Micron Technology
Vertically integrated IDM: designs and fabs its own DRAM and NAND in owned fabs (US, Japan, Taiwan, Singapore), sells to OEMs/hyperscalers/distributors. Deeply cyclical commodity historically, now partly de-commoditized via take-or-pay HBM/Strategic Customer Agreements (fact: 16 SCAs, ~$100B minimum contracted revenue through 2030, Q3 FY26 call).
The thesis on this name
State of AI Compute
Only US-listed integrated DRAM+NAND pure-play at ~10x forward with record 85% GM and $100B+ committed backlog — cheap ONLY if the super-cycle outruns a normal one; the verify flagged the 'buy the dip' entry rests on a mis-dated/mis-attributed selloff (the larger June-4/5 leg was a Broadcom AI-demand miss, not cycle fear), so gate entry on realized prints +…
State of AI Compute
Long Micron over 9-15 months: at ~$1,232 (7.4x fwd EPS on a $31±1 Q4 FY26 guide and a $100B contracted backlog) the market is pricing a normal cyclical roll-over, but a structurally supply-constrained, contract-locked AI-memory cycle (HBM sold out through 2027, memory rising from ~30% to ~48% of hyperscaler capex) kee…
State of the Memory Supercycle
The only US-listed pure-play memory maker and the cleanest expression of the cycle. FQ3'26 printed ~81% gross margin (vs 27% a year earlier), HBM4 revenue already >$1B, and the entire 2026 HBM book is sold out under multi-year contracts with ~$22B of hyperscaler cash deposits prepaid (Micron FQ3'26 prepared remarks, Jun 2026). FQ4 guide is ~$50B revenue and ~86% gross margin. Critically the bull case is no longer HBM-only: TrendForce flags DDR5 contract profitability surpassing HBM3e in 2026 (Oct 2025), and Micron's 3:1 HBM-to-DDR5 wafer conversion means every HBM ramp tightens the far-larger commodity DRAM pool Micron also sells. Through-cycle this is the lowest-cost-curve US name with the strongest free-cash inflection; own it for the roll, sized to a cycle-aware entry not the FQ4-margin peak.
State of the Memory Supercycle
Distinct from the core MU compounder call: the mispricing is timing, not the franchise. The Street treats MU as a peak-margin cyclical to be sold at ~86% GM, but the through-cycle FCF profile has structurally improved (HBM long-term contracts + prepayments smooth the trough) — TradingView/GuruFocus note 'the memory cycle has changed' on cash flow. The asymmetric trade is to add on cycle-fear pullbacks (a CXMT-glut headline, a single soft DDR5 print) rather than chase the margin peak. Size the rent for the roll: the entry band, not the franchise, is what's undervalued here.
State of the Memory Supercycle
Avoid CHASING memory makers into the FQ4'26 ~86%-gross-margin peak. The weighted cycle peak clusters around Q1'27 (range H2'26–H2'27) and TechInsights forecasts a 2027 semiconductor/memory downturn; no prior memory upcycle has lasted beyond ~30 months and this one is already at that threshold (useLuminix / blocksandfiles, 2026). Historically the moment DRAM feels safest — sold-out capacity, record prices, rising targets — is closest to the top. This is an avoid-the-entry call on the same MU we own: the franchise is fine, but paying peak-margin multiples at month-30 of an unprecedented upcycle, into a live CXMT glut falsifier, is the canonical way to lose money in memory. Buy the dislocation, do not chase the print.
State of the Memory Supercycle
Own the only US-listed pure memory maker through the cycle — HBM sold out 2026, record 84.6% GM, but buy the inevitable cyclical pullbacks, don't chase the post-print +16% spike.
State of the Memory Supercycle
The single best US expression of HBM + DRAM; sold-out 2026 HBM, record margins. Accumulate on >15% pullbacks — entry is late-cycle at an all-time high.
Earnings, margins, COGS & capex
Micron is in the steepest up-cycle in its history: revenue went $13.6B → $23.9B → $41.5B across Q1–Q3 FY26 (fact), gross margin 56.8% → 74% → 81.2% (fact), and Q4 FY26 is guided to ~$50B revenue at ~86% gross margin and ~$31 EPS (fact). The swing is driven by AI data-center DRAM, sold-out HBM, and a pricing spike, with ~$100B of minimum contracted revenue now locked through 2030 (fact). The key debate is not whether numbers are great today but how much of the ~86% margin and the price spike is peak-cyclical vs structurally durable — Q4 guidance itself flags 'meaningful moderation in the rate of price increases.'
