Are We Past the Peak?
Micron guides a single quarter to $50 billion at an 86% gross margin; the stock has sat below its June high for thirteen weeks; and the last two times the company reported a record margin, the stock had already peaked. Wednesday’s print cannot answer the title question — it tests fiscal 2027’s first quarter — and the price already sits at the middle of the distribution the house draws for it.
Prices and valuation as of the last regular-session trade on September 25, 2026; volatility, drawdown and percentile series run through the September 24 bar. Micron reports fiscal fourth-quarter 2026 results on Wednesday, September 30, 2026, after the close (company-verified); the quarter has fourteen weeks. This note carries a frozen pre-event scorecard — Addendum 1 — locked on September 25, before the print. The post-event scorecard is published after the print as a separate follow-up note, linked from this page and from the research library.
How to read it. Weekly closing price of MU in dollars per share, January 2016 to September 2026, on a logarithmic vertical axis (equal vertical distances are equal percentage moves). The ink line is the price; the two dashed gold horizontals are the weekly closing highs of the 2018 and 2022 cycles, carried across the chart as high-water marks; gold dots mark those highs and the June 2026 high close, oxblood dots the weekly closing lows that followed, labelled with the drawdown; the two dotted verticals mark the dates the company’s own release language turned. Compare the vertical distance from each crest to its trough, and the position of the last dot against the June 2026 crest. The one thing it shows: the current price sits thirteen weeks past a crest, above every prior high-water mark by an order of magnitude.
The verdict
Are we past the peak? Of margins, not this quarter — the guide is 86% against 84.9%, though the DRAM price series behind that margin (the second exhibit below) has already stopped accelerating. Of price, the high close is thirteen weeks old — in 2018 and 2022 that was about halfway through the lead the price had over the company's language, in 2021 a larger fall came with no turn in the language at all and in 2024 the turn came only after the fall was booked, and none of those sequences had a contract programme underneath it. Of the cycle, the print cannot say; it can show the direction of DRAM prices and the first fiscal-2027 guide. The house's answer is the distribution below, and the price sits at its middle. Fair value is $1,114 per share, probability-weighted from the cash-flow model: a memory downturn with contract floors at one tail ($554), a decade without one at the other ($1,875), and a base case in which the company's "structural transformation" survives the next down-leg at half strength ($1,013). The price, $1,082, is 3% below the weighted value and 7% above the base. That is not a verdict on the operations, which are the best in Micron's history by every line — record revenue, an 84.9% gross margin, $18.3 billion of quarterly free cash flow, net cash, sixteen multi-year take-or-pay contracts. It is a statement about where the price sits: the market and the house carry the same distribution, and the only argument worth having is over the weights on the tails, which one quarter cannot settle.
Four numbers carry the whole argument — the first two belong to the company's case, the last two to the question:
What Micron is now
Micron is one of three companies that make leading-edge DRAM — the memory every computer, phone and AI accelerator reads from — and one of six that make NAND, the storage behind them. Fiscal 2025 revenue was $37.4 billion, 76% DRAM; fiscal 2026, which ended September 3, will print about $129 billion, three and a half times that, on roughly the same number of bits. The difference is price. In the June quarter DRAM bits grew low-single digits and DRAM prices rose in the low-60s percent; NAND prices rose in the mid-80s. Revenue was $41.5 billion, up 74% in a quarter, and because a memory fab's costs are almost entirely fixed, 98 cents of every incremental revenue dollar reached gross profit. Gross margin 84.9%, operating margin 81.2%. The fourth-quarter guide is $50.0 billion at about 86%, $31.00 of EPS — on a fourteen-week quarter, so the 13-week-equivalent run-rate is $46.4 billion, about $186 billion a year.
How to read it. Grouped bars, one group per fiscal quarter from the August 2025 quarter to the May 2026 quarter, each bar the sequential percentage change Micron stated on its earnings call: gold is the DRAM price per bit, ink the NAND price per bit, grey the DRAM bits shipped; the vertical axis is the change in percent. The company states ranges in words (“low-60s”, “mid-teens”); each bar stands at the midpoint of the words with a whisker for the range and the range printed above it. The two hollow dashed bars at the right are TrendForce’s industry contract-price forecasts for the calendar quarters ahead (server DRAM for July–September, PC DRAM for October–December), a different source and basis, shown for direction only. Compare the heights of the price bars across quarters and against the bit bars. The one thing it shows: revenue growth has been price, not bits, and for DRAM the rate of price increase stopped rising between the February and May quarters while NAND’s was still rising.
Three things changed in fiscal 2026 beyond price. First, high-bandwidth memory (HBM) — stacked DRAM bonded to AI accelerators — became a business whose calendar-2026 supply was sold on price and volume by December, which the company now describes as booked through 2027; HBM4 has shipped over $1 billion for its lead customer's platform and Micron's stated share target — restated on the June call, in answer to a question — is HBM share "close to our DRAM share" (about 23%, per industry trackers). Second, the strategic customer agreements: sixteen take-or-pay contracts, typically five years through calendar 2030, covering roughly 20% of DRAM and a third of NAND volume today and targeted at "half or more" of revenue, the largest with a floor through the term and a ceiling at the June-quarter market price; fourteen of them carry about $100 billion of minimum revenue, and customers have committed $22 billion of deposits, $18 billion in cash, arriving this quarter. Third, the balance sheet: $24.4 billion of net cash at May, BBB+ from all three agencies, a 30% dividend increase, and a commitment to lift capital returns from December 9 and "over time... return 100% of our excess cash."
How to read it. Five horizontal bars, each a share of 100% (the outline), from the company’s June 24, 2026 prepared remarks and the fiscal Q3 10-Q: the DRAM and NAND volume under strategic customer agreements today, the share of revenue the company targets for coverage when all planned agreements are signed, the share of revenue at fixed or ceiling prices at completion, and (lighter) the roughly $23 billion a year of minimum contracted revenue as a share of fiscal 2026 revenue. Compare lengths only. The one thing it shows: at completion, half of revenue is targeted for floors and two-fifths sits at fixed or ceiling prices; today, minimum contracted revenue is about a fifth of fiscal 2026’s.
