B net income) and produced $1.67B of free cash. Capex ate 98% of five years of operating cash flow.
This is the §1.1 test ("FCF is the backbone… harder to fake than earnings") and it is the most damning number in the file. It is also why the trailing P/E flatters: the earnings are real, and they do not become cash.
The current cycle is converting better — for now
| ttm (Aug-25 → May-26) | |
|---|---|
| Revenue | $90.27B |
| Operating income | $59.29B (65.7% margin) |
| Net income | $50.47B (55.9% margin) |
| Diluted EPS | $44.17 |
| OCF | $51.43B |
| Capex | −$25.27B |
| FCF | $26.17B → $23.16/share |
FCF conversion is 52% of net income — vastly better than history, and still only half of earnings. And it gets worse from here: FY2026 capex is ~$27B (guided up twice, from $18B → $20B → ~$27B), and FY2027 quarterly capex is guided above $10B, implying >$40B. Capex is running at roughly 65% of the prior full year's revenue. That is textbook top-of-cycle capital deployment.
The balance sheet looks transformed — but a large part of the cash is not Micron's
| May-2025 | May-2026 | |
|---|---|---|
| Total debt | $16.14B | $6.38B |
| Cash | $10.16B | $25.00B |
| Net cash | −$5.98B | +$18.62B |
| Equity | $50.75B | $100.72B |
| Debt/assets | 20.6% | 4.8% |
| Non-debt liabilities | $11.51B | $27.01B |
✅ The deleveraging is real: debt cut by $9.8B in four quarters, and Micron enters whatever comes next with a far better balance sheet than the one it took into FY2023. That matters and it is the strongest bull fact in this section.
🚩 But the net-cash headline is substantially customer-financed. Over the same four quarters, non-debt liabilities rose +$15.5B ($11.51B → $27.01B) while cash rose +$14.8B — the two move almost in lockstep. Management disclosed $22B of customer deposits and financial commitments, ~$18B of it cash, received progressively (~$0.5B in FQ3, ~$10B expected in FQ4). These deposits are a refundable liability, returned to customers over the term and back-end weighted to 2029–30.
Strip the customer money and Micron is approximately net-cash-neutral — while committing to $40B+/yr of capex and a deposit-repayment wall that lands in the same years the SCAs expire and CXMT reaches its 17% share target. The "fortress balance sheet" is, in material part, a customer-financed working-capital loan being reported as strength. (Yahoo does not break out a customer-deposit line; the inference rests on the aggregate liability movement plus management's disclosure, and is flagged as such.)
Shares outstanding and capital allocation
Diluted shares 1.12B → 1.15B (+0.1%/yr over three years) — essentially flat, no meaningful dilution and no meaningful buyback. Buybacks were $0 in the May-2026 quarter and $350M the quarter before. Dividends are $0.17B/quarter (0.06% yield — immaterial).
Management is not buying back stock at $892 after an 8x run. Read generously, that is discipline; read plainly, management is not a buyer of its own stock at this price either. Capital-return restrictions from the CHIPS agreements lift on 2026-12-09, after which management intends to return "100% of excess cash."
Growth (§1.5)
| Metric | 3yr CAGR (FY22→FY25) |
|---|---|
| Revenue | +6.7% |
| Net income | −0.6% |
| FCF | −18.8% |
| Shares (diluted) | +0.1% |
Three-year CAGRs through a trough are almost meaningless for a cyclical — included because the framework requires them, and because they show the honest baseline the current quarter is measured against.
