MU

We rate Micron Avoid, close to a Hold: our fair value is $654 a share against a price of $878.67 (0.74x). Only $314 of that price is cash we can date, the free cash flow of the next two fiscal years plus net cash. The other $565 pays for what Micron earns after the current memory boom, and that is a judgment.

How to read the numbers.

• Cases. The central case is our best estimate. The bear case is a severe but plausible bad outcome and the bull case the matching good one. The central case sets the view; the bear and bull show how far the value could move.
• Markup and the market normal. The markup is Micron's revenue divided by its cash cost of goods sold, so a markup of 2.55 means the memory sells for 2.55 times its cash cost to make. The market normal is the markup we expect memory to earn on average after the shortage ends, outside the contracts below.
• Floors and the floored share. Micron's strategic customer agreements are take-or-pay contracts, mostly in price bands with floors. The floored share is the part of FY2031 volume sold at the floor price instead of the market. At the floor the markup is 4.39 (61% gross margin).
• Fiscal years. Micron's fiscal year ends about 1 September, so FY2031 runs from September 2030 to August 2031.
• The Hold line. $747, the fair value at which our rule would move from Avoid to Hold.
• HBM is high-bandwidth memory, used with AI chips, and CXMT is the Chinese memory maker gaining share.

What this report does

It revises our earlier report of 2026-10-16 (Avoid, $625), which stays on the record (link). This report answers two questions Dong asked on 2026-10-17. Was the earlier report anchored too hard on five specific years? Partly: the year prices turn moves value by about $78, and the earlier report hid the part of the value we actually know inside one judged number. And did it weigh what is really different this cycle? It did not value the one hard difference, the contract floors. This report values them and splits the price into a dated part and a judged part. The answer moves only $29, because most of the contracts end before FY2031, the year the exit price capitalises. The sector note on what is different this cycle is ../sectors/memory/note.md; its sources, including the 84 memory bookmarks in Dong's vault, are in ../sectors/memory/data/sources.md.

The price is $314 of dated cash and $565 of judgment

([F] marks a reported fact and [M] a management claim.)

The judged part is a question about returns. On this model's path, the central case earns 16.6% after tax on its plant in FY2031; Micron earned 15.4% (net income over net plant) pooled over FY2014-FY2025 and 24.3% in its best cycle, FY2017-FY2022 [F, FMP statements in our 2026-10-15 Micron report]. The price needs about the best cycle, held as normal. That is not an extreme ask, which is why the view sits near the line.

What is really different this cycle, and how much we lean on it

The full ranking is in the sector note, section 2. In short:

• Hard, and valued here: take-or-pay agreements over 35% of revenue through 2030, three quarters with defined pricing, mostly in bands with floors; $32B of customer commitments, mostly cash; the CFO: "even at floor prices, we expect margins meaningfully above any prior cycle peak margins" (Q4 FY2026 call). We take the floor at the old peak, a 61% gross margin, which at this model's FY2031 cost is a cash markup of 4.39.
• Hard, but short: the agreements mostly end in calendar 2030. FY2031 runs September 2030 to August 2031, so about 9% of its volume is inside today's terms; with extensions and new agreements we use 15% (bear 0%, bull 35%).
• Real but partial: HBM takes three to four times the wafers per bit and has three suppliers; demand comes from a few hyperscalers with long plans. These lift the market normal above Micron's long average, not above its best cycle.
• Cutting the other way: CXMT went from 7.6% to 9.5% of DRAM revenue in one quarter (TrendForce) and plans to double its wafers by 2028; SK hynix, Samsung, Micron and CXMT all add cleanrooms in 2028-30; customers already cut memory per server (Micron, Q4 call) and DeepSeek's newest model needs a quarter of the HBM for its cache.
• Not leaned on: "no line of sight to balance" (an outlook; the 2018 and 2022 peaks each fell 34-80 margin points within about six quarters), and Street price paths that never fall (Citi's DRAM path rises to 2031; Bloomberg Intelligence's HBM path implies about $630B of HBM in 2030).

Is a fall from 85% to 41% in three years unprecedented? Not in speed: Micron's quarterly gross margin fell from 47% to -33% in four quarters (2022-23) and from 61% to 27% in five (2018-19). What is new is the height it starts from.

Three cases: $219, $654 and $1,420

The bear holds a trough as the normal forever (1.8% on plant), which is harsh; it is the case where the new capacity overshoots and the floors are renegotiated. FY2027-FY2028 are the same in every case (the path from our earlier Buy report of 2026-10-16, linked here, which the Q1 guide and the agreements date).

The simulation's median ($779) sits above our central case ($654)

The simulation (5,000 draws, odds.json) has a median of $779 and a mean of $791, against the central $654; 29% of draws are at or above the price. The gap is ours by design: both judged inputs have longer tails up than down. The market normal runs from 2.1 to 3.74 around 2.55, because Micron's markup has a hard floor in practice (its latest cycle averaged 2.34) but has reached 3.74; the floored share runs from 0 to 45% around 15%, because the agreements can be renewed but cannot go below zero. The central case is our most likely single path; the median says the value is more likely above it than below. The two straddle the Hold line ($747): the view rule uses the central case and reads Avoid, but this is close to a Hold, and a reader who weights the upside tails more than we do will reach one.

Our 10th-90th range ($602-$1,002) is 22% as wide as the Street's targets ($1,200-$3,000). We did not widen it: $314 of the value is dated cash in every draw, and the Street's spread is about whether peak earnings are normalised at all, which we do not treat as uncertain in that direction (a judgment we disclose).

Sensitivities (central, one at a time)

Bridge from the earlier report

The two steps are one change: the contracts move from a reason for the normal to an explicit input, so they count once. The rest of the engine is unchanged and the earlier report reruns at $625.43 (output.txt, last line).

Our $654 is 0.41x the Street's target, and the gap is peak earnings never normalised

Our $654 is 0.41x the Street's average target of $1,588.85 (FMP, 12 targets in the last quarter, horizons absent). Running the Street's own revenue through our engine (street_path.txt), then letting prices reach our blended normal of 2.83 by FY2031:

So the Street's revenue path is not the gap: even its FY2030 revenue, which rises 49% in a year, is worth $816 if prices then reach our normal. The gap is the Street's targets (7.6x its own FY2028 EPS) valuing peak earnings with no normalisation after them. We do not adopt that; it is the same error our earlier Buy report of 2026-10-16 made, which the report we revise corrected.

Why the other side could be right

• The agreements are a new kind of contract, backed by cash. If buyers keep paying the floors past 2030 to secure supply, a floored share of 41% justifies the price, and management's goal is about 50% of revenue under agreements, three quarters priced.
• Micron's best cycle (24% on plant) is the right base if HBM and three suppliers make memory less of a commodity than it was in 2014-2025.
• Prices may turn later than FY2029: Citi's DRAM path rises every year to 2031, and Micron sees 2028 tighter than 2027. A year later is +$78.

What would prove us wrong

The full list is in record.json: the floors disclosed past 2030 (up), floors amended in a price fall (down to about $549), prices still rising after Micron's ID2 fab and SK hynix's Yongin Y1 fab start (later turn), CXMT's share past 12% with 420k wafers in 2027 (lower market normal).

Review: no independent reviewer, so a self-check

No independent reviewer read this report: Dong asked our research lead to do this work alone. We checked it ourselves against the standard checklist for research reports:

Files

model.py → output.txt (cases, two parts, grids); market.py → market-output.txt; street_path.py → street_path.txt; inputs.json holds every judgment with its basis; assumptions.md; drivers.json and odds.json for the simulation.