The cumulative revenue at minimum contract prices across fourteen of the sixteen strategic customer agreements, with floors the company says sit “well above our peak quarterly margins in any past cycle.” The company’s strongest number, and one no memory cycle has carried before.
Two arithmetic facts about the contracts deserve plain statement, because both sides of the argument quote them. The $100 billion is a floor, not a forecast — over the roughly 4.3 years left in the terms it is about $23 billion a year, 18% of fiscal 2026 revenue, and the company says revenue "will well exceed" it. And the ceiling cuts the other way: about 40% of revenue at completion is contracted at or near today's prices, so that share cannot participate in a 2027 price rise. A floor on half, a ceiling on 40%: the contracts make the business less cyclical in both directions on the covered share, and leave the other half exactly as cyclical as it has always been.
The supply plan is the other half of the file. Fiscal 2026 capital spending is about $27 billion (raised three times from an $18 billion estimate: $20 billion in December, above $25 billion in March, about $27 billion in June); fiscal 2027 quarterly capex runs "above" the fourth quarter's $10 billion, more than half of the increase construction, with a figure in the mid-$40 billions given on the call. The first new wafer capacity arrives mid-2027 (Idaho's first fab and the acquired Taiwan site), the second Idaho fab late 2028; the full build list is in the next section. Management's own forecast is that "the blended DRAM cost per bit" will rise from here — new products cost more to make and greenfield fabs carry startup costs. That is not a problem while prices rise faster. It is the mechanism by which margins compress when they stop.
How it got here: forty-eight years, other people's fabs, one trillion dollars
Micron was founded on October 5, 1978, as a four-person semiconductor design company in the basement of a Boise dental office; it broke ground on its first fab in 1980, shipped its first 64K DRAM in 1981 into the Commodore 64, and listed on Nasdaq in 1984 (the company's own timeline). It came through the mid-1980s DRAM price collapse that drove most American producers, Intel among them, out of the business, and for the next three decades it was the survivor that bought capacity at the bottom of other people's cycles: Texas Instruments' memory operations in 1998; Toshiba's Manassas, Virginia fab in 2002; a NAND joint venture with Intel in 2005 and Lexar in 2006; a DRAM joint venture with Nanya, Inotera, in 2008; Numonyx from Intel and STMicroelectronics in 2010; Elpida, Japan's last DRAM maker, out of bankruptcy on July 31, 2013, for ¥60 billion (about $612 million) in cash at closing plus its stake in Rexchip — the Hiroshima fab now being converted to EUV; the rest of Inotera by share swap in December 2016; Intel's half of the NAND venture for about $1.25 billion in October 2019 (8-Ks). It has also walked away: 3D XPoint, announced with Intel in 2015 as "the first new memory category in decades," was abandoned on March 16, 2021, and the Lehi fab sold to Texas Instruments; the Crucial consumer brand was retired at the start of this year. Four chief executives in forty-eight years — Joe Parkinson to 1994, Steve Appleton until his death in February 2012, Mark Durcan, and since 2017 Sanjay Mehrotra, a co-founder of SanDisk.
The industry consolidated around it. Dozens of DRAM producers in the 1990s became three at the leading edge — Samsung, SK hynix and Micron — with the state-backed CXMT the entrant the 10-K names. Consolidation did not end the cycle; it made each turn larger for the survivors. In nine fiscal years the company has reported three record margins and two troughs: fiscal 2018 ($30.4 billion of revenue, a 61% gross margin in the peak quarter, $11.51 of EPS) to fiscal 2020 ($2.37); fiscal 2022 ($30.8 billion, 47% in the peak quarter, $7.75) to fiscal 2023 (−$5.34); and now fiscal 2026 (about $129 billion, 86% guided for the peak quarter, about $73). Across fiscal 2017 to 2025 the company earned $232 billion of revenue and kept about $17 billion of it as free cash flow — 7% — because each up-cycle was reinvested into the next. The stock went from $14 at the end of 2015 to $1,082.
The history matters for one reason the valuation section returns to. Every prior expansion was bought cheaply at a trough — Elpida's fab, Inotera's, Intel's share. This one is built from the ground up at the peak: a $15 billion Idaho fab announced in 2022 and a second beside it, a New York cluster announced at up to $100 billion over twenty years, an acquired site in Taiwan, HBM packaging in Singapore, a NAND fab there, EUV in Hiroshima, assembly in India — $27 billion this fiscal year and more next. Micron has bought its way through troughs before and spent record capex at every peak (fiscal 2018, 2022 and now 2026); what it has never done is build greenfield at this scale — on decisions taken as the 2022 down-cycle began, a year before its trough and long before this up-cycle, which is the company's pattern rather than a break from it. What is new is who helps pay: $6.4 billion of CHIPS Act grants and $18 billion of customer deposits cover about a third of two years' capex — either the structural break the company describes, capacity pre-sold under take-or-pay before it is built, or the largest capacity commitment in Micron's history, made at a record margin.
Where memory sits in the build-out
The house's map of the AI build-out (published in July) reads it as a profit migration down the stack governed by two forces: a demand force — whether roughly $700 billion a year of capex earns its cost of capital — and a supply force, how long each input takes to add: about two years for silicon against five or more for power in the map's figures (the house's power note puts the range at three to seven-plus years). On the demand force the house's payback clock found the four largest hyperscalers guiding $700–725 billion of 2026 capex (from about $420 billion in 2025 and $230 billion in 2024), a 2026 vintage that needs roughly $221 billion of annual operating profit to clear its hurdle, and creditable AI operating profit of about $84 billion — a shortfall near $137 billion a year that the house's September rates note re-derived at about $143 billion on a 9.4% cost of capital after the Fed's hike, and that closes only in the family of futures where capex freezes at 2026 levels while AI revenue compounds.