2. Moat (§2) — NARROW, at the weak end
ROIC through the cycle — the framework's §2.1 quantitative base
| FY | NOPAT | Invested capital | ROIC |
|---|---|---|---|
| 2022 | $8.54B | $55.9B | 15.3% |
| 2023 | −$4.76B | $56.2B | −8.5% |
| 2024 | $1.14B | $56.5B | 2.0% |
| 2025 | $8.63B | $65.7B | 13.1% |
| 4yr average | 5.5% | ||
| ttm (peak) | $50.52B | $82.1B | 61.5% |
Long-run median ROIC ≈ 4%, median ROE ≈ 7%, with roughly one-third of all quarters in outright capital destruction. Against a cost of capital appropriate to a beta-2.21 business, that is a multi-decade record of value destruction interrupted by two-year windfalls. Gross margin by FY: 45.2% → −9.1% → 22.4% → 39.8% → 72.6% ttm → 84.6% latest quarter.
A company with a cost moat does not sell below cash cost in a downturn. Micron's FY2023 gross profit was −$1.42B.
Moat sources, rated
| Source | Rating | Evidence |
|---|---|---|
| Efficient scale | Present (strong) — but industry-owned | Three-player DRAM oligopoly, ~93% combined share; $20–30B and 4+ years to add a fab. This is the real moat, and it belongs to the industry, not to Micron — which is the smallest of the three holders. Being diluted by CXMT reaching ~350K wafer-starts/month by end-2026 vs Micron's ~375K. |
| Switching costs | Weak — protecting the wrong thing | HBM qualification is 9–18 months and design-locked (Samsung's 18-month HBM3E failure proves the cost is real). But Nvidia has now qualified all three suppliers on HBM4/Vera Rubin, so the barrier protects the supplier class, not Micron. Genuine per-customer stickiness exists only in Auto & Embedded — 11% of revenue. |
| Cost advantage | Weak / absent | #3 scale player, historical process follower. FY2023 negative gross profit. 1-beta/1-gamma node claims are parity at best. |
| Intangibles | Weak | Large patent estate, broadly cross-licensed. Memory sells on spec and price; no brand pricing power. |
| Network effects | Absent | A DRAM bit is no more valuable because someone else bought one. |
Is HBM a moat or a cycle? — a cycle with a gate that is closing
The bull case rests on HBM being a durable, qualification-gated premium business. The evidence says the gate is being opened by the customer who owns it:
- Nvidia has certified all three suppliers for HBM4 on Vera Rubin. Reported allocation: SK Hynix 60–70%, Samsung 25–30%, Micron single-digit to low-teens. Micron's stated 20–25% HBM share is an aspiration, not a position.
- TrendForce reported (Feb 2026) that Nvidia may relax HBM4 specifications because Samsung and SK Hynix face yield limits. Every relaxation converts a qualification moat into a purchase order. Nvidia's interest is three healthy suppliers, not one privileged one.
- Micron caps its own HBM share deliberately — Mehrotra: "we strategically are choosing it to be close to our DRAM share… because of the trade ratio."
What is real and durable: HBM consumes roughly 3x the wafer area per saleable bit versus DDR5, which structurally absorbs industry supply and tightens commodity DRAM. That is the best structural argument in the memory bull case — but it is an industry effect, and Micron is the smallest of its three beneficiaries.
🚩 The SCA structure — the genuinely new fact, and it cuts both ways
Independently verified (FQ3 FY26 call, Investing.com/Yahoo/wccftech corroboration):
- 16 Strategic Customer Agreements, typically 5-year, CY2026 → CY2030, take-or-pay with binding volume commitments
- Cover ~20% of DRAM volume and ~33% of NAND volume
- ~$100B of contracted minimum revenue across 14 of the 16
- $22B of customer deposits and financial commitments, ~$18B in cash, held by Micron for the term
- Management: the floor price yields gross margins "well above our peak quarterly margins in any past cycle" (prior peak ≈ 62%)
The $18B of customer cash on deposit is the hardest available evidence that the shortage is real. Customers do not post collateral for a shortage they doubt.
But the pricing mechanism is a collar, not a floor:
The largest agreements carry a price CEILING pinned at the CQ2-2026 market price — approximately today's peak — with a floor beneath, and quarterly market resets inside the band.
