Business

Know the Business

Micron is a global commodity producer with an AI overlay. The economic engine is bit shipments × ASP at a ~70%-fixed-cost structure, which is why through-cycle gross margin swings from −9% (FY2023) to 75% (Q2 FY2026). The single most important variable today is HBM mix: a sub-segment of DRAM that has re-rated the consolidated business from a PC/mobile commodity into a data-center memory franchise, and where market expectations are most likely to be wrong in either direction.

1. How This Business Actually Works

Micron sells bits, not chips. Revenue is bit shipments × global ASP per bit. Cost per bit falls about 15–20% per year from process shrinks (1ß → 1γ DRAM in FY2025, G8 → G9 NAND); if ASP falls slower than that, margin expands; if ASP collapses, margin goes negative within a quarter because roughly 70% of cost is depreciation, EUV tools, and skilled labor that does not flex with volume.

The non-obvious mechanics. Three details a generalist almost always misses. First, HBM costs roughly 3× the wafer area per bit of standard DDR5, so every HBM unit sold silently consumes capacity that would otherwise produce ordinary server and PC DRAM — HBM tightens the rest of DRAM whether or not demand for standard parts grows. Second, inventory write-downs are a feature, not an exception: in FY2023 Micron wrote down $1.8B of finished memory to net realizable value, which then lowered FY2024 cost of sales by $987M as that inventory sold — the cycle distorts gross margin in both directions. Third, government incentives offset capex but do not eliminate it: $6.4B of US CHIPS grants and the 35% Advanced Manufacturing Investment Credit (per FY2025 10-K, p.56) make $50B+ of US fab spend possible at a lower after-tax burden, but Micron still books the gross capex and the bit supply still hits the global market — the subsidy changes who builds, not how many bits get built.

2. The Playing Field

DRAM is a three-firm oligopoly; NAND is four-firm. Micron is the only US-headquartered scale player and the only pure-play memory name on the NASDAQ now that SanDisk has been spun out as a NAND-only stub.

What this peer set actually reveals. The "memory peer" label hides three different businesses. SK hynix and Samsung are the only direct DRAM+NAND comparables — both with HBM, both at hyperscaler scale. SK hynix is the right benchmark for what Micron is trying to become: a memory company whose margin profile is dominated by HBM mix, not standard DRAM. Samsung is the right benchmark for manufacturing scale and process leadership, but its memory segment hides inside a conglomerate so consolidated multiples are misleading. Kioxia and SanDisk are NAND-only — useful for tracking the NAND cycle but not for understanding the AI memory story. The genuinely concerning peer is the one not on this table: CXMT (DRAM) and YMTC (NAND), private state-backed Chinese producers on subsidized capital that the FY2025 10-K names as competitors but whose financials cannot be benchmarked. They are the long-tail risk to oligopoly discipline.

What "good" looks like in this industry. Look at the SK hynix 48.6% operating margin: that is the upper bound of what a memory pure-play can earn when its HBM mix is highest. Micron's 26.1% operating margin in FY2025 is mid-cycle by SK hynix standards because Micron's HBM share is still ramping. The peer comparison says Micron has roughly 20–25 percentage points of operating margin upside if it executes the HBM4 ramp in CY2026–27 — and roughly the entire 26 points to lose if HBM share stalls or AI capex breaks.

3. Is This Business Cyclical?

Yes, and the cycle hits everywhere — ASP first, gross margin second, inventory third, capex fourth. The chart below makes it impossible to miss.

The pattern is uncomfortably regular. A trough every 3–5 years (FY2008–09, FY2012, FY2016, FY2023), each preceded by industry-wide capex above $60B and bit-supply growth out-running content growth, each followed by a 1–2 year recovery, then a peak. Peak-to-trough gross margin moves 40–60 percentage points; peak-to-trough revenue moves roughly 50%. The FY2018 peak (59% gross margin, $14B net income) and the FY2023 trough (negative 9% gross margin, $5.8B loss) are not outliers — they are the geometry of the business.

Why this cycle may behave differently. Three structural changes are worth taking seriously without taking literally. (1) HBM is sold on multi-year strategic customer agreements, not quarterly spot — Micron signed its first 5-year SCA in Q2 FY2026, which dampens the high-frequency ASP swings on the highest-margin portion of the book. (2) The three DRAM oligopolists have shown unusual capex discipline in 2023–2025 — Samsung actually cut output in late 2023 alongside SK hynix and Micron, the first time it has done so in size. (3) HBM die-area economics structurally tighten DRAM at any given level of demand. None of this repeals the cycle. All of it potentially extends the upturn by 12–24 months versus history — and concentrates the eventual oversupply when the combined CY2026 capex (above $70B industry-wide) lands in 2027–2028.

