Micron Technology, Inc. (NASDAQ: MU)

6 September 2026 · Built from Micron's SEC filings — FY2021–FY2025 Forms 10-K and Forms 10-Q through the quarter ended 28 May 2026 — supplemented by independent industry sources for market structure, competitor economics and price history. Fiscal 2026 ended in early September 2026 and had not been reported as at this date. Not a valuation and not a recommendation.

Sources: FY2025 10-K; FY2026 Q3 10-Q.

1. What the company does

Processors compute faster than they can be fed data. DRAM holds working data close to the processor so it can be read in nanoseconds; NAND stores it durably and cheaply when the power is off. Every computer, phone, car and AI accelerator needs both, no volume-production substitute exists for either, and only three companies manufacture both at the leading edge.

Micron buys a 300mm silicon wafer and runs it through a fab operating twenty-four hours a day, seven days a week. The wafer yields several hundred dies, each holding a fixed number of bits. The filings state that the primary determinants of manufacturing cost are “process line-width, 3D non-volatile layers, NAND cell levels, process complexity … and manufacturing yield” — so cost per bit falls with technology transitions, not with volume. That is why the industry re-engineers its process roughly every eighteen months, and why depreciation rather than labour or materials is the dominant cost. Dies are packaged — for High-Bandwidth Memory, stacked twelve high and wired vertically through the silicon — tested, shipped, and collected shortly afterwards.

DRAM was 76% of revenue in the most recent quarter ($31.3bn), NAND 24% ($9.9bn). The four reporting units are organised by customer type, not by product, and were reorganised in Q4 FY2025 with prior periods restated.

Source: FY2026 Q3 10-Q. Business units reorganised in Q4 FY2025; prior-period segment figures retrospectively restated and not comparable to earlier filings as originally reported.

The line being entered is the data centre; the line being vacated is not a product but a customer set. CMBU grew from 12% of revenue in FY2023 to 36% in FY2025 while MCBU fell from 48% to 32%. The mechanism was deliberate allocation, not new capacity: Micron “shifted a portion of our DRAM supply to the data center and hyperscale cloud markets,” while MCBU “product supply was constrained to meet demand from higher-value segments” (FY2025 10-K). Mobile customers received fewer bits because data-centre customers paid more for them — and in FY2025 CMBU earned a 45% operating margin against MCBU's 17%.

2. Industry, competitive position and moat

The memory industry sells interchangeable bits built to a common JEDEC standard. A Samsung DDR5 die and a Micron DDR5 die at the same specification do the same job in the same socket. Because the product is fungible, price is set by whichever supplier is willing to sell the marginal bit; and because fixed costs dominate cash costs, every producer keeps its fabs running even when price falls below full cost. That combination is the whole reason memory cycles are as violent as they are.

DRAM is a three-firm oligopoly with a fourth, state-funded entrant. In the second calendar quarter of 2026, DRAM revenue share was Samsung 39%, SK hynix 26%, Micron 25%, CXMT 10% (Counterpoint Research, reported Aug–Sep 2026). NAND is more fragmented: Samsung $23.1bn, SK hynix and Solidigm $14.3bn, Micron $11.9bn, Kioxia $10.7bn, SanDisk $9.0bn in the same quarter, the top five together up 77% sequentially (TrendForce, 18 Aug 2026). In HBM — the highest-value product in the industry — SK hynix held 50%, Samsung 33% and Micron 18%, against 62% / 17% / 21% a year earlier (Counterpoint). Micron is third in DRAM, third in NAND and third in HBM.

The barriers that actually bind are capital, lithography and qualification. A leading-edge DRAM fab costs on the order of $10bn — Nanya's new EUV fab is budgeted at roughly $10.7bn, Micron's Hiroshima modernisation at roughly ¥1.5 trillion. EUV lithography is denied to Chinese producers by export control, leaving CXMT on DUV at 193nm, two to three process generations behind and carrying a cost-per-bit disadvantage estimated above 30%. HBM must additionally be qualified by the accelerator vendor, which takes multiple quarters. What does not bind is intellectual property in the exclusionary sense: Micron holds more than 60,000 granted patents, but the industry litigates over them rather than being kept out by them.

The margin test — the most important evidence here

Sources: Micron FY2026 Q3 10-Q; SK hynix, Samsung and SanDisk quarterly releases; Kioxia earnings call as reported, Aug 2026.

Kioxia's own decomposition settles the mechanism: blended average selling price rose approximately 70% year over year on low-single-digit bit growth. The margin is coming from price, and the price is coming from the industry — not from any one operator's execution. The consequence is that Micron's current economics are an industry condition, not a company-specific advantage. They are being earned simultaneously by every producer, including the smallest and least technically advanced, which means they are reproducible by competitors and reversible by the same mechanism that created them. Nothing in the peer data supports treating an 85% gross margin as evidence of a moat.