⚠️ Important qualification that most bear framing (including my own first draft) gets wrong: the ceiling applies to existing products only. HBM, DDR6 and LPDDR6 are explicitly excluded and negotiated separately. So the cap does not choke the AI growth line. Anyone arguing "Micron capped itself at the peak" is overstating it — the cap binds the legacy book.
The sharper and still-valid point is about who chose to hedge. In July 2026 SK Hynix removed the price cap from its own long-term agreements entirely, becoming the only major supplier without one — while holding 60–70% of HBM4 allocation against Micron's single-digit-to-low-teens.
Micron sold away the legacy upside from the weaker competitive position, to buy a floor it needed because it has never earned its cost of capital. A hedge bought by the player with the worst hand is a rational trade and a poor moat. It converts Micron from an option on memory prices into a spread — and on cyclicals, the option is where the returns live. You capture the floor; SK Hynix captures the shortage.
🚩 Do take-or-pay contracts actually hold in a bust? The historical record
This is the load-bearing question under the entire "the cycle is different now" thesis, and it has a real answer.
| Precedent | What happened |
|---|---|
| Hemlock v. SolarWorld (2011–17) — the only cleanly litigated semiconductor take-or-pay | Polysilicon LTAs at fixed prices; subsidized Chinese supply crushed the market. Parties renegotiated in 2011; when Hemlock demanded the original 2012 price, the customer refused. ~$800M awarded, affirmed by the 6th Circuit (867 F.3d 692). Take-or-pay held in law — but SolarWorld went insolvent and Hemlock recovered minimally or not at all. |
| 2017–19 memory | Forward purchase agreements signed in the 2017 shortage. When demand rolled, DRAM fell 40%+ in 2–3 quarters, customers delayed taking delivery, and contract prices were renegotiated back toward spot. Prepayments then were under 5% — there was nothing holding them. |
| 2001 / 2008 / 2015 / 2022–23 | No enforceable LTA regime existed. Memory LTAs were, in the industry's own phrase, "gentleman agreements in a business where there's no gentlemen." Both sides reneged routinely. |
The honest synthesis — and it lands in the middle. Take-or-pay is enforceable, and Micron's counterparties (Microsoft, Google, Meta, Amazon, Nvidia) are vastly better credits than SolarWorld. That asymmetry is real and favours Micron. But Hemlock won and still lost: the contract converts price risk into counterparty and relationship risk, and you do not sue Microsoft in Q3 and sell it HBM5 in Q4. Renegotiation in a bust is likely, not impossible — but it now starts from a much higher floor than 2019, because $22B of posted collateral makes walking away genuinely expensive.
The contracts compress the trough. They do not abolish it. That is the correct weight to give the single most important new fact in the bull case.
Adversarial stress-test (§2.2)
- As CXMT: I don't fight for HBM. I take the commodity floor out from under you. I hold ~7–10% of global DRAM, I reach ~350K wafer-starts/month by end-2026 — approaching your own capacity — I target 17% share by 2028, I just IPO'd on the Shanghai STAR market, and I am not required to earn a return on capital. I already out-priced Samsung on some DDR5 server parts in early 2026. Your answer was to lobby Congress to tighten tool exports — an admission that the defence is political, not competitive. Political moats can be revoked by an election.
- As Samsung: I have the balance sheet to lose money longer than you can. I did it in 2007–08 and 2015–16 and it worked both times. I lost the Blackwell cycle and I'm angry about it; I'm at 25–30% of Vera Rubin HBM4 already, and my P5 Pyeongtaek fab lands in 2028 — precisely when your greenfield bits arrive. If HBM demand disappoints even modestly, I convert HBM capacity back to commodity DRAM and price you into another FY2023.
- As SK Hynix: I don't need to attack. I hold 60–70% of HBM4 allocation and I just removed my price cap while you kept yours. You capture the floor; I capture the shortage.