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~15¢ is cost of goods and ~4¢ operating expense, leaving ~81¢ of operating profit (~58¢ net).
Revenue trend
Margins
up (56.8%→74%→81.2%→~86%)
up (+12pts QoQ; opex ~$1.6B fixed vs surging rev)
up
up (operating CF $25.4B)
COGS structure
COGS is dominated by silicon-wafer fab cost: depreciation of multi-billion-dollar fabs (the single largest fixed cost — every node/capacity build adds D&A), raw silicon wafers, ultra-pure chemicals/gases, photomasks, and high power/water for fab operations, plus advanced packaging/test for HBM (TSV stacking is yield-limiting and cost-heavy). In a memory up-cycle, price runs far ahead of largely-fixed COGS, which is why gross margin explodes from 57% to 86%; in a down-cycle the same fixed D&A crushes margin negative. HBM carries 3-4x the silicon content of standard DRAM per bit (est.), so HBM mix is cost-dilutive on bits but margin-accretive on revenue.
Capex
FY26 capex ~$27B net of government incentives (fact, Q3 call), with ~$10B guided for Q4 alone — funding HBM/advanced-DRAM capacity (1-gamma/EUV nodes), HBM4/HBM4E TSV packaging, and US greenfield fabs (Idaho, New York) backed by CHIPS Act incentives. This is the highest capex in company history and the central bear risk: it adds future D&A and supply that could re-commoditize the market when AI demand normalizes.
Latest earnings
Beat on both lines. Revenue $41.46B (fact) vs consensus of $35.25B-$35.82B (est., varies by provider); non-GAAP EPS $25.11 (fact) vs consensus $20.28 (est.). Consensus figures are analyst data, not filing data.
Q4 FY26 (fact, issued Jun 24 2026): revenue $50.0B ± $1.0B; non-GAAP gross margin ~86%; non-GAAP opex ~$1.65B (~$1.86B GAAP); diluted EPS $31.00 ± $1.00 non-GAAP ($30.73 ± $1.00 GAAP). Capex: FY26 ~$27B net of government incentives, FY27 'above the mid-$40 billion range' with more than half tied to construction (fact). No price-moderation language is verifiable in the Q3 materials — the phrase is absent from the earnings press release and from a transcript pass; on the call management pointed instead to supply tightness persisting well beyond 2027, alongside the ~86% gross-margin guide.
- DRAM revenue (Q3 FY26)
- $31.3B, 76% of revenue, +343% YoY, +67% QoQ (fact)
- NAND revenue (Q3 FY26)
- $9.9B, ~+361% YoY, nearly doubled QoQ on mid-single-digit % bit growth (fact)
- Data-center revenue (Q3 FY26)
- >$25B in the quarter, a >$100B annualized run rate; data-center SSD revenue >$5B, more than doubled sequentially (fact, Q3 FY26 call)
- Contracted backlog (SCAs)
- 16 Strategic Customer Agreements, ~$100B minimum contracted revenue, covering ~20% of DRAM volume and up to one third of NAND volume; $22B+ in total cash and financial commitments including ~$18B of upfront cash deposits already received (fact, Q3 FY26 call — a slide summary instead reported '$22B in upfront customer cash'; the transcript figure is used here). Management targets ~half of company revenue under SCAs over time.
- Cash generation & balance sheet (Q3 FY26)
- Operating cash flow $25.39B, adjusted FCF $18.30B (~44% of revenue), net capex $7.08B (fact). Cash + marketable + restricted investments $30.13B vs $5.72B total debt ($582M current, $5,140M long-term) = net cash ~$24.4B as of May 28 2026 (fact).
- Capital returns
- Buybacks paused in fiscal Q3 (after $300M in Q1 and $350M in Q2 FY26); quarterly dividend raised 30% to $0.15/share, with a stated intent to return 100% of excess cash over time, primarily via buybacks (fact; the raise appears to date from the Q2 FY26 report — exact announcement date not verified).