Three things are specific to Micron. It is the only leading-edge memory manufacturer headquartered in the United States, which converts industrial policy into cash: up to $6.4bn of CHIPS Act grants, a 35% investment tax credit on qualified US semiconductor investment, and a non-binding New York State term sheet worth up to $5.5bn over twenty-plus years. It shipped the industry's first 1-gamma DRAM node — its first using EUV — in FY2025. And it carries roughly $24bn of net cash, which is the balance sheet a memory company needs to keep investing when prices collapse. Against that, 40% of its property, plant and equipment sits in Taiwan and only 18% in the United States, while 65% of FY2025 revenue came from US-headquartered customers. The geography of its assets is far more concentrated than the geography of its politics.

CXMT went from 4% to 10% of global DRAM revenue in twelve months. It cannot win on economics — it is generations behind and more than 30% more expensive per bit — and it does not need to. With roughly 36% state-linked ownership before its IPO, about $8.6bn raised and a stated target above 600,000 wafer starts per month by 2030, it can add supply on a policy signal rather than a price signal. Oligopoly discipline, which is what made the pricing of the last two years possible, has no mechanism for disciplining a participant indifferent to returns.

The industry rewards the lowest-cost, highest-share producer with the balance sheet to keep investing through a trough. In the 2018–19 downturn, Samsung and SK hynix gained DRAM share while Micron lost it. Micron now has the balance sheet and has closed much of the technology gap, but remains the third-largest producer in all three of its markets. It is closer to being the kind of company that wins here than at any point in the period examined; it is not yet that company.

3. Growth engine

There is no acquired growth to separate out. Micron has purchased no revenue over the period; its only material transaction was the March 2026 acquisition of a wafer fab in Tongluo, Taiwan from Powerchip for $1.8bn in cash, which adds capacity from mid-calendar 2027 and contributed nothing to reported revenue. Every point of the 203% nine-month revenue increase is organic. The decomposition that matters is therefore price against volume — and Micron discloses it precisely.

Sources: FY2025 10-K and FY2026 Q3 10-Q, MD&A. Micron expresses price and volume changes as ranges; figures are as filed. Bit shipments made up the remainder of each revenue change.

Roughly four-fifths of the nine-month revenue increase came from average selling prices. Bit volume — the variable Micron actually controls — grew about 30% in DRAM and in the low twenties in NAND, rates a memory company would consider unremarkable. FY2024 is the instructive contrast: revenue rose 62% that year on a reversed composition, with DRAM bits up in the mid-40% range against a low-teens price increase. Same company, same fabs, opposite engine.

The honest summary: this is a price event occurring on a capacity base that has barely grown. Management's contribution has been to aim existing supply at the highest-priced demand — a real skill with a real margin consequence, but not the same thing as having created the growth.

4. Margin, cash and capital allocation

Two numbers carry the margin story. In the quarter ended May 2026, cost of goods sold rose 5% sequentially while revenue rose 74%. Depreciation, $8.28bn in FY2025, is the largest single cost and is fixed the moment the fab is built. Operating expenses have become almost invisible against revenue — R&D was 3% of revenue in the latest quarter against 20% in FY2023 — but that is entirely a denominator effect: absolute R&D rose 36% year over year.

Sources: FY2023 and FY2025 Forms 10-K; FY2026 Q3 10-Q. Comparability: business units were reorganised in Q4 FY2025 with prior periods restated; FY2023 predates HBM at scale, so the product mix behind these lines is not the same business; the 9M and Q3 FY2026 columns are unaudited interim periods, not annual figures; capital expenditure is gross of government incentives ($2.01bn in FY2025, $2.99bn in the nine months to May 2026).

Cash conversion is close to one-for-one — $45.7bn of nine-month operating cash flow against $47.3bn of net income — despite receivables ballooning from $9.3bn to $31.0bn. The offset came from a $3.3bn increase in payables and accrued expenses, higher tax payables, and higher consideration payable to customers for pricing adjustments. Inventory tells the sharpest story: $8.57bn at May 2026 against $8.36bn nine months earlier — effectively flat while revenue rose four and a half times. Micron is selling everything it makes, and that fact, more than any margin figure, defines the current moment.

Capex was $19.6bn gross in nine months, guided to approximately $27bn for FY2026 net of incentives, against $7.68bn at the FY2023 trough. Micron does not disclose a maintenance-versus-growth split NOT DISCLOSED, but it discloses the destination, and the timing matters more than the amount: Boise fab one produces its first DRAM in mid-calendar 2027, a second Boise fab in late calendar 2028, the first New York fab supplies from 2030, Singapore HBM packaging from the first half of calendar 2027, Tongluo from mid-calendar 2027. Not one dollar of the current programme produces a saleable bit before mid-2027.