- As a hyperscaler / ASIC designer: I don't build fabs — that's not the attack. Custom HBM is. Marvell's architecture redesigns the base die to dovetail with my compute die. The base die is where differentiation and margin migrate — to Marvell, Broadcom and TSMC. I'm not disintermediating Micron; I'm relegating it to supplying the commodity stack that sits on top of my proprietary logic.
- Architecture: SOCAMM2 / LPDDR6 modules launch in 2026, projected at 25–30% of AI inference memory by 2027 at near-HBM performance and materially lower cost. Genuinely two-sided — Micron sells LP server DRAM too — but it caps HBM's TAM expansion with a cheaper, less-gated product.
Evergreen assessment (§2.5) — memory is evergreen; Micron's earnings are not
DRAM demand in 2036 will be far larger than today; that is not the question. The question is whether a supplier retains a share of the value it creates, and thirty years of evidence says the industry gives its economics to customers through the price mechanism and keeps only what the capital cycle allows — a 4% median ROIC.
This is an infrastructure business that must spend to survive, in which spending is competitive rather than accretive. It is closer to an airline or a shipping line than to a compounder. It is not a hold-forever asset. It is a position you size for a cycle and exit on a multiple — and the framework's core principle applies with unusual force: how good the company is matters far less than how the market has priced it.
3. Valuation (§3)
DYT and DDM are N/A (0.06% yield, 1% payout). Bogle is N/A in any useful form — "earnings growth" for a cyclical at peak is cycle noise, not a trend, and plugging +1,369% YoY earnings growth into an expected-return model produces nonsense. The work is carried by normalized earnings and Graham.
What the price requires
| Basis | EPS | P/E at $892.67 |
|---|---|---|
| FQ4 FY26 guide annualized | $124.00 | 7.2x |
| Consensus NTM (implied by Yahoo's 5.74x fwd) | $155.52 | 5.7x |
| ttm actual | $44.17 | 20.2x |
| Pre-HBM decade average (FY2016–25) | $3.94 | 227x |
| EPS required for 15x | $59.51 sustained | 15.0x |
| EPS required for 20x | $44.64 sustained | 20.0x |
The inversion — and it is the crux
A 5.74x forward multiple is not the market offering value. It is the market pricing a collapse. Cyclicals earn low multiples at peak earnings and high multiples at trough earnings. The correct question is therefore not "is 5.7x cheap?" but "is the discount the market is already applying severe enough?"
The payback test answers it:
| Assumed sustainable EPS | Net income | Years of earnings to repay the $1.01T market cap |
|---|---|---|
| $155.52 (consensus NTM) | $178.8B | 5.7 |
| $124.00 (FQ4 guide) | $142.6B | 7.1 |
| $60.00 (strong normalized) | $69.0B | 14.7 |
| $35.00 (good normalized) | $40.2B | 25.2 |
| $12.00 (3x the pre-HBM decade) | $13.8B | 73.4 |
And on cash rather than earnings: at the FQ4 guided run-rate (~$143B net income) with the ttm 52% FCF conversion — which is optimistic given capex is heading above $40B — free cash flow is ~$74B, so the market cap is ~13.7 years of peak free cash flow. That is not a cheap asset even if you assume the peak lasts.
Normalized earnings, built on the actual SCA split
Rather than one blended guess, model the contracted slice (floored, near-peak margin) separately from the exposed slice (fully cyclical):
Assumes ~25% of revenue contracted at a 70% gross margin, opex ~$6.6B/yr, 15% tax, 1.15B shares.
| Revenue | Uncontracted GM 25% | 35% | 45% | 55% |
|---|---|---|---|---|
| $55B | $10 (90x) | $13 (69x) | $16 (56x) | $19 (47x) |
| $70B | $14 (64x) | $18 (50x) | $22 (41x) | $26 (35x) |
| $85B | $18 (50x) | $23 (40x) | $27 (33x) | $32 (28x) |
Every cell in that table is a P/E above 28x on normalized earnings. Even the most bullish corner — $85B of revenue with the uncontracted book still earning a 55% gross margin, which would be above Micron's all-time peak margin for the whole company — leaves the stock at 28x.