Growth drivers
- HBM (high-bandwidth memory) for AI accelerators — sold out through CY2026, booked into 2027, HBM4 revenue already >$1B and ramping 2x faster than HBM3E (fact, Q3 FY26)
- Data-center DRAM + high-cap server DIMMs — data-center revenue exceeded $25B in Q3 FY26 (fact), now the majority of the business
- Severe DRAM/NAND pricing — tight industry supply (Micron can fill only half-to-two-thirds of HBM demand) driving ASPs up across the board
- Data-center SSD / enterprise NAND — Q3 data-center SSD revenue more than doubled QoQ (fact); NAND revenue +361% YoY
- Strategic Customer Agreements — take-or-pay contracts (~$100B minimum, through 2030) converting commodity sales into contracted, margin-floored revenue
- HBM4E custom base-die with TSMC (2027) — moves Micron up the value chain into customized, stickier AI memory
Reported financials — SEC EDGAR
Audited GAAP figures pulled from SEC filings · latest filing 2025-10-03. The audited primary-source spine — not financial advice.
Revenue — annual (GAAP)
Margins & balance sheet — FY’25
Bull & bear
At $823.03 (fact, Jul 31 2026 close) and a $929.52B market cap (fact) — roughly $400B of value off in five weeks — the market pays ~6.6x the annualized Q4-guide run-rate ($31.00 non-GAAP EPS x 4 = $124, fact/guide) and 18.6-19.8x trailing (fact) for a business whose reported economics did not break. Q3 FY26 non-GAAP gross margin was 84.9% (84.6% GAAP) (fact), running ~3.7pt above the 81.2% non-GAAP operating margin. HBM4 is in high-volume production for NVIDIA Vera Rubin, CY2026 HBM supply is sold out with pricing and volume locked, and the 16 SCAs carry ~$100B of minimum contracted revenue through 2030 plus $22B+ of total cash and financial commitments including ~$18B of upfront cash deposits ALREADY RECEIVED (fact, Q3 FY26 call) — the strongest de-risking fact in the case.
- The July de-rate was multiple compression, not an earnings event: MU fell ~28.7% in July (from a $1,154.29 Jun 30 close to $823.03, fact; -29.4% from the ~$1,166 Jun 25 close) inside a sector-wide move — the SOX closed Jul 31 2026 at 11,311.08, -22.7% from its Jun 22 2026 record close of 14,634.72 (fact, Nasdaq official index page) — while Q3 DRAM contracts were reported settling 20-30% HIGHER through the selloff and Q4 remains guided to $50.0B ±$1.0B revenue at ~86% non-GAAP gross margin and $31.00 ±$1.00 non-GAAP EPS (fact, guidance issued Jun 24 2026)
- Correcting our own error strengthens the case: 84.9% non-GAAP gross margin (84.6% GAAP) on $41.46B revenue in Q3 FY26 (fact) — a company record — with 81.2% non-GAAP operating margin (80.4% GAAP), non-GAAP net income $28.86B, $25.39B operating cash flow and $18.30B adjusted FCF (~44% of revenue, fact). The 81.2% we had published as gross margin was the operating margin in the wrong field, a 370bp understatement of the single metric this thesis argues about
- The contracted book is bigger than we described: 16 Strategic Customer Agreements, ~$100B minimum contracted revenue through 2030, covering ~20% of DRAM and up to one third of NAND volume, plus $22B+ in total cash and financial commitments including ~$18B of upfront customer cash deposits ALREADY RECEIVED (fact, Q3 FY26 call; a slide summary instead reported '$22B in upfront customer cash' — transcript figure used, conflict flagged). Management targets ~half of company revenue under SCAs over time. Cash in hand ahead of shipment is a materially harder floor than a signed minimum
- The HBM qualification risk we were watching has cleared: high-volume production of HBM4 (36GB 12-high) designed for NVIDIA Vera Rubin at >11 Gb/s pin speed and >2.8 TB/s bandwidth (~2.3x HBM3E) with >20% better power efficiency, and 16-high 48GB samples shipped to customers (fact, Jul 2026); calendar-2026 HBM supply is fully sold out with pricing and volume locked, HBM4 included. Data-center revenue exceeded $25B in the quarter (a >$100B annualized run rate) with data-center SSD revenue >$5B, more than doubled sequentially (fact)
- The balance sheet absorbs the capex step-up without dilution and on a long timeline: ~$24.4B net cash at end Q3 FY26 ($30.13B cash, marketable and restricted investments vs $5.72B total debt, with $4.4B of debt retired in the quarter, fact) against $18.30B of quarterly adjusted FCF and ~$18B of customer cash already received. More than half of the FY27 build is construction — first wafers mid-2027 at Boise, the second fab late-2028 (fact, Jul 9 2026) — so it is a self-funded multi-year commitment, not a leveraged race to add near-term bits
The bear case does not rest on this pass's corrections — all three of them cut the other way, and we say so plainly. It rests on two things that verify. First, essentially all of the sequential growth is price: DRAM bit shipments rose only low-single-digit % QoQ while DRAM ASPs rose low-60s % QoQ (fact, Q3 FY26 call), and the same fixed fab D&A that produces an 84.9% gross margin produces a negative one when price reverts. Second — and this is where a high conviction is genuinely exposed — the 'structural, contract-locked' claim rests on floor prices covering only ~20% of DRAM volume (fact), so roughly four fifths of DRAM volume prices outside that mechanism, and the mechanism has never been run through a memory down-cycle. The FY27 capex guide of 'above the mid-$40 billion range' versus ~$27B in FY26 (fact) and a now-listed, funded CXMT (fact, Jul 27 2026) are real additions to the file, but both land with a lag and on the conventional-DRAM tail the thesis already discounted. We retire the 'meaningful moderation in the rate of price increases' quotation outright: it was the stated basis of our bearish ASP signal and it cannot be found in the Q3 press release or in a transcript pass; on that call management pointed instead to supply tightness persisting well beyond 2027 while guiding gross margin UP to ~86%.
