Thesis

Micron beats the S&P 500 and the AI sector — with less risk.

Matt Hargrove · August 2026

Five things the market is missing

1
Demand is twice supply right now — and widens through 2028AI data centers +60% a year · supply +20%
2
Prices keep rising — HBM reprices in 2027no new fabs before mid‑2027 · lines switching from DRAM to HBM · AI buyers pay whatever it costs
3
Half of revenue floored at 70–75% to 2031take-or-pay · $18bn cash deposits
4
$110bn of cash for buybacks by early 2027≈10% of the company · opens Dec 2026
5
6× forward P/E — NVIDIA 22×10× on today's earnings = +50%
The bet
+50%
in the next 12 months
The risk
floored
a crash reaches Micron 6–12 months late

To beat the market, disagree with it — and be right.

the August 2026 quarter is company guidance

The bear case, at its strongest

CRUX 1
Memory is a commodity. It always booms — and always busts.
CRUX 2
A wall of new supply — and China — crushes prices while AI spending cools.
CRUX 3
85% margins and 6× are what a cycle top looks like.

Twelve bear claims — each one answered

CLAIM 01
Memory is a commodity
the answer
CLAIM 02
Always boom, always bust
the answer
CLAIM 03
AI demand is a bubble
the answer
CLAIM 04
A wall of new supply
the answer
CLAIM 05
China floods the market
the answer
CLAIM 06
Hyperscalers cut spending
the answer
CLAIM 07
The stock already ran
the answer
CLAIM 08
Debt-financed build-out
the answer
CLAIM 09
Cheaper inference, less memory
the answer
CLAIM 10
The contracts aren't real
the answer
CLAIM 11
85% margins can't last
the answer
CLAIM 12
6× is right for a peak
the answer
The claim
Why it's wrong
    full text: research web app → Bear Claims tab

    Every past bust needed two things. Both are gone.

    Every past bust
    1
    A finite market — PCs, phones — that supply could overshootPC units fell seven years running (2012–18) · phones +0–3% a year · memory per device ≈ +15% a year
    2
    No price protectionprices reset every quarter · volume deals with no price · prepayments under 5% · no take-or-pay
    Now
    1
    Demand growing far faster than the old markets ever didAI data-center memory demand +60% in 2027 · more racks × more memory per rack
    2
    Supply physically fixed for two years3 makers · a fab (chip factory) takes 3+ years · nothing new before 2028
    3
    Half of revenue going under 5-year price floorstake-or-pay · ≈$18bn cash deposits · mid‑2026 to mid‑2031
    take-or-pay = pays for the minimum volume even if not taken

    Three kinds of memory — the warehouse, the filing cabinet, the desk

    NAND — long-term storage · keeps data with power off · slow · cheap per GBthe warehouse across town
    DRAM — working memory · fast · forgets when power is offthe filing cabinet beside the desk — close, but you have to get up
    HBM — DRAM stacked in towers on the same desk as the GPUthe desk — everything within arm's reach

    In one year conventional DRAM ×8, NAND ×3.5 — HBM, still on 2025 contracts, has not moved

    sources: TrendForce (DRAM spot, SSD contract) · DigiTimes (HBM) · Jan–Jun 2026 DRAM interpolated

    One HBM stack ≈ 3 DRAM wafers' worth of silicon today — going to 4

    wafer = the 12-inch silicon disc chips are cut from

    Where each kind of memory went — then, and now

    share of bits · NAND: Q2 2025 → Q2 2026 (Counterpoint) · DRAM: ≈2018 → 2026 (TrendForce) · HBM revenue: SK hynix

    Two terms set the price — both rise through 2027

    Four terms move EPS. The market prices one of them correctly.

    EPS =
    Sales price × Volume − CostsShares

    Rises through 2027, very likely 2028, then stabilizes

    1  AI data centers set the price — demand ≈ 2× supply
    2  Factory lines switching from conventional DRAM to HBM
    3  Newer, more customized memory

    About +20% a year in total — about +10% of conventional DRAM

    3 makers of advanced memory
    A fab takes 3+ years to build
    Everything producing through 2028 is built or under construction · Idaho (mid-2027) is inside the +7%
    CXMT (China): ≈12% of world bits in 2027 · no HBM · smaller than one year of data-center demand growth (+27%)

    Two costs, both known

    Fall 10–20% through 2027 via buybacks

    CHIPS Act terms block buybacks until Dec 2026 · special dividends restricted to 2029 → cash returns as buybacks
    Only $2.16bn left on the old authorization — a big new one is the announcement to watch

    Three buyers — AI data centers set the price because they pay the most

    Apple raised device prices citing memory

    Data centers want about twice what reaches them

    bn GB a year · 2025 derived from growth rates · excludes cars, robots, AI on phones and laptops