Scenario-weighted fair value
| Scenario | Prob | Normalized EPS | Multiple | Value | vs spot |
|---|---|---|---|---|---|
| Supercycle holds to 2030 | 15% | $90 | 12x | $1,080 | +21% |
| Structural re-rate, softer | 35% | $45 | 14x | $630 | −29% |
| Normal good cycle | 30% | $25 | 13x | $325 | −64% |
| Cycle turns hard | 20% | $8 | 15x | $120 | −87% |
| Probability-weighted | ~$504 | −44% |
Graham cross-check (√(22.5 × EPS × BVPS), BVPS $89.22): at ttm EPS $44.21 → $298; at $25 normalized → $224; at $12 → $155. Graham on peak EPS already says the stock is 3x its intrinsic value.
Normalized EPS, built bottom-up from the actual contract structure
The debate round rebuilt this from the disclosed structure rather than a blended guess:
| Layer | Basis | Normalized EPS | Confidence |
|---|---|---|---|
| Contracted floor | ~$20B/yr revenue @ ~67.5% GM, taxed | $8.90 | High — legally floored |
| HBM | ~12% share of an $80–100B TAM @ 55–60% GM | $3–6 | Medium — share loss to SK Hynix binds |
| Uncontracted DRAM/NAND | $60–80B @ 30–35% through-cycle GM | $2–14 | Low — this is the swing, and it is the CXMT-exposed book |
| Total normalized | ~$20 (range $14–30) |
$20 is 5.1x the pre-HBM decade average of $3.94. That credits a genuine five-fold structural step-up — from consolidation to three players, from HBM existing at all, from the new contract regime, and from rising capital intensity as a barrier. That is an aggressive credit and it is on the record that it was given. The bear's $3.94 anchor is too harsh; the bull's $124–155 is peak-of-peak.
🔑 The decisive calculation — run the bull's own best argument to its conclusion
The strongest bull case is the payback: "grant three years at the current run-rate and 100% of excess cash returned from 2026-12-09." Grant it in full:
- Three years at $75–105B of annual FCF = $225–315B returned, or 22–31% of the market cap
- You are then left holding the residual: $700–785B against ~$23B of normalized net income = 30–34x through-cycle earnings
- Even on the generous branch ($30 normalized EPS), the residual is ~21x
The payback math does not rescue the valuation — it caps the bull case. Even if everything goes right for three more years, you end up owning a deep cyclical at 21–34x mid-cycle earnings. This is the single most important calculation in the file.
Fair value
$280 – $560, midpoint ~$400. Reconciled from two independent methods that broadly agree: - Normalized EPS × cycle-appropriate multiple: $20 (range $14–30) × 12–18x = $240–540. Deep cyclicals with 40%+ capital intensity historically earn 8–12x mid-cycle; the 12–18x used here already credits the contract regime and the three-player structure with a premium. - My scenario-weighted table above: ~$504, which is higher because it carries a 15% weight on the supercycle branch.
The range is deliberately enormous because the honest distribution is enormous — a meaningful probability of $120 alongside a meaningful probability of $1,080 should not be collapsed into a point estimate. Spot at $892.67 is roughly 2.2x the midpoint and sits above the top of the primary band.
For scale: $892.67 requires $59.51 of sustained EPS at 15x — 3x the normalized estimate, and 15x the pre-HBM decade average.
4. Sentiment — euphoria in estimates, fear in the tape
Sell-side positioning is at outright euphoria. 46 ratings: 9 Strong Buy, 32 Buy, 5 Hold, 0 Sell. Mean target $1,507.79 (+68.9%), median $1,550, high $2,200. Every rating action in the last four months was a price-target raise — not one downgrade, not one cut. UBS moved $535 → $1,625 and Stifel $550 → $1,500 in ten weeks. Those are not analyses; they are catch-up. The disciplined reads are the lowest targets: Goldman Sachs is the only Neutral at $1,100, and Morgan Stanley — historically the most cycle-aware memory desk — carries the lowest Overweight at $1,200. When Morgan Stanley's bull case is 35% below the consensus mean, the consensus mean is the outlier.