- Growth is price, not bits: DRAM bit shipments +low-single-digit % QoQ against DRAM ASPs +low-60s % QoQ (fact, Q3 FY26 call). A +345% YoY / +74% QoQ revenue line built almost entirely on price is not a steady-state growth rate, and this is the load-bearing number for the mean-reversion debate
- Coverage, not capex, is the actual structural vulnerability: the SCAs floor ~20% of DRAM volume and up to one third of NAND volume (fact), which leaves ~80% of DRAM volume pricing outside the floor mechanism, and take-or-pay floor pricing has never been tested through a down-cycle. If the 'structural' half of this thesis fails, it fails here — not on a competitor's IPO and not on our own file errors
- The FY27 capex step-up, restated to what it supports: guided 'above the mid-$40 billion range' vs ~$27B in FY26 (fact, disclosed Jun 24 2026), plus >$250B of planned US investment through 2035, first concrete at Clay NY more than a quarter early, two Boise fabs under construction (first wafers mid-2027, second late-2028) (fact, Jul 9 2026). More than half is construction, so it is neither near-term bit supply nor near-term D&A, and it is self-funded against $18.30B quarterly FCF and ~$24.4B net cash. The tell is not the capex number but whether contracted coverage expands alongside it — uncontracted capacity is the glut mechanism; contracted capacity is not
- China supply is now a named, funded entity — with the mechanism stated correctly: CXMT listed on the Shanghai A-share market Jul 27 2026, raising RMB 57.92B (~$8.55-8.6B) at RMB 8.66/share, the largest mainland-China semiconductor IPO on record, closing up ~466% on debut; it is on track to exit 2026 at ~350,000 WSPM, within ~25,000 WSPM of Micron's own capacity, at ~7.6% global DRAM share, up from ~4.7% a quarter earlier (fact). The IPO is a capital-markets event — the ~350,000 WSPM plan pre-dates it — and >98% of CXMT revenue is COMMODITY DRAM with effectively no HBM presence (fact). That is confirmation of a risk this thesis already carried, on the segment it already discounted, not falsification of it
- The competitive-position leg of the bear case is currently UNQUANTIFIED, not disproven: our prior '#3 in HBM at ~20% share, ~20% of NVIDIA's HBM4/Rubin allocation vs SK Hynix's ~two-thirds' has no primary source and is withdrawn (a P-5 recurrence). Verified share is OVERALL DRAM — Samsung ~39%, SK Hynix ~29%, Micron ~22% (fact) — which is not an HBM figure. The honest statement is that we do not know Micron's HBM share, and that gap sits on the bear side of the ledger
- Capital returns were pulled in while FCF peaked: buybacks were PAUSED in fiscal Q3 (after $300M in Q1 and $350M in Q2 FY26) against $18.30B of quarterly adjusted FCF, with the quarterly dividend raised 30% to $0.15/share — immaterial against that cash generation (fact). Cash is being routed to a fixed-cost base that is being enlarged for a decade at the point in the cycle when current-quarter economics are at a record
What it is worth
Forward earnings / through-cycle normalization (P/E on an annualized run-rate, sanity-checked against mid-cycle EPS) — unchanged, because the debate is still peak-vs-normalized earnings, not asset value. What changed is the ANCHOR, not the ladder. All figures est. except cited inputs.