4. The Metrics That Actually Matter

For Micron, standard P/E and ROE are noise. The five metrics below explain almost everything about why the stock works or doesn't.

5. What Is This Business Worth?

The right lens is normalized through-cycle earnings power plus an explicit HBM premium, not trailing P/E and not a single-year DCF. Spot earnings power at Q2 FY2026 run-rate (75% non-GAAP gross margin, 69% non-GAAP operating margin) overstates true value by perhaps 2× because no memory business sustains those margins; trailing TTM through the FY2023 trough understates value because the HBM franchise rebuilt during the trough and is now an established product line. The answer sits between them, and is mostly about how many years of HBM-driven margin uplift you are willing to underwrite.

Why I do not run a sum-of-the-parts here. Micron's segments are not separable businesses — they share fabs, share R&D, share processes, and have only existed in their current form since the Q4 FY2025 reorganization. The 45% CMBU operating margin is not "the HBM business's margin"; it is the margin earned on bits allocated to hyperscalers under current pricing. Move those bits to MCBU and the margin falls. The right way to use the segment view is as a mix barometer: every quarter, how much wafer capacity is going to the 45%-margin pool versus the 12–17%-margin pools. That ratio is the single best operating predictor of consolidated margin.

What would make this stock cheap or expensive. Cheap: the through-cycle earnings power most analysts use (roughly $4–6 EPS averaged across cycle, FY18 through FY25) is too low because it under-weights the HBM franchise that did not exist during most of that period. Expensive: trailing earnings power at Q2 FY26 run-rate ($30+ annualized EPS, implied from 75% GM) is not sustainable through the next trough, and the planned FY27 capex step-up will pressure FCF for two more years before harvest. The intelligent investor sets a normalized HBM-share assumption (somewhere between 21% and 30%), a normalized DRAM ASP assumption (FY18 was a peak, FY23 was a trough, FY25 was mid-cycle expansion), and a normalized capex assumption (40% near-term, 30% long-term), and accepts that the answer is a range, not a number.

6. What I'd Tell a Young Analyst

Watch capex, not earnings. When the three DRAM oligopolists' combined CY-year capex is below $50B, the cycle is healing. Above $70B, the next glut is being built. CY2026 is already booked above $70B — that is the asymmetric risk this cycle, regardless of what AI demand does.

HBM share at NVIDIA is the franchise. Not the company's overall DRAM share; not even total HBM industry share. Whether Micron is in the dominant qualification tier for HBM4 and HBM4E at NVIDIA (and increasingly AMD, Google TPU, Microsoft Maia) is the one piece of news that genuinely moves the long-term economic engine. SK hynix lost a quarter of a year on HBM3E qualification in CY2024 and the market punished it; Micron gained share. The next round is HBM4 in CY2026–27. Track it directly.

Negative gross margin is a real outcome, not a tail risk. FY2008–09, FY2012, FY2023 — three trough years in 15 with negative gross margin. Anyone underwriting Micron at 20× trailing EPS in good years is implicitly assuming the next trough never comes. It will. Size positions accordingly.

Government grants are not gross capex relief. A 35% Advanced Manufacturing Investment Credit (per FY2025 10-K) on $50B of US capex is real money — roughly $15–17B of after-tax value — but it offsets after the spend, not at the time of the spend. Free cash flow in FY26–FY28 will look ugly even with grants. That gap is when the stock typically gets cheap relative to the underlying franchise.

Ignore consolidated P/E. Use mid-cycle EPS plus an HBM premium. A reasonable through-cycle EPS — using FY2017 through FY2025 simple average — sits in the $4–6 range; this understates HBM. Adjust by a per-share HBM contribution that you think survives the next downturn. That is the anchor. Spot multiples on trailing or forward numbers have led generations of analysts to buy this stock at peaks and sell at troughs.

The thesis breaks on three things, in this order of likelihood. (1) AI capex inflection — a six-month pause from hyperscalers in CY2027 would collapse HBM pricing power. (2) CXMT scaling DRAM faster than expected — Chinese state-backed capacity is the structural overhang the financials cannot show you. (3) Micron losing an HBM4 generation at NVIDIA — has not happened, but has happened to Samsung on HBM3E, which is the closest historical analogue.