The tape disagrees with the sell-side. The stock is −28.9% from its $1,255 high and fell ~20% in July alone, on Samsung's earnings, the CXMT IPO and a Kospi rout. This is a multiple de-rating running ahead of a fundamental turn.
🚩 Insiders: ~$233M sold in six months, essentially zero buying
Net −306,707 shares (−9.4% of insider holdings); insiders own just 0.26% of the company.
- CEO Mehrotra ≈ $144M since Feb 2026, selling in essentially every month since Sept 2025 at $130, $160, $200, $220, $340, $520, $960, $1,150, $930. The cadence is consistent with a 10b5-1 plan, which blunts the signal considerably.
- Two transactions do not look programmatic: Mehrotra sold $46.3M at $1,128–1,192 on 2026-06-26 — two days after the largest beat in company history. And officer April Arnzen sold 40,000sh at $1,077–1,096 on 2026-07-01, within 13% of the all-time high and 5x her prior clip size at $345 three months earlier.
- 🟢 The one genuine positive: Director Mark Liu (ex-TSMC chairman) bought 23,200sh at ~$337 on 2026-01-14 (~$7.8M open-market) — high-quality, +165% in the money, and seven months stale.
Read: weakly bearish. Programmatic selling by an executive team whose stock rose 654% is normal diversification. The absence of any buying at −29% off the high is the real negative — if management believed $892 were the bargain the sell-side claims, the December buyback unlock would not be the only capital they are willing to commit.
Cycle state — still rising, decelerating hard
| 2Q26 | 3Q26 forecast | |
|---|---|---|
| Conventional DRAM contract | ~+60% QoQ | +13–18% QoQ |
| Server DRAM contract | — | +13–18% QoQ |
| NAND contract | ~+85% realized | +10–15% QoQ |
The reason for the deceleration is bearish-flavoured: TrendForce attributes it to buyer affordability limits — PC, smartphone and consumer OEMs can no longer absorb further increases — not to improving supply. That is demand destruction at the edge, which historically precedes the turn. Management's own FQ4 language flags "a meaningful moderation in the rate of price increases" while blended DRAM cost per bit is projected to rise. Cost per bit rising while price increases moderate is the definition of margin peak formation, even if the peak is a high plateau.
NAND is where the crack is: TrendForce forecasts NAND supply growth outpacing demand in 2027, with 2026's 4–5% deficit flipping positive. NAND is 24% of revenue and carried +85% pricing last quarter — the highest-beta line to a turn.
Sentiment verdict: EUPHORIA in narrative and estimates, FEAR in positioning.
5. Verdict
🔴 TRIM (existing holder) · AVOID (new capital) — conviction [4.0]
The [4.0] is a quality-and-attractiveness rating on a 0–10 scale, not a confidence level. Confidence in the TRIM call itself is high (~7.5/10). Those are different numbers and conflating them is how people talk themselves out of acting.
The crux: is the market's own 5.74x forward multiple already discounting the bear case?
This is the question the whole analysis turns on, because a 5.74x forward P/E looks like the market has already done the bears' work. The answer is no — not adequately — and the low multiple is a warning label on the denominator, not a margin of safety.
The arithmetic. 5.74x on $892.67 implies forward EPS of $155.52 — which is above the $124 annualized FQ4 guide. Consensus is not modelling the peak holding; it is modelling the peak rising another 26%. The multiple is low because the E is peak-of-peak-plus-growth.
The market is not naive, though. If it believed $155 were durable, even 12x puts the stock at $1,860. It trades at $892.67. Reverse-engineered, the market is capitalizing roughly $59.51 of sustainable EPS at 15x — it has already priced a 62% peak-to-normalized decay. That is a serious discount and the bear case has to acknowledge it.
