Business

Know the Business — Micron

Bottom line. Micron is a high-fixed-cost, oligopolistic commodity producer of DRAM and NAND that is currently living through the fattest pricing window in its 47-year history because High-Bandwidth Memory (HBM) for AI accelerators has bent the entire industry's profit pool. Gross profit just swung from negative $1.4B in FY2023 to positive $14.9B in FY2025 — and the Q3 FY2026 guide is for an 81% gross margin, a level Micron has never come within 20 points of in any prior cycle. The investment debate is not "is this a good company"; it is whether multi-year HBM Supply Capacity Agreements (SCAs) and the structural shortage of leading-edge memory wafers have made the next downcycle materially shallower than FY2023, or whether this is the same cycle Micron has lived through five times before, only with bigger numbers.

1. How This Business Actually Works

Revenue every quarter equals bits shipped × average selling price per bit (ASP). Bit shipments grow ~15–20% per year structurally. ASPs swing 50%+ peak-to-trough because three suppliers (Samsung, SK hynix, Micron) make essentially all leading-edge DRAM and each one's capex decision moves global supply by 5–15%. The cost line is dominated by depreciation: a leading-edge fab depreciates over 5–7 years whether it ships a wafer or not, so when prices fall, suppliers cut price, not volume. That is why gross margins move from negative to 60%+ in the same business across an 18-month window.

The simplest analogy is an oil refinery, not a software company. Capacity is fixed in the short run, marginal cost is well below average cost, and the same plant prints either a 60% gross margin or a -10% gross margin depending on whether the world wants 5% more or 5% less of what it produces. The structural twist is that Micron must redesign its product (move to a smaller node) every 18 months at a cost of billions, because if it doesn't, a Korean competitor will and Micron's products become uncompetitive within two years.

The big change since FY2024 is HBM — DRAM stacked vertically with through-silicon vias, sold to NVIDIA, AMD, and Broadcom for AI accelerators. It consumes ~3x the wafer area per bit of commodity DDR5, so scaling HBM cannibalizes standard DRAM supply across the whole industry. Because each HBM design is qualified to a specific GPU platform 12–18 months in advance, customers are now signing multi-year volume + price commitments rather than the one-quarter contracts that used to define memory. Management closed its first five-year SCA in Q2 FY2026 and is in discussions with multiple other customers. If those agreements stick, the business model moves a meaningful slice of revenue from "spot commodity" to "contract industrial" — the most important shift in the company's economic model in 30 years.

CMBU (cloud + HBM) almost quadrupled in one year and is now the largest segment. That single line is the entire AI thesis on Micron — and the reason consolidated gross margin went from 22% to 40% to a guided 81%.

2. The Playing Field

Micron is one of three at-scale leading-edge DRAM makers worldwide, but the peer set extends to NAND-only players (Sandisk, Kioxia) and storage-adjacent (WDC). Two facts matter more than the rest: SK hynix (the closest economic peer) just posted a 49% operating margin and 44% net margin on $66B of revenue for FY2025 — the cleanest read on what a focused, well-positioned memory pure-play looks like in this cycle. And Samsung's memory business sits inside a $1T+ conglomerate, which gives it the chaebol balance sheet to overbuild capacity through downturns and historically discipline the cycle by absorbing pain Micron cannot afford.

Market caps as of May 2026. SK hynix and Samsung financials reported in KRW; FY2025 (calendar) for both. Micron FY2025 ends August 2025. Sandisk FY2025 net loss reflects $1.6B of charges including a goodwill impairment from the early-2025 spin from WDC. NVIDIA is excluded — it is a major HBM customer, not a competitor.

Two takeaways. Micron sits in a unique spot: a clear #3 in DRAM but a credible #2 in HBM, where it has roughly tripled its share over the past year (9% → 21%) — a meaningful share grab in the highest-margin segment. And SK hynix is the comp that matters most. It is the only peer with the same product mix, the same end-market exposure, and full financial disclosure — its 49% operating margin in FY2025 is the right benchmark for what Micron's earnings power can look like if HBM execution holds.

What "good" looks like in this industry is a 40–60% peak operating margin and a balance sheet strong enough to survive the trough that always follows. SK hynix and Micron qualify on the second criterion today; Sandisk, Kioxia, and Western Digital are sub-scale or sub-segment plays whose cycle survival depends on exiting the bottom of the next downturn before a Korean competitor tries to break them.

3. Is This Business Cyclical?

Yes — violently so. Memory is the most cyclical sub-sector in technology, and Micron is the cleanest expression of the cycle because it has no logic, foundry, or services revenue to mute the swings.

The cycle hits five places in sequence:

The FY2023 downcycle is the most recent textbook example: revenue halved (from $30.8B to $15.5B), gross profit went from +$13.9B to -$1.4B in a single year, free cash flow was -$6.1B, and Micron raised $6.7B of debt to bridge the bottom. Shareholders' equity took a 12% hit. That is the realistic downside case any investor must underwrite when paying for an AI-cycle peak.