Memory is where the supply force landed in 2026. Three suppliers, a two-year add time for wafers, and an HBM generation that consumes more conventional-DRAM wafers per bit sold every year turned a demand shock in accelerators into a price shock in every memory market at once: conventional DRAM contract prices roughly doubled in the March quarter (TrendForce), the DRAM industry's June-quarter revenue was $154.7 billion — annualizing near $620 billion, close to the four hyperscalers' entire 2026 capex budget, which is another way of saying the memory bill is now paid by everyone who buys a server, a phone or a PC, not only by the AI builders. Micron's data-centre revenue passed $25 billion in the June quarter, "an annualized run rate of over $100 billion"; data-centre bits exceed half of industry demand for the first time in calendar 2026 (company); the HBM market the company sized at about $35 billion in 2025 is forecast at $100 billion by 2028 — "larger than the size of the entire DRAM market in calendar 2024" — and on the June call that milestone was pulled to 2027. Third-party teardowns put HBM at roughly 45–50% of an accelerator's bill of materials (secondary). The binding silicon input in 2026 is not the GPU; it is the memory bonded to it — power, by the house's map, is scarcer still and slower to add.
Three consequences follow for the valuation. First, by the house's own map memory is the most cyclical rent in the stack — a shortage rent with a roughly two-year half-life, which is exactly the interval to Micron's greenfield dates (Idaho and Taiwan mid-2027, the industry's additions through 2028); the base case in this note is that map applied to one company; the company's reply is that the additions are already spoken for — HBM booked through 2027 and the contracts covering the volume as it arrives — which is the upside case. Second, the demand-force risk lands on memory first: in the one family of futures where the hyperscalers' payback gap closes, capex freezes, data-centre bit growth slows from doubling-in-two-years to something three suppliers can meet, and the spot half of Micron's revenue reprices — that is the mechanism inside the downside scenario. Third, concentration flows downstream: the house's NVIDIA note found that company's top two direct customers at 36% of revenue; Micron takes half of its revenue from ten customers, four "very large" ones anchor the contracts, and the $18 billion of deposits is those customers paying to hold their place in an allocation. The contrast with the accelerator layer is instructive. The house's NVIDIA valuation found the entire gap to that price was the discount rate and the terminal multiple, not growth. Micron's gap is the opposite shape: the discount rate is contested but bounded, and the disputed input is an operating one, the recovered margin — in the one layer of the stack where the market already applies a cycle discount. On trailing earnings Micron and NVIDIA trade at 24.5x and 28.4x; on next year's consensus Micron is 6.8x and its Korean peers mid-single digits, because the market expects the earnings to more than double and then discounts them as a peak.
What the stock did
MU closed 2026-09-24 at $1,080.53 and last traded at $1,081.69. Twelve months ago it was $155; the high close, $1,213.56, printed on June 25, the day after the June quarter's results (+15.7%). The stock is 11% below it. Realized volatility is 54% over thirty days and 95% over ninety; the largest drawdown inside the last fifteen months was 39%. Four sessions into September 22 added 18%; on the 23rd the stock fell 2.2% after Michael Burry — who opened a short in July, saying the AI memory boom "might differ from previous cycles, but not nearly different enough" — said he had added to it "in some size," citing an Acer remark that DDR4 has "more sellers than buyers," expanding Chinese supply, and a report of CXMT DDR5 yields above 90%. The same day Wells Fargo cut its target to $1,400 and Citigroup raised to $1,300. Over ninety days the target tally is three raises and three cuts; the revision cycle is flat. Fifty-four analysts rate the stock a buy, two hold, one sell; the mean target is $1,516–1,576, 40–46% above the price. On our reading that target is the upside path, not the base: the Street's fiscal 2027 — $248 billion of revenue, $159 of EPS — is within rounding the first year of our upside case ($244 billion, $148), and its "6.8x forward" multiple is a multiple of that path.
The rest of the complex tells the same story at different volumes. SanDisk, Seagate and Western Digital, which ran with DRAM, sit 20–43% below their June highs. SK hynix and Samsung both reported record June quarters (76% and 70% operating margins on memory) and both guided undersupply through year-end. TrendForce's contract-price series is the cleanest description of where the cycle stands: conventional DRAM contract prices rose 93–98% in the March quarter and 58–63% in the June quarter; server DRAM is forecast up 13–18% this quarter, PC DRAM 3–8% next. The rate of increase is falling; the level is not. Micron's own guide says the same thing in its own words: the 86% margin "reflects a meaningful moderation in the rate of price increases."
The last two times
How to read it. Three horizontal timelines, one per cycle, in weeks from the week of the stock’s peak weekly close (gold dot at zero). The ink diamond marks the release in which the cycle’s record gross margin was reported; the oxblood cross marks the release whose language first described weakening conditions, with the price drawdown from the peak to that date; the hollow circle on the 2026 row is today. Compare the lengths of the gold bars (weeks of confident language after the price peak) and the drawdown labels. The one thing it shows: in both prior cycles the price turned first and the language last, and the 2026 row is thirteen weeks along with no turn in the language.
Micron has reported a record gross margin three times in nine years. The first two are dated in the company's own filings and in the tape, and the sequence is the same both times.
In 2018 the stock's high close was $61.39 in the week of June 4. Two weeks later the company reported a record quarter and "ongoing momentum and healthy industry fundamentals"; in September it reported the cycle's peak gross margin, 61.0%, and called its position "a tremendous catalyst... in 2019 and the years ahead." On December 18 the language turned — "weak near-term industry supply-demand dynamics entering calendar 2019" — with a guide down 25% in a quarter. By then the stock was $30.32, down 51%. Gross margin bottomed at 26.6% a year later; fiscal EPS went from $11.51 to $2.37.
In 2022 the high close was $97.36 in the week of January 10. The company reported records in March. On June 30 it reported another record quarter and, in the same release, "recently, the industry demand environment has weakened," with a guide down 17%; on August 9 it said the quarter would land at or below the low end. The stock was $53.65 that day, down 45%. Gross margin went from 47.2% to −32.7% within a year; fiscal EPS from $7.75 to −$5.34.
The lead of the price over the language was 28 weeks and 24 weeks. Today's high close is thirteen weeks old; the record margin has been reported and a higher one guided; the language — "tight conditions... beyond calendar 2027," "no line of sight" to supply catching demand — is at its most confident. The sequence conditions on a record-margin quarter; unconditioned, the stock fell 29% from an April 2021 high with no turn in the language at all, and 39% from a June 2024 high (54% by March 2025) before the language softened in December 2024 — "consumer-oriented markets are weaker in the near term," with a 9% guide-down — 27 weeks after the peak and with no margin cycle behind it. So the clock is a sequence to watch, not a probability: nothing here says the cycle has turned, only that the sequence happened twice — and that what is different this time is in the contracts, not in the confidence.