It is still not enough, for three reasons:
- 62% is a normal-cyclical haircut applied to an abnormal peak. Micron's actual peak-to-trough EPS decay has been 90–100%+ — FY2023 produced negative gross profit. A defensible through-cycle $20 is 87% below the forward figure, not 62%.
- The discount contains no CXMT. Nothing in a 38%-of-peak normalization prices a state-financed competitor taking ~10 points of share into a market whose total bit supply grows only 6–8%/yr — aimed squarely at the 60–80% of volume that has no contractual floor. CXMT's cost of capital is a policy variable, not a market one; it does not need to clear a hurdle rate to keep building.
- It contains no capex reckoning. $40B+/yr against any plausible normalized revenue of $90–110B is 40–50% of sales — a level Micron has never sustained. The market is implicitly assuming either that revenue never normalizes or that capex flexes costlessly. Neither has ever been true in this industry. And the new capacity lands in 2028–2030 — exactly when the SCAs expire and CXMT peaks.
The market is pricing a cyclical downturn. It is not pricing mean reversion, and it is not pricing CXMT.
Why this is a TRIM and not an AVOID-and-exit-everything
Three things genuinely changed, and they deserve to be stated as clearly as the negatives:
- The contract regime is real and unprecedented. $100B of take-or-pay RPO with $22B of posted collateral is not a gentleman's agreement. It compresses the next trough.
- The HBM/DDR6/LPDDR6 exclusion from the price cap means Micron's AI line is not capped — the bear framing on this was too strong.
- The balance sheet, even discounted for customer deposits, is far better than the one that met FY2023.
These are why normalized EPS gets credited at 5.1x the pre-HBM decade rather than 1x. They change the floor. They do not justify the price.
A portfolio-specific passage was removed from the public build.
❌ No trim multiple is set for this name — deliberately
CLAUDE.md asks for trims as a valuation multiple (Trim NNx fwd). That convention is invalid here and would produce a dangerous instruction. The site computes the dollar level as multiple × (price ÷ forward P/E); with MU's forward P/E of 5.74 the implied EPS is $155.52, so Trim 8x fwd renders at $1,244 — approximately the all-time high — and Trim 10x fwd at $1,555. The mechanism would faithfully capitalize an EPS figure this entire analysis rejects, and instruct holding through the top of the cycle.
Cyclicals at peak earnings need an explicit instruction, not a multiple. Filed as pitfall-multiple-trim-inverts-on-peak-cycle-cyclicals. The instruction for MU is trim on strength / exit, not a price level.
Break triggers (falsifiable, dated)
| # | Trigger | Date | Falsifies |
|---|---|---|---|
| 1 | FQ1 FY27 guide + 4Q26 DRAM contract pricing. Bear confirmed if the FQ1 revenue guide is <$48B or GM guide <80%, or 4Q26 DRAM contract pricing is <+5% QoQ. Bull confirmed at >+15% QoQ with GM ≥84%. | FQ4 print ~2026-09-23 | The affordability-ceiling thesis |
| 2 | HBM4 allocation at Vera Rubin volume ramp. Bear confirmed if disclosed HBM share stays <15%; bull confirmed at ≥20% or a named at-scale win. Watch whether Nvidia actually relaxes HBM4 specs. | CQ4 2026 – CQ1 2027 | Whether Micron participates in AI economics at all |
| 3 | What management does when CHIPS restrictions lift. A >$20B buyback executed near current prices is management ratifying $892 with real money — and would directly contradict $233M of insider selling with zero open-market buying. Cash routed to capex instead confirms top-of-cycle deployment by revealed preference. | 2026-12-09 | The insider signal, and the capex thesis |
Secondary: CXMT wafer starts vs the ~350K/mo plan exiting 2026 · any CXMT DDR5 qualification at a Western hyperscaler · NAND contract pricing turning negative.