The HBM-driven hope is that this time is different — not because demand is permanent, but because a meaningful slice of revenue is now under multi-year volume + price contracts (the SCAs), customer-specific qualification creates switching costs, and HBM wafer cannibalization keeps commodity DRAM supply structurally tight. Management has publicly said the gap between memory demand and supply is the largest it has ever seen, and HBM is sold out for calendar 2026. That doesn't eliminate the cycle; it might make the next downturn 30–40% gross margin instead of -10%. The investor question is how much of that compression to underwrite.

4. The Metrics That Actually Matter

The single biggest mistake retail and even some institutional investors make on memory is anchoring on P/E, EPS, or dividend yield. Those numbers have a 10x range across the cycle and are nearly meaningless on a snapshot basis.

Two of these warrant emphasis. HBM mix is the only metric that determines whether this cycle's earnings power is durable — it is the share of revenue protected from spot pricing by customer-specific qualification and SCA contracts. The other five metrics tell you where you are in the cycle; HBM mix tells you what cycle you are in. Combined industry capex is the warning light. Each of the three suppliers has individually committed to growing supply; collectively, $70B+ of memory capex in 2026 is more than the entire NAND industry's annual revenue at trough pricing. If 2027 capex moves above $80B without commensurate AI demand, the next downturn comes in 18 months.

The 81% Q3 FY2026 gross margin guide is approximately 20 points higher than the highest GM Micron has ever posted in any prior cycle (FY2018 super-cycle peaked at ~58%). That is not a sustainable run-rate; it is the visible mark of a structural supply shortage colliding with AI demand. The right way to use it is as a ceiling, not a normal.

5. What Is This Business Worth?

This is a single-engine, through-cycle earnings business — not a sum-of-the-parts story. All four reportable segments make memory chips that share fabs, R&D, and packaging; the only meaningful product distinction is HBM versus everything else, and HBM still depends on the same DRAM wafers. There is no listed subsidiary, no holding-company discount, no separable real-estate or financial services arm.

The right valuation lens is normalized through-cycle EPS × an appropriate cyclical multiple, with explicit credit for HBM contract durability.

The math you should do in your head: across the eight fiscal years FY2018–FY2025, Micron earned on average roughly $4.5B of net income per year — through one super-cycle peak (FY18: $14.1B), one full bust (FY23: -$5.8B), one COVID disruption, and one AI super-cycle (FY25: $8.5B). At ~1.12B shares, that's ~$4 per share of through-cycle EPS. The current $640 share price is therefore ~160x simple-average eight-year EPS, but only ~85x trailing FY25 EPS, and a low-double-digit multiple of any reasonable estimate of FY26–FY27 earnings power. Whether the stock is cheap or expensive depends almost entirely on which of those numbers is the right denominator — and that depends on how durable HBM contract structure proves to be.

What would justify paying a premium versus the cyclical average: (1) genuine evidence that HBM SCAs put a higher floor under trough-cycle earnings — a 30–40% gross margin trough versus FY23's negative-9%; (2) sustained ROIC above cost of capital across a full cycle, not just the peak; (3) discipline in industry capex such that combined Samsung + SK hynix + Micron capex stays below ~30% of industry revenue. What would crush it: a Korean over-build, a hyperscaler digestion quarter, or a Chinese DRAM (CXMT) qualification at a top-five buyer.

6. What I'd Tell a Young Analyst

Watch the four signals that move the stock: TrendForce monthly contract DRAM prices (the leading indicator that's published weeks before any Micron release); hyperscaler quarterly capex guidance from AWS/Microsoft/Google/Meta (50–60% of DRAM bit demand); HBM qualification announcements at NVIDIA/AMD/Broadcom (each unlocks long-duration revenue, each loss is structural); and combined industry capex (the supply-side warning light when it crosses ~35% of industry sales).

The bull case to test, not assume: that HBM's customer-specific qualification + multi-year SCAs convert a meaningful slice of memory from spot commodity to contract industrial. If true, Micron's normalized EPS is materially higher than the eight-year average and the stock deserves a higher multiple than it has historically carried. The way to falsify this is to track how many SCAs Micron actually signs, how long their durations are, and whether non-HBM DRAM also moves to longer contracts. One five-year SCA is a thesis; six is a business model change.

The bear case to test: China. CXMT's DDR5 ramp is the long-tail risk that doesn't show up in any quarterly print until it does, and the May 2023 CAC ban is a reminder that Beijing can take ~$1B of addressable revenue out of Micron's market with a press release. Watch CXMT's qualification news at top-10 buyers — that, not US export controls, is the leading indicator of structural Chinese pressure on margins.

The thing that would change the thesis fastest: a guide-down on HBM volume or pricing for 2026. Management has explicitly stated HBM 2026 is sold out at negotiated price; walking that back invalidates the AI memory framework, and the next 12 months look more like FY2023 than FY2025.

The single biggest mistake to avoid: treating Micron like a "tech stock." It is a high-fixed-cost, oligopolistic commodity producer in the best pricing window of its history. Treat it like a cyclical industrial — value it through the cycle, weight the bear case heavily when margins are at all-time highs, and be willing to own it in size when gross margin is negative and the consensus says it's broken.