~$600-750
cyclical roll-over: capacity adds (Micron's own FY27 >mid-$40B build plus a funded CXMT at ~350,000 WSPM) re-commoditize conventional DRAM, margins mean-revert against a permanently larger fixed-cost base, and the multiple compresses on falling normalized EPS. Held at the same range: the SCAs floor revenue on ~20% of DRAM volume but not the ~86%-guide spot economics on the rest, and that uncovered ~80% — not the capex line — is what this rung is really pricing.
~$1,200-1,300
super-cycle real but moderating, with contracted visibility priced against eventual normalization. NOTE: the prior 'current price roughly fair' framing was written at ~$1,166 and is withdrawn at $823.03; this rung no longer describes the tape.
~$1,600+
structural-cycle case: the ~$100B contracted book plus ~$18B of upfront customer cash holds, HBM stays sold out beyond CY2026, FY27 EPS sustains near the $124 annualized run-rate, and the market pays a structural multiple on contracted earnings. The corrected 84.9% non-GAAP gross margin makes this rung's premise stronger than when we set it.
LEVEL HELD. Re-anchored from the ~$1,166 Jun 25 2026 close to $823.03 (fact, Jul 31 2026 close) and a $929.52B market cap (fact) after roughly $400B came off in five weeks. Three items the read turns on: (1) Q3 FY26 gross margin is 84.9% non-GAAP / 84.6% GAAP, ~370bp above the 81.2% non-GAAP operating margin — the metric this thesis is about; (2) the 'meaningful moderation in the rate of price increases' quotation is unsourceable, so no bearish ASP signal rests on it; (3) the '#3 in HBM at ~20% share vs SK Hynix ~two-thirds' claim has no primary source and is withdrawn. At $823.03 the stock is ~6.6x the annualized Q4-guide run-rate ($31.00 x 4 = $124, fact/guide) and 18.6-19.8x trailing (fact). The market now sits BETWEEN the bear and base rungs, so the current price is below the base rung rather than roughly fair against it, and the '$1T+ / $1.3T memory cap' editorial device is inoperative at $929.52B. Rung levels are held unchanged: the 84.9% gross margin raises the earnings base the bull rung assumes, while the FY27 capex guide raises the fixed-cost base the bear rung discounts — but on a mid-2027 / late-2028 wafer timeline, self-funded, and matched against ~$100B of contracted minimums, so we decline to tilt the bear rung on it. The question the ladder actually turns on is unadjudicated until the Q4 FY26 print (due ~late Sep 2026, est.), guided to $50.0B ±$1.0B at ~86% non-GAAP gross margin and $31.00 ±$1.00 EPS.
SWOT
Strengths
- Only US-headquartered integrated DRAM+NAND IDM — strategic/geopolitical premium, CHIPS Act incentives, and a >$250B US investment commitment through 2035 with a 40%-of-DRAM-in-the-US target (fact, Jul 9 2026)
- Record reported economics, corrected upward — $41.46B revenue, 84.9% non-GAAP gross margin (84.6% GAAP), 81.2% non-GAAP operating margin (80.4% GAAP), $28.86B non-GAAP net income, $25.39B operating cash flow and $18.30B adjusted FCF in Q3 FY26 (fact) — versus the 81.2% gross margin we previously published in error
- 16 SCAs / ~$100B minimum contracted revenue through 2030 plus $22B+ of total cash and financial commitments including ~$18B of upfront cash deposits already received (fact, Q3 FY26 call); management targets ~half of company revenue under SCAs over time
- Record balance sheet — ~$24.4B net cash ($30.13B cash, marketable and restricted investments vs $5.72B total debt; $4.4B of debt retired in the quarter, fact) against a history of leverage — funds the FY27 build without dilution
- HBM4 in high-volume production for NVIDIA Vera Rubin (36GB 12-high, >11 Gb/s, >2.8 TB/s, >20% better power efficiency; 16-high 48GB samples shipped) with CY2026 HBM supply sold out and pricing/volume locked (fact, Jul 2026)
Weaknesses
- Structurally a commodity cyclical — the sequential growth is essentially all price — DRAM bits +low-single-digit % QoQ vs ASPs +low-60s % QoQ (fact, Q3 FY26 call) — so earnings power is regime-dependent, not steady
- Contract coverage is thin where it matters and untested — SCA floor prices cover ~20% of DRAM volume and up to one third of NAND volume (fact), leaving ~80% of DRAM volume outside the floor mechanism, which has never been run through a down-cycle