    Each rack generation is planned to carry far more memory

    2027 AI demand
    +23% × +30% ≈ +60%
    AI power 11.0 → 13.5 GW × conventional DRAM per MW 135 → 176 TB
    NVIDIA's announced specs, not shipments · NAND per rack: estimates — only Vera Rubin has a public figure · Feynman: no spec yet, estimate

    Two costs — both known

    Existing fabs · per bit
    cost per bit falls with each generation — they get better at making it

    On the EPS trend, cash to buy back about 41% of the company by August 2028

    Quarter (calendar)Micron FQNet cash
    start, $bn    		Free cash flow
    on the EPS trend, $bn    		Net cash
    end, $bn    		Cumulative buyable
    at $926, % of shares
    Jun–Aug 2026FQ4 20262423475% · blocked
    Sep–Nov 2026FQ1 20274728767% · blocked to Dec 2026
    Dec 2026–Feb 2027FQ2 2027763411010% · buybacks open
    Mar–May 2027FQ3 20271104014914%
    Jun–Aug 2027FQ4 20271494519419%
    Sep–Nov 2027FQ1 20281945124523%
    Dec 2027–Feb 2028FQ2 20282455630129%
    Mar–May 2028FQ3 20283016236335%
    Jun–Aug 2028FQ4 20283636743041%
    Dec 2026
    buybacks open
    CHIPS Act terms until then · special dividends restricted to 2029 → cash returns as buybacks
    Watch for
    new authorization
    only $2.16bn left on the old one
    FCF = 65% of trend EPS (FQ3 2026 ratio) · trend: the quarterly EPS chart · % of shares at $926 · UBS: >$400bn FCF through 2028

    Half of revenue is heading under 5-year price floors

    16 agreements · SCAs
    ≈ $100bn
    of revenue at minimum prices · mid‑2026 to mid‑2031 · take-or-pay
    Customer commitments
    ≈ $18bn cash
    deposits · of $22bn committed
    SCA = Micron's name for its 5-year customer agreements

    At the floor, half of Micron's revenue earns NVIDIA-like margins for five years

    company: floor "well above" any prior peak quarter · 70–75% from the disclosed terms · outside research range 65–75%

    The other half: a price crash would reach Micron 6–12 months late

    conventional DRAM contract index, Sep 2025 = 100 · after Sep 2026 both paths are illustrations, not forecasts

    The market pays 6× for Micron — and 19–22× for everyone else

    From "commodity, boom-bust" to "bottleneck with runaway demand"

    RPO = contracted revenue not yet delivered

    Micron vs NVIDIA — faster growth, a fraction of the multiple

    Next quarter guided: Micron $50bn (FQ4 2026) · NVIDIA $108bn (Q3 FY27) · Micron high target $2,200

    Two rivals in the world — versus a crowd that includes its own customers

    Micron
    NVIDIA
    Competitors
    SamsungSK hynixCXMT — years behind
    AMDIntelBroadcom custom chipsGoogleAmazonMicrosoftMetaTesla / xAIOpenAIAnthropic
    Moat
    A fab: 3 years and $20bn+ · HBM qualified per customer chip
    CUDA software ecosystem
    Future
    Every new GPU / TPU design carries more memory
    If NVIDIA succeeds, Micron succeeds — not the reverse

    Micron holds the contracts and the pricing power; NVIDIA holds the cash

    Micron
    NVIDIA
    Contracts
    16 × 5-year floors
    ≈$100bn at minimum prices · ≈$18bn deposits
    No floors
    "at least $1 trillion" expected through 2027
    Pricing power
    Prices ↑
    Apple, the hyperscalers, NVIDIA — because of memory
    75% → 71–72%
    gross margin, on memory prices
    Cash · debt
    $24bn net cash · 2.3% of the company
    cash $30.2bn − debt $5.7bn · market cap ≈$1.05T
    $23bn net cash · 0.4% of the company
    cash + securities $56.6bn − debt $33.4bn · market cap ≈$5.6T · plus $42.8bn of equity stakes
    "$1 trillion" — Huang, GTC March 2026

    Quarterly EPS: the trend of the last year, continued — and what lands on it

    trend = straight line through the last 5 quarters, ≈ +$7.6 a quarter · could be lower · FY27 ≈ $200 vs consensus ≈ $155 · FY28 ≈ $322