The strongest case for the price (steelman — stated in its own terms, before we test it)
The case for owning Micron at $1,082 is that the memory cycle, as the file records it, has ended, for reasons that are physical rather than rhetorical. Supply is constrained by things that do not respond to price in a year: greenfield fabs take three to four years, skilled construction labor and permitting are binding constraints, each EUV node delivers less bit growth per wafer than the last, and every HBM generation consumes more conventional-DRAM wafers per bit sold. Management says it has "no line of sight" to supply catching demand and has moved that horizon out a year every quarter. Demand is structural: data-center DRAM and NAND bits doubling in two years, agentic AI adding CPU racks and context stores to the accelerator racks, on-device AI lifting content in phones and PCs, robotaxis and humanoids carrying five to ten times a car's memory. And the business model has changed in a way the tape has not priced: take-or-pay contracts with floors "well above any past peak margin," $100 billion of minimum revenue, $18 billion of customer cash on deposit, half of revenue targeted for coverage — customers "recognizing that their product roadmaps rely on... committed long-term memory supply." On this view a company earning $125 of EPS at the current run-rate, net cash, returning all of its excess cash from December, at 8.7x that run-rate and 6.8x next year's consensus against an index at 19x, is not expensive; it is a cyclical being priced as a cyclical after the cycle has been contracted away. The highest sell-side targets, $2,000–2,200, are what a decade of that looks like.
The test. We enter that case the only way a cash-flow model can: as a scenario in which every operating claim in the June prepared remarks comes true and persists — tightness with no supply intercept through the decade, gross margins in the 80s through fiscal 2029 and 78% at the terminal, revenue rising past $400 billion, capex absorbed. That construction is worth $2,639 per share, of which 64% is terminal value, and it is what the top of the sell-side range requires. It is not probability-weighted, because its premise is that a memory cycle has ended, and the file contains two that did not. Which of these claims the note tests: the supply constraint partially, through the supply-arrival dates in the forecast; the demand claim partially, through bit growth; the contracts directly, through the floor-and-ceiling arithmetic and the recovered margin; the balance sheet fully, in net cash; the valuation claim directly, in the reverse DCF. Which it cannot test within its frame: whether the demand curve for memory has permanently steepened, and whether the floors were struck at levels that survive a customer's own downturn. Between it and the price sits the question the note is about: how much of the volume is on floors when the spot price falls — and that is a fact the print cannot disclose, because the floor prices are not disclosed.
What $1,082 has to believe
Reverse-engineer the price. The market capitalization is $1,222 billion; less $46 billion of net cash at the fiscal year-end (deposits excluded — they are a liability), the operating value in the price is $1,176 billion. At the house cost of equity of 11.6% — a 5.18% ten-year Treasury after the Federal Reserve's September increase, a five-year weekly beta of 1.65 adjusted to 1.43, a 4.5% equity risk premium — that operating value is:
$90 billion a year of free cash flow to equity, sustained forever and growing 3.5%.
At 10.6% it is $80 billion; at 12.6%, $101 billion. Fiscal 2026 will produce about $48 billion; our base case's fiscal 2027 — the peak year — produces $99 billion.
We use the five-year beta window because a ten-year discount rate should not be set by the fifteen months in which the stock rose sevenfold; the two-year and daily windows put the cost of equity at about 14% to 16%, and at 14% the weighted value is about $881 — the skeptic's discount rate is printed rather than settled.
The price requires the peak year's cash flow to be the permanent level. Put the same requirement in the model's own language — hold the base revenue path through the fiscal 2028–2030 down-leg and solve for the recovered mid-cycle gross margin that reproduces the market cap — and it is 63.5% (55% at 10.6%, 73% at 12.6%). The prior-cycle peak was 61%. Our base case assumes 60%; our upside case 70%. The price sits just above the base and well below the upside case, which is what a 50/25/25 weighting says. Read this way the market is neither deluded nor smarter than the model: it is carrying the same distribution.
Valuation
How to read it. A vertical gauge of value per share in dollars, $300 at the bottom to $2,900 at the top, linear. Ink ticks with labels are the three probability-weighted scenarios and, in grey, the unweighted steelman-maximum; the gold triangle is the probability-weighted fair value; the oxblood triangle is the last regular-session trade on September 25, 2026; the shaded gold band is the pre-committed range inside which any single-print outcome may move the fair value. Compare positions on the gauge only. The one thing it shows: the price sits between the base case and the weighted value, inside the frozen band.
We value the equity with an explicit-path FCFE model (free cash flow to equity, discounted at the cost of equity; the balance sheet is net cash and the dividend is 1% of earnings) through a full memory cycle: each scenario carries its own revenue, gross-margin and capex path through a down-leg and a recovery, and the terminal value is struck on the recovered year, after growth has landed on its terminal rate. A growth-fade from fiscal 2026 is rejected on principle — that base is a cyclical peak, and a fade from it would bake the peak into the terminal. Stock compensation is expensed (it is 1.2% of revenue), the share count is frozen at the company's 1.15 billion guide, one cost of equity applies across scenarios, terminal growth is 3–4%, and the fiscal 2027 base year starts from the fourteen-week quarter's 13-week-equivalent run-rate. Fiscal 2026 is $129 billion of revenue and $73 of EPS; the exit run-rate is $186 billion.