The ranking is the finding. In the largest cash-generating period in the company's history, management repaid $9.4bn of debt, took gross borrowings from $14.6bn down to $5.7bn, and let cash build to $30.1bn — while buying back $650m of stock against $47.3bn of earnings. Buybacks are restricted under the CHIPS Act funding agreements, which is part of the explanation, but the pattern is consistent with a management team fortifying the balance sheet for a downturn rather than distributing a windfall. That is the behaviour of people who have run this business through FY2023.

The strategic customer agreements — announced, and post-dating most of the reported figures

From Q3 and Q4 FY2026, Micron began entering multi-year take-or-pay agreements with binding volume commitments — a structural departure from the short-term negotiated pricing that had defined the business for decades. Most are fixed-price or carry floor and ceiling bands; the largest have a ceiling for existing products approximating the market price in the second calendar quarter of 2026, and a floor running through the term. Micron expects $22bn of cash deposits and related financial commitments from agreements concluded to date, roughly $18bn of it in cash.

Two things belong beside that claim. The disclosed remaining performance obligation is only about $5bn, because it is measured at minimum volumes at minimum prices and excludes the agreements without price bands — it is a floor on a floor, not a backlog, and roughly one third is expected to convert to revenue within twelve months. And the floor prices, counterparties, contract lengths and cancellation provisions are not disclosed NOT DISCLOSED. The most consequential claim in Micron's current story is, from the filings, untestable.

5. Cyclicality, constraints and what to monitor

The evidence for how bad this gets sits inside the document set itself. Between FY2022 and FY2023 Micron's revenue fell 49%, gross margin went from 45% to negative 9%, and the company lost $5.8bn. It suspended the company-wide bonus, cut executive salaries, reduced wafer starts significantly below peak capacity, and cut capital expenditure from $12.1bn to $7.7bn. Industry-wide, DRAM revenue fell from $25.6bn in Q2 CY2022 to roughly $11.4bn a year later — about 55% in four quarters — with Micron's own DRAM revenue falling 41% in Q4 CY2022 alone (TrendForce). The 2018–19 downturn was milder: industry DRAM revenue fell 18.3% in a quarter, and SK hynix's operating margin went from 66% to 58%. The two most recent downturns differed by a factor of three, which is itself worth holding onto.

Where the business sits now: at an unprecedented peak on price and margin, and at an ordinary point on volume. The 85% gross margin is 40 points above the previous cycle peak of 45% in FY2022 and 94 points above the FY2023 trough, while bit shipments grew about 30% in DRAM over nine months. A record margin reached on ordinary volume is a price peak — and price peaks in this industry have never been the durable part.

Supply is genuinely tight today, and the constraint is mechanical rather than rhetorical: SEMI projects 300mm memory wafer capacity rising only from about 4.1 to 4.2 million wafers per month between 2026 and 2027, even as memory equipment spending rises from $52bn to $57bn, because capital is going into more expensive nodes and HBM packaging rather than raw wafer starts. But tightness of this kind is capital, not physics. Samsung has committed roughly $73bn across chips in 2026, SK hynix's first-half capex rose 73% year over year, and Micron's own FY2026 capex is roughly double FY2025's. That is the same synchronised capacity response that ended each of the last two cycles, arriving on the same eighteen-to-twenty-four-month lag — which places it in 2027 and 2028.

A second temporary item is easy to miss. Singapore tax incentives reduced the FY2025 tax provision by $1.05bn, worth $0.93 of diluted EPS. Singapore's implementation of the Pillar Two 15% minimum tax “largely offsets the benefit” from FY2026, and the effective rate has already moved from 11.6% to 15.0%.

6. Risks and unknowns

Cyclicality is covered above. What follows is what cyclicality alone does not capture, ordered so that the risks which compound come first.

What the sources could not answer

• Absolute average selling price or cost per gigabit, and absolute bit volumes — so Micron's cost position against Samsung's and SK hynix's cannot be computed.
• HBM revenue, HBM gross margin, and HBM as a share of DRAM bits — never disclosed separately, despite HBM being the stated reason for the entire mix shift.
• The maintenance-versus-growth split of capital expenditure.
• Strategic customer agreement counterparties, floor price levels, contract lengths, volume commitments and cancellation provisions.
• FY2026 fourth-quarter and full-year results — fiscal 2026 ended in early September 2026 and had not been reported as at this date.

Before forming a thesis, three things would need resolving: the floor-price mechanics of the strategic agreements at least in aggregate; a bottom-up view of industry bit supply growth for calendar 2027 and 2028 built from announced capacity rather than from Micron's characterisation of it; and Micron's cost per bit relative to its two larger competitors — the only number that determines who emerges from the next trough with share intact.

7. Takeaways