    Where each quarter's prices come from — the 2027 HBM step lands in Mar–May 2027

    Fiscal quarterEPSHBM · $ per GB
    contract year · when set    		Server DRAM · index
    priced from · CQ3-25 = 100    		NAND · index
    priced from    		vs FQ3 26
    HBM / DRAM / NAND
    FQ1 26Sep–Nov 2025$4.78reported$15.0CY2025 contract (HBM3E ≈$15) · reset lower for 2026113INDEXCQ3-25 (⅔) + CQ4-25 (⅓)106INDEXCQ3-25 (⅔) + CQ4-25 (⅓)+24% / -65% / -57%
    FQ2 26Dec 2025–Feb 2026$12.20reported$13.1CY2025 (⅓) + CY2026 (⅔)182INDEXCQ4-25 (⅔) + CQ1-26 (⅓)149INDEXCQ4-25 (⅔) + CQ1-26 (⅓)+8% / -43% / -40%
    FQ3 26Mar–May 2026$25.11reported · baseline$12.1CY2026 contract · set Q4 2025322INDEXCQ1-26 (⅔) + CQ2-26 (⅓)247INDEXCQ1-26 (⅔) + CQ2-26 (⅓)baseline
    FQ4 26Jun–Aug 2026$31.00guided$12.1CY2026 contract · set Q4 2025452INDEXCQ2-26 (⅔) + CQ3-26 (⅓)331INDEXCQ2-26 (⅔) + CQ3-26 (⅓)0% / +40% / +34%
    FQ1 27Sep–Nov 2026≈ $39trend · now$12.1CY2026 contract · set Q4 2025513INDEXCQ3-26 (⅔) + CQ4-26 (⅓)368INDEXCQ3-26 (⅔) + CQ4-26 (⅓)0% / +59% / +49%
    FQ2 27Dec 2026–Feb 2027≈ $46trend$18.5 (high $27.4)CY2026 (⅓) + CY2027 (⅔) · set Q4 2026555INDEXCQ4-26 (⅔) + CQ1-27 (⅓)393INDEXCQ4-26 (⅔) + CQ1-27 (⅓)+53% / +72% / +59%HBM high +126%
    FQ3 27Mar–May 2027≈ $54trend$21.7 (high $35.0)CY2027 contract · set Q4 2026583INDEXCQ1-27 (⅔) + CQ2-27 (⅓)410INDEXCQ1-27 (⅔) + CQ2-27 (⅓)+79% / +81% / +66%HBM high +189%
    FQ4 27Jun–Aug 2027≈ $61trend$21.7 (high $35.0)CY2027 contract · set Q4 2026612INDEXCQ2-27 (⅔) + CQ3-27 (⅓)422INDEXCQ2-27 (⅔) + CQ3-27 (⅓)+79% / +90% / +71%HBM high +189%
    DRAM and NAND priced a month before shipment (⅔ earlier CQ + ⅓ later), shown as an index (no single $/GB) · HBM = quoted $/GB · beyond Aug 2026: trend and estimates

    EPS follows the trend, P/E stays at 6× → $1,750 by the Mar–May 2027 quarter

    price = 6 × the next four quarters' EPS · baseline = consensus $155 (today's $926) · later quarters on the trend line · buybacks not included

    Any quarter, any multiple — 10× on today's earnings already clears +50%

    QuarterForward EPS6×8×10×12×15×22×
    FQ3 26Mar–May 2026$155forward EPS · consensus$926$1,240$1,550$1,860$2,325$3,410
    FQ4 26Jun–Aug 2026$200forward EPS · trend$1,200$1,600$2,000$2,400$3,000$4,400
    FQ1 27Sep–Nov 2026$230forward EPS · trend$1,382$1,843$2,304$2,765$3,456$5,069
    FQ2 27Dec 2026–Feb 2027$261forward EPS · trend$1,565$2,086$2,608$3,130$3,912$5,738
    FQ3 27Mar–May 2027$291forward EPS · trend$1,747$2,330$2,912$3,494$4,368$6,406
    FQ4 27Jun–Aug 2027$322forward EPS · trend$1,930$2,573$3,216$3,859$4,824$7,075
    price = forward EPS × P/E · baseline = consensus $155 · later quarters on the trend line · 22× = NVIDIA and the sector · excludes buybacks

    Micron beats the S&P 500 and the AI sector — at least +50% in the next 12 months, with less risk

    EPS rises
    1
    Demand = 2× supply now — increasing in 2027 and 2028AI data centers +60% a year · supply +20%
    2
    HBM jumps up in price in early to mid 20272026 HBM was priced before the run-up
    3
    Stock buybacks begin in 2027≈10% of the company by early 2027UBS projects 40% of shares by end-2028
    P/E rises
    4
    6× is the bust multipleNVIDIA 22× · S&P 20×
    5
    Every quarter without the bust removes discount30 Sep results · Oct 10‑K · Dec: buyback window opens
    Less risk
    6
    Half of revenue protected through 2031take-or-pay · 70–75% gross margin at the floor
    7
    The other half of revenue protected by 6–12 month agreementsa price crash would reach Micron 6–12 months late
    Close

    Matt Hargrove

    Financial and valuation analyst · 20 years in personal finance
    Master's degree in finance
    20 years in personal finance and financial analysis
    More than 100 companies valued professionally
    Passed the CFP and CPA examsnot certified or licensed
    Talk to me
    linkedin.com/in/matt-hargrove-03973929
    matt.hargrove5@gmail.com
    214-592-5326
    questions · the workings behind any number · work like this for you
    Disclosure
    I hold Micron call spreads. This is my own research, not investment advice.
    the research app and the source for every number come with this video
    1 / 29