| Scenario (weight) | The story | FY27 revenue / EPS | Trough margin | Mid-cycle margin | Per share |
|---|---|---|---|---|---|
| Downside (25%) | The cycle is a cycle: price increases stop in fiscal 2027, Chinese and greenfield supply meet a digestion, spot prices fall a third across FY28–29; floors hold on ~40% of revenue, so the trough is a margin, not a loss | $194B / $104 | 35% (FY29) | 46% | $554 |
| Base (50%) | The floor holds and spot normalizes: FY27 grows on the exit run-rate plus bit growth with flat prices; greenfield supply turns spot pricing down through FY28–30 while the contracted half sits on floors; recovered margin 60% — above every prior peak, the structural claim at half strength | $219B / $124 | 54% (FY30) | 60% | $1,013 |
| Upside (25%) | Supercycle: tight through 2029, HBM at the DRAM share, prices rise again on the spot half in FY27, a shallow late-decade correction from a much larger base | $244B / $148 | 68% (FY31) | 70% | $1,875 |
| Probability-weighted | ≈ $1,114 | ||||
| Maximum defensible (unweighted) | The steelman case above, in full | $251B / — | none | 78% | $2,639 |
The weights are a judgment, and here is the argument for them: half on the base because the two facts that matter — the contracts exist, and every prior cycle also existed — offset rather than cancel; a quarter on each tail because the file holds two cycles that ended and one contract programme that has not yet been through one, and no reading of that record puts either tail above the middle. The tails are equal because the one claim that separates them — that the floors pay in a customer's downturn — is untestable until one occurs, and a weight is not a vote count.
Sensitivities, stated: each point of cost of equity is worth −$115 / +$149 on the weighted value; each 5 points of recovered mid-cycle margin about ±$68 on the base value; the base case's three-year down-leg costs about $109 a share against a path that holds margins in the 60s. Terminal value is 52–56% of operating value in every scenario, and the assumption carrying it is the recovered margin. The relative lens cannot adjudicate this question: every DRAM maker trades at a single-digit forward multiple (SK hynix and Samsung both in the mid-single digits on the aggregators, Micron 6.8x on consensus) because the market applies the same cycle discount to all three at once, so a comps table inherits the sector's judgment rather than testing it; Micron's one-turn premium to its Korean peers is consistent with a U.S. listing, a U.S. footprint customers pay for, and the only disclosed contract program. The defensible range across reasonable weightings is $850–1,450; the point estimate is $1,114, which is 15x fiscal 2026 EPS, 9x the base fiscal 2027, and 10.7x the base mid-cycle earnings.
Risks — in both directions
The case against the price, in its own terms: every memory cycle has ended with supply arriving after the price signal; the contracts guarantee a fifth of revenue and cap two-fifths of it at today's prices; cost per bit is rising by the company's own forecast; receivables are three-quarters of a quarter's revenue and $18 billion of deposits is, among other things, a form of buying ahead; and in both prior cycles the price turned first. The note tests the cycle repeating (the downside scenario, floors holding on about 40% of revenue, $554); it cannot test the floors failing, but it can price it — the same revenue path with a 20% trough gross margin on all of revenue — between the 2019 trough of 27% and the 2023 one, which was negative — is worth about $500, and a literal 2023 replay would be lower.
The risk we rank first is the one the base case already contains: the price-increase cycle ending in fiscal 2027 — TrendForce's deceleration, the company's own "meaningful moderation," PC and mobile softening (Samsung's phrase). What separates base from downside is the supply side — earlier or larger arrivals from CXMT (its DDR5 yields are a secondary-sourced claim), from Korean capex, or from Micron's own $45 billion — and the demand side's oldest mechanism, customer digestion, which the receivables line would show first. The contract floors are a legal right, not a cash certainty, and cannot be tested until a customer has a downturn of its own. Micron is third of three in HBM, with secondary reports putting its share of the lead accelerator platform's HBM4 in the teens; management says the ramp is twice as fast as the last generation and the target is the DRAM share — we carry the allocation as a range and freeze a gauge on whether a number is disclosed. China is both a lost customer set (the 2023 CAC decision) and the marginal supplier. The discount rate — beta between 1.5 and 3.0 depending on the window — is the largest single sensitivity and cannot be settled by a print.
The other direction is as concrete. If contract coverage reaches half of revenue with floor prices disclosed, if HBM share reaches the DRAM share, if the March-2027 gross margin is still above 82% with contract prices rising, the weights move toward the upside case and the value toward $1,875.
Two governance data, neither a valuation input. Insiders sold $387 million of stock in twelve months — the CEO $210 million on sixteen sale dates across eleven filings, at prices from $180 to $1,192: a cluster of eleven dates in the five weeks to early November 2025, then one a month from May to August 2026, every filing checked ticks the Rule 10b5-1 box and cites plans adopted in August 2024 and January 2026 — with one director's $7.8 million purchase at $337 in January the only buy; consistent with diversification at any price level. Short interest is 2.45% of shares at the 15th percentile of its own two-year range and down a third since June: no crowded short to squeeze, no fresh short pressing — borrow fee and utilization are unobserved — so on this data whatever moves the tape Wednesday is long-only money changing its mind.
Wednesday night: what the print decides — and what it can't
The September 30 report is the last quarter before the fiscal 2027 base, the first with contract deposits on the balance sheet in size, the first fourteen-week quarter of the cycle, and — by the company's own calendar — the quarter in which the rate of price increase moderates. The quarter itself is close to pre-announced: $50.0 billion ± $1.0 billion, about 86%, $31.00 ± $1.00 (consensus $31.59 on $51.2 billion), and a prediction market prices a beat of $32.22 at 96.6% (a thin contract, graded as color). The information is in the first-quarter guide, which is a thirteen-week quarter: a guide below $46.4 billion is a decline in the weekly rate, whatever the headline says.
We froze our scorecard on September 25, before the print (Addendum 1 — thresholds immutable; each gauge carries a reading rule so the post-event note can score it in a sentence). The eight questions, stacked:
Revenue at or above $50.0 billion and a 13-week equivalent (revenue × 13/14) at or above $46.4 billion; below $45.5 billion is the downside tell. Forty-two percent of the guided sequential growth is calendar.
A midpoint at or above $46.4 billion clears; $50 billion is the upside tell; below $44 billion the downside. No revenue guide is not-given — and a posture change from a company that has guided every quarter for a decade.
A first-quarter gross-margin guide at or above 84% clears; below 80% is the downside tell.
The prepared remarks always state the direction; up clears, "flat" or "down" is the downside tell, +20% or more the upside.
A guide at or under $48 billion clears (the base uses $45 billion); above $50 billion is the downside tell. No number is not-given; the "above fourth-quarter levels" language stands.