- Capital intensity is extreme and rising — ~17% of revenue in Q3 FY26 ($7.08B net capex, fact; $19.6B nine-months-to-date), ~$27B FY26 net of incentives, stepping to 'above the mid-$40 billion range' in FY27 with more than half construction (fact) — future D&A and future supply, on a mid-2027 / late-2028 wafer timeline rather than a near-term one
- Competitive position in DRAM is third — Samsung ~39%, SK Hynix ~29%, Micron ~22% (fact). Any HBM-specific share figure is UNVERIFIED and must not be published — our prior '~20% HBM share / ~20% of Rubin allocation' is withdrawn, leaving the competitive leg unquantified
- Heavy customer concentration in a few hyperscalers and NVIDIA for HBM — de-qualification on one platform is material
- Capital returns are subordinated to the build — buybacks paused in fiscal Q3 despite $18.30B of quarterly adjusted FCF (fact); the dividend at $0.15/share (raised 30%) is immaterial against that cash generation
Opportunities
- AI memory cycle extends — if HBM/data-center demand stays supply-short, contracted margins persist far longer than a normal cycle — data-center revenue already exceeded $25B in Q3 FY26 (a >$100B annualized run rate) with data-center SSD revenue >$5B, more than doubled sequentially (fact)
- Convert cleared HBM4 qualification into share — HBM4 shipping in high volume for Vera Rubin and 16-high 48GB sampling (fact, Jul 2026); HBM4E custom base die with TSMC (2027) moves Micron up-stack into differentiated, stickier custom memory
- Grow SCA coverage beyond ~20% of DRAM volume toward management's ~half-of-revenue target (fact, Q3 FY26 call) — more take-or-pay plus more upfront cash is the cleanest available de-risking of the cycle, and the single item that would most strengthen the structural claim
- US onshoring as a commercial asset, not just a cost — >$250B through 2035, Clay NY first concrete a quarter early, two Boise fabs (first wafers mid-2027, second late-2028), and up to $3B committed to the US supply chain including $500M of strategic financing to GlobalWafers (fact, Jul 9 2026)
- Enterprise/data-center SSD and high-capacity DIMMs as a second AI-leveraged leg — NAND revenue $9.9B in Q3 FY26, ~+361% YoY on mid-single-digit % bit growth (fact); DRAM revenue $31.3B, 76% of the total, +343% YoY (fact)
Threats
- CXMT is now a funded, near-Micron-scale DRAM entrant — ~$8.55-8.6B raised in the largest mainland-China semiconductor IPO on record (Jul 27 2026), ~350,000 WSPM exiting 2026 (within ~25,000 WSPM of Micron), ~7.6% global DRAM share up from ~4.7% a quarter earlier (fact) — concentrated in commodity DRAM (>98% of its revenue, effectively no HBM). The listing capitalized a plan that pre-dated it; the threat is to the conventional-DRAM base, not the AI-memory franchise
- Industry-wide capacity adds — Micron's own FY27 step-up to 'above the mid-$40 billion range' (fact) alongside SK Hynix and Samsung expansion — are the textbook glut setup; the tell to watch is CY2027 supply-growth guidance against bit-demand growth, and specifically whether added capacity arrives contracted or uncontracted
- AI capex digestion — demand is hostage to hyperscaler and NVIDIA buildout. One early behavioural straw — reports that CoreWeave is exploring financial hedges against a decline in memory costs (Jul 2026) — is secondary-source only and UNCONFIRMED, and cannot carry weight until primary-sourced
- Permitting/delay risk on the US roadmap — a group of local residents filed suit against the ~$100B Clay, New York megafab campus (Jul 2026; filing date and case status NOT verified) — low near-term earnings impact, relevant to the long-dated supply picture
- Export-control / geopolitical whiplash affecting China revenue, equipment access, or customer routing — no new action was found on search, which is absence of evidence rather than confirmed absence
Moats, dependencies & bottlenecks
Moats
EUV 1-gamma, in-house HBM base die) real lead in cost/bit and HBM, but SK Hynix leads HBM and Samsung has deeper resources; parity is contested every node. Memory tech leadership is rented, not owned — it must be re-won each node with capex.
high capital barriers keep entrants out, but the three incumbents still compete on price into gluts, and China is funding entrants. The discipline is real this cycle (sold-out, contracted) but has broken in every prior cycle.
genuinely new for memory and floors revenue/margin, but covers only ~20% of DRAM volume and is at floor prices, not spot. A cycle-dampener, not a permanent pricing moat; the bull case rests heavily on it holding.