The single number that would most change this answer
The SCA floor gross margin. Management refused to quantify it three times on the FQ3 call. It is the largest swing factor in FY2028 earnings, it determines whether the "floor" is worth $8.90 of EPS or $20, and it is the one number that would move normalized EPS enough to change the verdict. Everything else in the bull case is already credited.
Data-quality notes
- 🚩 roic.ai reports MU's FCF as identical to operating cash flow because its
cf_cap_expendituresfield isnull— FY2025 FCF shown as $17,525M against an actual $1,665M (10.5x overstatement), FY2024 as $8,507M against $121M (70x). Derived fields inherit it:free_cash_flow_per_sh$15.70 vs a true $1.49,pr_to_free_cash_flow7.6x vs ~80x. Verified directly against the API. Filed aspitfall-roicai-fcf-field-returns-ocf. Businessquant is also wrong here (FY2023 FCF +$1.39B vs actual −$6.12B — wrong sign in the worst year). - 🚩 Yahoo reports MU's ttm
operatingMarginsat 80.37%, above itsgrossMarginsof 72.57% — arithmetically impossible. The 80.4% figure is the latest quarter's operating margin, not ttm. True ttm operating margin is 65.7%. Do not mix the two fields. - ⚠️ Micron does not disclose total HBM revenue. The only company figure is ">$1B of HBM4 revenue" cumulative. Any "HBM revenue was $X" number in circulation is an analyst estimate.
- ⚠️ The SCA floor gross margin is deliberately undisclosed — management deflected three separate analyst attempts to quantify it. It is the single largest swing factor in FY2028 earnings and it is the number we do not have.
- ⚠️ CXMT IPO proceeds are reported between $4.2B and $8.6B across sources. The $8.6B figure (RMB 57.92B, Shanghai STAR, 2026-07-27) has the better sourcing.
- ⚠️ Correction to an internal figure. The prior watchlist entry cited "Samsung/SK Hynix −$290B" alongside CXMT's funding in a supply context. The $290B is combined market capitalization lost in a single day on 2026-07-28 — an equity-value event, not committed capex. The real 2026 capex figures are Samsung ~$73.3B (KRW 110T, chip capex + R&D, +22% YoY) and SK Hynix ~$32–35B (high-KRW-40s trillion). Corrected in the watchlist by this run.
Sources
Yahoo Finance MCP (quarterly income statement, cash flow, balance sheet, recommendations, insider transactions) and .mcp/fin.py, retrieved 2026-08-04 · roic.ai get_cash_flow NASDAQ:MU, retrieved 2026-08-04 · Micron FQ3 FY26 press release · Micron FQ3 FY26 8-K (SEC) · Micron FQ3 FY26 prepared remarks · MU 10-Q FY2026 Q3 · Investing.com — $100B customer agreements · wccftech — 16 SCAs, 5-year, non-cancellable · Slashdot — Micron locks in prices for five years · TrendForce — SK Hynix removes LTA price cap · TrendForce — Nvidia may relax HBM4 specs · Nvidia certifies all three for Vera Rubin HBM4 · TrendForce — 3Q26 memory pricing · TrendForce — NAND supply to outpace demand in 2027 · Tom's Hardware — affordability limit · CNBC — CXMT STAR debut +466% · SemiAnalysis — CXMT challenges DRAM · Counterpoint — Q2 2026 DRAM share · Bloomberg — Samsung $73B 2026 chip spend · Digitimes — SK Hynix capex · Marvell custom HBM · NIST — Micron $200B US investment · Manufacturing Dive — MATCH Act · stockanalysis.com — MU cash flow
Related agency knowledge: pattern-ai-levered-fields-trade-above-own-band · principle-down-a-lot-is-not-cheap · principle-primary-source-beats-vendor · principle-size-against-both-portfolio-files · agency precedent: SNDK /analyze 2026-07-28 — AVOID at conviction 3.5 on a 78.3% NAND gross margin, the same trade one rung down the quality ladder.