Any quantified HBM metric — revenue, run-rate, share, bits — clears. Adjectives only is not-given by construction: this gauge grades disclosure posture, so silence is the answer, and the third-party allocation stays a range.
A count above sixteen, minimum revenue above $100 billion, deposits received of $10 billion or more, or a floor price disclosed, clears. No update at a fiscal year-end is not-given — and worth noting from a company that calls the program transformational.
A context gauge for a house call that resolves October 5; not mentioned is not-given and moves nothing.
What the print cannot settle — frozen in advance so no outcome gets over-read: the discount rate; the floor prices in the contracts; when greenfield and Chinese supply arrive; whether the demand curve for memory has permanently steepened; whether a take-or-pay contract pays in a customer's downturn. Accordingly we pre-committed a bound at the freeze: no single-print outcome moves our central fair value outside $959–1,332 — the re-weighting bounds on the frozen scenario values (downside to 40%, or upside to 45%). Scenario values themselves move only from an archived rerun of the model, and only if the release changes a model input.
The market's own posture into the print, briefly: options price ±8.2% for the event against a realized median of ±9.1% over the last eight prints — the first print in this run where the market pays less than the stock's own history; the stock closed down after five of those eight, including after a 31% earnings beat in March. The 25-delta skew is fractionally to the calls. (Our frozen file also contains a fitted probability density from the option chain; it extracts to a standard deviation a third of what the implied volatility says, so we set it aside as unreliable rather than lean on it.) Peers printed record margins and guided undersupply, so a soft Micron quarter would be company-specific. One asymmetry the reader should hold: our guide gauge clears at $46.4 billion (a flat weekly rate), while the Street's number for the same quarter is about $57.3 billion of revenue and $35 of EPS (31 and 32 analysts) — a guide near $50 billion would clear the scorecard and miss consensus by about an eighth, which is the beat-that-disappoints setup. These are reads, not positions; each has a grader and settles on the ledger.
How to read it. Bars are the absolute percentage change of the stock in the session after each of the last eight earnings releases, dated by release month (jade: the stock closed up; oxblood: it closed down), in percent on the vertical axis. The dashed gold line is the event-specific move the options market priced on September 25 for the September 30 print; the dotted grey line is the median of the eight bars. Compare bar heights to the two lines. The one thing it shows: the market is paying slightly less than the stock’s own recent history for this print, and five of the eight bars are oxblood.
What would change our view (reversal)
We move toward the upside case — and would raise the fair value toward $1,875 — on any of: a fiscal 2027 gross margin at or above 82% through the March 2027 quarter with DRAM contract prices still rising; HBM revenue or share disclosed at or above the DRAM share; contract coverage reaching half of revenue with floor prices disclosed as numbers; or a durable ~150bp fall in the ten-year Treasury (in which case our discount rate, not the company, was the error — stated symmetrically). We harden toward the downside on: a quarter-on-quarter DRAM price decline before the September 2027 quarter; receivables above 80% of quarterly revenue or inventory days above 140; a customer publicly disputing or renegotiating a contract; CXMT shipping DDR5 into a top-ten Micron customer at volume; or fiscal 2027 capex guided above $50 billion without a matching revenue path.
Recommendation
Fair value $1,114; the price is the base case; no directional call into the print. For a reader who must act: if you believe the contracts and the supply constraint have ended the memory cycle, the stock is worth $1,875–2,639 and the sell-side targets are right; if you believe every cycle ends the same way, it is worth $554 with floors and about $500 without them, and the 2018–2022 sequence has thirteen weeks on the clock; if, like the house, you weight the middle, the stock is fairly priced and Wednesday is information about fiscal 2027's first quarter, not about the answer. At 54% realized volatility and a fair-value distribution centered on the price, a directional position here is a bet on the weights, and is sized accordingly under house rules — which is to say, small or not at all. The house's interest is on the other side of the print: the first-quarter guide on a thirteen-week base, the direction of DRAM prices, and whether any of the three things this note cannot see — the floor price, the HBM share, the supply date — becomes visible.
Sources & Method
All figures trace to the Phase 0 evidence manifest (data/MU/2026-09-25-120000/manifest.json; 26 feeds, snapshot September 25, 2026) and to archived computation scripts; no number is restated from memory. Primary sources: the FY2025 10-K (filed 2025-10-03; revenue by technology, customer concentration, competition, CHIPS funding); FY2026 10-Qs (December 2025, March 2026, June 2026 — the fourteen-week fourth quarter, SCA deposits and remaining performance obligations, debt, the balance sheet); fourteen 8-Ks June 2025–August 2026 including the five earnings releases with their guidance blocks, the March–April 2026 note tender, and the August 26 president appointments; the DEF 14A of November 2025 (compensation design, ownership); 73 Form 4s (insider trades, per-trade prices); Micron's prepared remarks for the FQ4-25 through FQ3-26 calls (Investor Relations CDN, archived); the FQ3-26 call Q&A from three independent transcriptions, used only where two agree and marked call-only in the draft (the fiscal 2027 capex figure, the HBM share target 'close to our DRAM share', HBM TAM timing, startup costs); the 2018 and 2022 cycle-turn releases (8-K EX-99.1, June–December 2018; March 2022–June 2023). Market data: Robinhood (quotes, fundamentals, verified earnings date, analyst ratings, daily bars June 2025–September 2026 and weekly bars 2016–2025 for MU and SPY), FINRA short interest via Massive, FRED (rates), yfinance (estimates, targets, revisions), Schwab option chain (implied move — frozen in research/mu-q4-fy26/data/rubric.json), Polymarket (one contract kept of 225 scanned), SEC XBRL company facts for MU, SanDisk, Seagate and Western Digital. Secondary, labeled where used: TrendForce and Counterpoint contract-price and share releases; SK hynix and Samsung newsroom results; sell-side and short-seller statements as reported by wires. Model: explicit-path FCFE through a full cycle on the 52-week-equivalent fiscal 2026 base ($125.4 billion), mid-year discounting, SBC expensed once, constant 11.62% cost of equity across scenarios (beta 1.65 weekly 2021–25, Blume-adjusted 1.43; the four beta windows and a ke grid are printed in the draft), explicit paths landing on terminal growth ≤4%, terminal capex by the steady-state identity; the house DCF gate reports WARN (two exit-multiple cross-checks imply perpetual growth above 4% and are carried as written-down assumptions; no FAIL). Known gaps, stated: the fourth quarter's free cash flow and working-capital build are author estimates (±$10 billion of net cash is ±$9 a share); fiscal 2027 capex rests on call transcriptions; the HBM4 allocation is secondary-sourced and carried as a range; the fitted risk-neutral density is set aside. Scenario weights, the 4.5% equity risk premium and the Blume adjustment are house judgment inputs, labeled as such; D&A is modelled at 16% of net PP&E against 18% in fiscal 2025 ($8.35 billion of D&A on $46.6 billion of year-end net PP&E; conservative in the explicit years, neutral at the terminal); buybacks are treated as value-neutral at fair value, so the share count is frozen; one cost of equity applies across scenarios because the risk is carried in the weights. The exhaustive working draft this note distills is archived alongside (MU_exhaustive_draft.md). Standard disclosure block attaches at publication.