Customer qualification / switching cost (HBM design-in to Nvidia/AMD accelerators) once qualified on a GPU platform, HBM is sticky for that generation, but re-qualified each generation and allocation is re-contested. Stickier than commodity DRAM; far less sticky than a logic foundry relationship.
durable as a geopolitical/scarcity asset and incentive stream, policy-dependent. Only Western integrated DRAM+NAND maker — a scarcity premium, not a cost moat.
Dependencies
HBM and data-center DRAM are the entire growth thesis; concentrated in a handful of AI buyers. A capex pause de-rates the stock fast.
Earnings power is regime-dependent; 86% gross margin assumes tight supply. The single largest risk factor.
Applied Materials, Lam, KLA, Tokyo Electron) EUV/etch/deposition tools gate node and HBM capacity; long lead times and export controls add risk.
Moving the HBM4E base die to TSMC adds a foundry dependency and shares value, but enables customization for AI accelerators.
Two-way: controls on equipment access and on China revenue/customer routing; geopolitical whiplash can hit supply and demand simultaneously.
Advantages
- Only US-listed integrated DRAM+NAND IDM — geopolitical/secure-supply premium plus CHIPS Act incentives no pure-play foreign rival captures
- Record net-cash balance sheet (~$24.4B) — self-funds record capex and survives the next down-cycle without dilution, a first for cyclical Micron
- Contracted, margin-floored backlog (~$100B take-or-pay through 2030) that dampens the historical boom-bust earnings volatility
- Both DRAM (incl. HBM) and NAND under one roof — captures the full AI-memory bill of materials (HBM + high-cap DIMMs + data-center SSD)
- Operating leverage proven this cycle — opex held ~$1.6B while revenue quadrupled — incremental revenue drops through at ~80%+ margins
- Competitive HBM4 (>11Gbps, in-house base die) and a TSMC HBM4E custom roadmap — a credible #2/#3 with a path up the value chain
Weaknesses
- Commodity-cyclical core — earnings swing from record profit to losses on price; no escape from the cycle, only dampening
- #3 in HBM (~20% share) and only ~20% of Nvidia HBM4/Rubin allocation vs SK Hynix ~two-thirds — a follower in the highest-value segment
- Extreme, rising capital intensity (~$27B FY26 capex) that adds future supply/D&A and can seed the next glut
- High customer concentration in a few AI hyperscalers/GPU makers — de-qualification or a single platform loss is material
- SCA backlog covers only ~20% of DRAM and ~1/3 of NAND volume at floor prices — most of the book still rides spot ASPs
- Valuation leaves no room for a normal cyclical disappointment at a $1.3T+ cap on peak margins
Bottlenecks
- HBM/TSV advanced-packaging capacity — the binding constraint; Micron can fill only half-to-two-thirds of HBM demand (fact, Q3 FY26), so output, not orders, caps revenue
- Leading-edge wafer capacity (1-gamma DRAM/EUV) — limited and capex/time-gated; new fabs (Idaho/NY) take years to ramp
- EUV and advanced etch/deposition tool availability from ASML/AMAT/Lam — long lead times throttle how fast capacity can be added
- HBM yield (TSV stacking is yield-limiting) — improving yield is a constraint on effective HBM bit supply and margin
- Skilled fab labor and US construction timelines for greenfield fabs — execution risk on the onshoring buildout
- Power and water for fabs — large, growing utility footprint constrains siting and ramp speed
Top signals & trends
Top signals
The qualification milestone is cleared: Micron is in high-volume production of HBM4 (36GB 12-high) designed for NVIDIA Vera Rubin at >11 Gb/s pin speed and >2.8 TB/s bandwidth (~2.3x HBM3E) with >20% better power efficiency, and has shipped 16-high 48GB samples to customers (fact, Jul 2026); calendar-2026 HBM supply is fully sold out with pricing and volume locked, HBM4 included. What remains unresolved is allocation SHARE — the HBM-only split across Micron / SK Hynix / Samsung has no primary source, so treat any HBM-share figure as unverified. Verified OVERALL DRAM share (not HBM): Samsung ~39%, SK Hynix ~29%, Micron ~22%.
No verified evidence of ASP moderation as of the Jul 31 2026 vintage. Q3 FY26 DRAM ASPs rose low-60s % QoQ (fact, Q3 call), Q3 DRAM contracts were reported settling 20-30% HIGHER through the July selloff, CY2026 HBM pricing and volume are locked, and Q4 gross margin is guided to ~86% (fact). The cycle-top tell is unchanged — the inflection from decelerating-increase to outright decline — but nothing verified marks it yet. One early behavioural straw: reports that CoreWeave is exploring financial hedges against a decline in memory costs (Jul 2026, secondary source only — UNCONFIRMED).