The scorecard, locked before the print
Frozen 2026-09-25T18:06:55 · 8 gauges · pair PENDING
This addendum is the commitment half of the event pair: the scoring rubric, the market reads at freeze, and the bound on how far the fair value may move on any single-print outcome — all fixed before the evidence exists. The post-event note scores it afterward. Fair value at freeze: $1,114 probability-weighted (downside $554.30 / base $1,013.12 / upside $1,875.45). Spot at freeze: $1,081.69, the last regular-session trade on September 25 (the Schwab quote at the freeze read $1,085.02 after hours; the official September 24 close was $1,080.53 — all three documented). Pre-committed move cap: no single-print outcome moves the central fair value outside $959–$1,332: downside-to-40% = 0.40×554.30 + 0.45×1,013.12 + 0.15×1,875.45 = $958.94; upside-to-45% = 0.45×1,875.45 + 0.40×1,013.12 + 0.15×554.30 = $1,332.35. Scenario values move only from an archived model rerun, and only if the release changes a model input.
The frozen gauge table (every gauge has a silence branch and a reading rule; the post scores all eight, none added, none dropped)
| # | Gauge | Cleared if | Silence branch | Reading rule |
|---|---|---|---|---|
| 1 | mq4-weekly-rateFQ4 revenue >= $50.0B AND 13-week-equivalent (x13/14) >= $46.4B; bear tell: 13-week-equivalent < $45.5B | revenue >= 50.0e9 and revenue*13/14 >= 46.4e9 | Revenue is always reported; if the release does not state the week count, the 10-Q's 'fourteen weeks' governs the basis — never not-given | basis = 13-week equivalent; tolerance +/-1%; base FY27 growth re-anchors to the 13-week exit run-rate x 1.18 |
| 2 | mq4-fq1-guideFQ1 FY27 revenue guide midpoint >= $46.4B (flat weekly rate on a 13-week quarter); bull tell >= $50B; bear tell < $44B | FQ1-27 revenue guide midpoint >= 46.4e9 | No revenue guide -> not-given AND a posture change (the company has guided every quarter for a decade), flagged in the post | basis = guide midpoint vs FQ4 13-week equivalent; the base 1H FY27 price assumption is set from the guide's implied price change |
| 3 | mq4-gm-pathFQ1 FY27 non-GAAP gross-margin guide >= 84%; bear tell < 80% | FQ1-27 non-GAAP gross margin guide >= 0.84 | No GM guide -> read the EPS guide's implied margin; neither -> not-given | basis = non-GAAP GM guide; tolerance +/-1pt; base FY27 GM moves 1:1 with the FQ1 guide minus 4 points of in-year fade |
| 4 | mq4-dram-priceDRAM ASP stated UP quarter on quarter in the prepared remarks; bull tell >= +20%; bear tell 'flat' or 'down' | DRAM ASP change q/q > 0 | The prepared remarks always state the direction; a percentage range absent -> grade direction only | basis = company-stated DRAM ASP change; the FY27 price path is re-set to the stated momentum |
| 5 | mq4-fy27-capexFY27 capex guide <= $48B (base $45B); bear tell > $50B | FY27 capex guide <= 48e9 | No FY27 capex number -> not-given; the 'above FQ4 levels' language stands and the post uses $45B | basis = net of incentives, company definition; tolerance +/-$3B; each $5B above $48B = -$4/share of PV; construction share recorded |
| 6 | mq4-hbm-numberA quantified HBM metric for the quarter is disclosed (revenue, run-rate, share or bits) | any quantified HBM metric disclosed (binary) | Adjectives only -> not-given by construction; the gauge grades disclosure posture, so silence IS the negative branch and the Vera Rubin allocation stays a secondary-sourced range | binary disclosure; if disclosed, the HBM share input moves to the stated figure |
| 7 | mq4-sca-coverageSCA programme advanced: count > 16, or RPO > $100B, or deposits received >= $10B in the quarter, or a floor price disclosed | SCA count > 16 or RPO > 100e9 or deposits received >= 10e9 or floor price disclosed | No SCA update at a fiscal year-end -> not-given, and the post notes that a programme described as transformational went unmentioned | company figures; base mid-cycle margin unchanged unless a floor price is disclosed, then re-derived from it |
| 8 | mq4-cpu-serverManagement attributes server-DRAM or LP-server demand growth to CPU / agentic servers WITH a number (units, content or revenue) | CPU/agentic server attribution with a numeric figure (binary) | Not mentioned -> not-given, and the house CPU-vs-accelerator call is unaffected by this print | attributed statement with a number; no model consequence — a context gauge |
Module reads at freeze (each has a grader; probabilistic reads settle as ledger rows, never “missed”)
How to read it. Bars are the absolute percentage change of the stock in the session after each of the last eight earnings releases, dated by release month (jade: the stock closed up; oxblood: it closed down), in percent on the vertical axis. The dashed gold line is the event-specific move the options market priced on September 25 for the September 30 print; the dotted grey line is the median of the eight bars. Compare bar heights to the two lines. The one thing it shows: the market is paying slightly less than the stock’s own recent history for this print, and five of the eight bars are oxblood.
Options. ATM straddle $93.38 on the $1,082.50 strike (October 2 expiry), spot $1,082.28: event-specific implied move ±8.2% by the forward-variance method; ATM IV 77.8%; 25-delta skew −1.4 (calls fractionally bid). Realized median ±9.1% over eight prints, max ±16.2% over the last twelve; up after three of eight (five of twelve). The module’s implied/realized diagnostic (0.90) is a ledger entry, not a claim. Grader: realized reaction vs the frozen numbers.
Risk-neutral density. The fitted density extracts to a 3.3% standard deviation against the ~10% a 78% implied volatility implies over six days; set aside as unreliable. Grader: the realized close’s percentile, recorded for calibration and labelled unreliable.
Prediction markets. One contract kept of 225 scanned (Polymarket): “Will Micron beat quarterly earnings?”, threshold $32.22 non-GAAP EPS, YES 96.6%, $5.1K volume (thin; color). Grader: before → after probability and resolution.
Short interest. 27.6M shares (September 15) = 2.45% of cover-page shares, days to cover 1.1, 15th percentile of two years, down 1.24 points since June; no convertibles, single class. Read: no crowded short, no fresh short pressing. Grader: positioning outcome at the next settlement.
Analyst revisions. One target raise and one cut in 30 days (three and three in 90); FQ4 EPS $31.59 vs $31.33 thirty days ago, 0 up / 2 down; 54 buy / 2 hold / 1 sell; mean target $1,516–1,576, 40–46% above spot. Grader: actual vs consensus-at-freeze; post-print target drift.
Insiders. 11 Form 4s in the 90 days before the print; $387.1M of open-market sales in twelve months (CEO $209.8M on sixteen sale dates across eleven filings, under Rule 10b5-1 plans adopted August 2024 and January 2026); one director purchase ($7.8M, January). Grader: Form 4s in the 30 days after.
Filings. No shelf, no prospectus supplements in 400 days, no ATM, no lock-up mechanics; recent 8-Ks are results, a note tender and appointments. Grader: any capital-markets 8-K in the window.
Peers. June-quarter operating margins: SK hynix 76.0% (record), Samsung memory 70.0% (record); Micron’s FQ4 guide implies ~82.7%. Frozen peer-implied range 70–83%. Grader: Micron’s FQ4 operating margin vs the range.
Sensitivity — named channels, not a lookup table
- The FQ1 guide and the DRAM price direction (gauges 2–4): moves the base FY27 revenue between $205B and $235B and its gross margin between 76% and 84%; the fair-value consequence is bounded by the cap; the market’s is not.
- Fiscal 2027 capex (gauge 5): each $5B above $48B is about −$4 a share of present value; the information is in the construction share.
- The contract programme (gauge 7): a disclosed floor price is the only print-night outcome that re-derives the base mid-cycle margin (each 5 points ±$68 on the base value).
- The multiple regime — explicitly not a model channel; a ±10% reaction with no change in the gauges changes nothing in the framework.
What this print cannot settle
The discount rate; the floor prices in the contracts; when greenfield and Chinese supply arrive; whether the demand curve for memory has permanently steepened; whether a take-or-pay contract pays in a customer’s downturn. No outcome Wednesday validates or refutes the $1,114 central value; the print tests fiscal 2027’s first quarter and the direction of price. Fair-value moves, if any, stay inside $959–1,332.
Reversal, frozen
Toward the upside on: a FY27 gross margin ≥82% through the March-2027 quarter with contract prices still rising; HBM revenue or share at or above the DRAM share; contract coverage at half of revenue with floor prices disclosed; or a durable ~150bp fall in the ten-year (a discount-rate error, stated symmetrically). Toward the downside on: a quarter-on-quarter DRAM price decline before the September-2027 quarter; receivables above 80% of quarterly revenue or inventory days above 140; a customer disputing or renegotiating a contract; CXMT DDR5 at volume into a top-ten customer; FY27 capex guided above $50B without a matching revenue path.
Disclosures
Information only. TON618 Capital. This report is for information purposes only. Nothing here is an offer to sell or a solicitation of an offer to buy any security, fund interest, or digital asset, and nothing here is personalized investment advice or a recommendation regarding any instrument.
Publisher’s exclusion. All research is published solely as general, impersonal information of regular circulation. It is not tailored to the objectives or circumstances of any individual and is not issued in connection with compensation from any client. The Fund has no clients and distributes all research free of charge. On that basis it publishes in reliance on the publisher’s exclusion from the definition of “investment adviser” under the Investment Advisers Act of 1940 (§202(a)(11)(D); cf. Lowe v. SEC, 472 U.S. 181 (1985)).
Registration & conflicts. TON618 Capital is not registered as an investment adviser or broker-dealer in any capacity. The Fund is a Bitcoin fund and may hold or transact in the securities or digital assets it discusses; material conflicts are disclosed where they exist. Ownership: the Fund holds no position in MU as of the report date, and MU is not a Fund holding or a BTC-correlated instrument; this report also references, among others, SanDisk (SNDK), Seagate (STX), Western Digital (WDC), NVIDIA (NVDA), Broadcom (AVGO), Advanced Micro Devices (AMD), Intel (INTC), SK hynix and Samsung Electronics; the Fund holds no position, long or short, in any of them. Compensation: the Fund received no compensation from any party in connection with this report and charges nothing for it.
Use of AI. Artificial intelligence is used in the creation of this research. All methodology and data integrity are reviewed and approved before publication by TON618 Capital’s Chief Investment Officer, Keyth Beck; errors may nonetheless occur, and readers should verify independently.
CFA. This report was prepared to align with CFA Institute analytical standards (methodology only). CFA® and Chartered Financial Analyst® are registered trademarks owned by CFA Institute. That reference describes the analytical framework applied; it does not imply the report was prepared, reviewed, or authored by a CFA charterholder, and the report is not issued, reviewed, endorsed, certified, or approved by — nor affiliated with — CFA Institute.
Risk & feedback. Past performance is not indicative of future results. Digital assets and equities are volatile and may result in total loss of capital. Corrections and feedback are welcome — please direct them to CIO Keyth Beck at keyth@ton618capital.com. Version 1.0.