The glut mechanism now has a named entity. CXMT (ChangXin Memory Technologies) IPO'd on the Shanghai A-share market on Jul 27 2026, raising RMB 57.92B (~$8.55-8.6B) at RMB 8.66/share — the largest semiconductor IPO in mainland China history — and closed up ~466% on debut (intraday as much as +531%). CXMT is on track to exit 2026 at ~350,000 wafer starts per month, within ~25,000 WSPM of Micron's own capacity, with global DRAM share ~7.6%, up from ~4.7% a quarter earlier (against Samsung ~39%, SK Hynix ~29%, Micron ~22%). The nuance that cuts the other way: >98% of CXMT revenue is COMMODITY DRAM with effectively no HBM presence, so the threat lands on Micron's conventional DRAM base, not the AI-memory franchise that carries the multiple. Micron's own FY27 capex step-up to 'above the mid-$40 billion range' (fact) is supply the industry must also absorb. Still watch CY2027 supply-growth guidance against bit-demand growth.
The demand engine. Sustained or rising AI capex extends the tight cycle; any digestion/pause hits HBM first.
More take-or-pay coverage beyond ~20% of DRAM volume would further de-risk the cycle; watch for new SCAs and rising minimum-revenue figures.
Faster HBM4 yield/capacity ramp converts sold-out demand into revenue. Management targets HBM share near overall DRAM share.
Trends
Every Nvidia/AMD GPU needs multiple HBM stacks; HBM bit demand growing far faster than total DRAM. Core driver of the super-cycle.
Drove gross margin 57%→86% in three quarters. Positive now; reverses to negative if/when supply catches up.
Micron ~$27B + SK Hynix/Samsung adds + China entrants — the classic mechanism that ends every memory up-cycle.
SCAs convert commodity DRAM/NAND into contracted, margin-floored revenue — structurally new, dampens future volatility.
Funds US fabs and positions Micron as preferred secure-supply vendor; offset by export-control whiplash risk.
Subsidized capacity at the commodity low-end — a slow, structural commoditization threat over the long run.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
EUV/DUV lithography — gates leading-edge DRAM (1-gamma) and HBM capacity. Sole EUV supplier.
Deposition/etch/CMP and broad fab equipment for DRAM/NAND/HBM process steps.
Etch and deposition; named Micron customer relationship. Critical for 3D NAND and DRAM.
Process control, inspection, metrology — yield-critical for HBM TSV and leading-edge DRAM.
Japan-listed (8035.T). Coat/develop, etch, deposition equipment — major fab tool supplier.
Foundry for Micron's HBM4E custom base logic die from ~2027 — supplier of the logic layer.
The anchor HBM customer; HBM3E/HBM4 designed into GPUs (Blackwell/Rubin). Drives the AI-memory demand thesis.
HBM customer for Instinct AI accelerators; co-target of HBM4E custom base dies.
Custom AI ASIC / accelerator ecosystem (hyperscaler silicon) consuming HBM and high-cap DRAM.
Hyperscaler — data-center DRAM, high-cap DIMMs, and enterprise SSD demand; AI capex driver.
Hyperscaler + custom silicon (Trainium/Graviton) — server DRAM, HBM and data-center SSD buyer.
Server OEMs (also HPE, SMCI) — channel for data-center DRAM DIMMs and SSDs into enterprises.
Korea-listed (000660.KS). The HBM leader: ~50-62% HBM share and ~two-thirds of Nvidia HBM4/Rubin allocation. Micron's most important competitor — sets HBM pricing and pace.
Korea-listed (005930.KS). #1 in total DRAM, pushing hard on HBM4 to regain share; Micron has recently edged ahead of Samsung in HBM. Deepest balance sheet of the three.
Mainland-China DRAM entrant (named for analysis, not a recommendation). Subsidized capacity threatening the commodity-DRAM low-end over time.
Mainland-China NAND maker (named for analysis, not a recommendation). Long-run commoditization threat to NAND; constrained by US export controls on equipment.
Japan-listed (285A.T); JV with Western Digital's flash. NAND-only competitor; relevant to the ~24% of Micron revenue that is NAND, not the HBM growth engine.
US-listed NAND/flash pure-play (post WD-SanDisk split). Competes in NAND/SSD; no DRAM or HBM exposure.