Peak Memory

Three findings do the work, and none of them is “memory is a cycle and cycles end.” Everyone knows that, including the author, who names supply as the main risk.

1. The 6× multiple is evidence against the thesis, not for it. Back out what $1,039 pays for on normalized earnings and the market has already capitalised a permanent re-rating to six or nine times Micron’s best pre-AI year ever. And the last time MU traded at a mid-single-digit forward P/E was the exact top of the 2018 cycle, from which it fell 56% while earnings were still rising.
2. You are buying commodity DRAM, not HBM. HBM is roughly 23%calc of Micron’s revenue. The other ~77% is commodity DRAM and NAND, priced on the spot and contract cycle. The AI story is the minority of the P&L; the cyclical commodity is the majority.
3. The supply response is dated, funded, and lands inside the holding period. The thesis needs pricing to hold through 2028–2029. That is precisely when SK Hynix Yongin, Samsung P5 and Micron’s Idaho fabs come online — paid for by a $37bn DRAM equipment year being booked right now.

The pitch opens with “even after the massive run in memory stocks this year, I still think the market is underestimating how strong this could become.” Here is the run it asks you to look past.

Read that table twice. Every name in the complex trades at six to ten times forward earnings after moving 340% to 650%. That is not a Micron-specific mispricing the author spotted and the market missed. It is the global market, uniformly, declining to capitalise a set of earnings it does not believe are permanent — which is the judgment the thesis needs to be wrong about in five places at once.

Worth steelmanning before taking it apart, because the structural argument is not empty and the pitch under-sells its own best evidence.

• The demand is contracted, not forecast. Hyperscaler capex for 2026 lands near $700–725bn for the big four — up about 77% on 2025’s $410bn — and roughly $830bn across the top nine cloud providers. Street projections for 2027 sit near $934bn for the big four; Nvidia’s CFO cites third-party forecasts above $1tn for 2027 and $3–4tn a year of total AI infrastructure by 2030.
• The shortage is arithmetic, not sentiment. 2026 bit supply grows about 16% against roughly 35% demand — a 19-point gap. DRAM, NAND and HBM supply-demand gaps of 4.9%, 4.2% and 5.1% are the widest since 2011.
• HBM cannibalises the commodity pool. HBM now consumes 23% of DRAM wafers, up from 19% in 2025, at a 3:1 wafer ratio. Every HBM die shipped removes roughly three dies of commodity capacity. This mechanism did not exist in 2018, and it is the single best argument in the bull case.
• Three players, and they remember 2019. Samsung, SK Hynix and Micron control over 95% of DRAM, and 2026 capex is disciplined by historical standards: Micron $13.5bn (+23%), SK Hynix $20.5bn (+17%), Samsung $20bn (+11%). Compare Samsung raising capex more than 50% in a single year in 2017.
• Pricing has moved into contracts. Micron holds 16 long-term agreements worth a $100bn revenue minimum, with $18bn of upfront customer cash deposits ($22bn including financial commitments), on take-or-pay terms with no cancellation clause. All of 2027’s output across all three suppliers is reportedly booked.

That is a real case. It supports “this cycle is longer and higher than history.” It does not support “this is no longer a cycle” — and the pitch slides quietly from the first claim to the second.

The thesis: “we still value memory stocks like this — the same old cycle, and prices are about to collapse again. That is why Micron trades at about 6 times forward earnings.” The inference is that the low multiple is the mispricing. Run it the other way.

At $1,039, what sustainable EPS is being paid for?

Micron’s best year before AI — FY2018, the top of the last great DRAM cycle — was roughly $11.50 of EPS. FY2023 was a loss. So the market at $1,039 is already paying for permanent earnings six to nine times Micron’s all-time previous peak. That is not a market pricing “the same old cycle.” The structural re-rating the thesis argues for has largely already happened — 555% of it.

What the 6.9× declines to capitalise is only the spike portion of FY27, and look at what that spike requires. FY27 consensus EPS has run from $112 ninety days ago to $150–156 today. At Micron’s current margin structure that implies revenue near $270–290bncalc, against a $203bncalc annualised run-rate exiting FY26 (Q4 consensus $50.8bn, reporting 30 September). The denominator of that “cheap” multiple needs another 35–40% of price and mix stacked on a level that is already unprecedented.

None of this proves the stock goes down. It reframes what kind of bet it is: not a mispricing you are arbitraging, but a momentum-and-estimate-revision bet that consensus keeps chasing upward. That can work. It is simply a different trade with a different risk profile than the one the pitch describes, and it wants a different exit discipline.

The pitch is entirely about AI: Goldman’s 24× usage growth, KB’s memory share of AI infrastructure spend, HBM sold out. Then look at the revenue line.

In fiscal Q3 2026, DRAM was $31.3bn, or 76% of Micron’s revenue. HBM is estimated at just over 30% of DRAM revenue. Multiply through: HBM is roughly 23% of the companycalc. The other ~77% is commodity DRAM and NAND — DDR5 modules, mobile LPDDR, enterprise SSDs — priced on exactly the contract and spot cycle the thesis asks you to look past.

Then the margin. Micron printed an 84.6% GAAP gross margin in FQ3 2026 and guided toward 86%. The 2018 cycle peak — the best memory pricing in a generation until now — topped out near 60%.

Take the two charts together and the shape is unmistakable: all-time-peak margins, on a business that is three-quarters commodity, with the rate of price increase already collapsing — and consensus extrapolating another 35–40% of revenue on top. Every one of those is a normal feature of a cycle top. None proves the top is here. All of them mean the margin of safety is thin.

The pitch says: “the biggest risk is supply. If new capacity comes online faster than expected and supply finally catches up, the memory price will fall and the thesis will change.” Correct, and it is the right risk — but it is framed as a contingency. It is not. It is a dated, funded, already-ordered schedule, and it lands in the same window the thesis needs prices to hold through.

That is roughly 20–30% of industry capacity arriving from late 2027 through 2028. And it is not a forecast — you can watch it being bought. Wafer fab equipment spending for calendar 2026 was just raised to the low-$150bn range from $135–140bn. DRAM equipment alone is up 29% to $37bn; NAND up 28% to $14bn. Memory jumped to 46% of Lam Research’s systems revenue, and Applied Materials’ DRAM-including-HBM-packaging line grew 52% year over year.

The honest counterweight, and it is a real one: this may not be enough. Against a 40%-plus HBM growth rate and 3:1 wafer cannibalisation, 20–30% more capacity may not close the gap. UBS does not expect DRAM supply and demand to balance until Q2 2028. That is why the read here is “late, not over” rather than “short it.” But “the glut arrives in 2028 instead of 2027” is a very different thesis from “memory is no longer cyclical,” and it prices very differently.

This is the finding the pitch misses entirely, and it may matter most for Micron specifically.

Through HBM3 and HBM3E, SK Hynix held something close to a monopoly on qualified Nvidia supply. Samsung repeatedly failed qualification; Micron was late. That scarcity is where the extraordinary HBM margins came from.

On 5 June 2026, Jensen Huang confirmed that Samsung, SK Hynix and Micron had all passed certification for HBM4 on Vera Rubin. This is the first HBM generation where all three vendors qualified at the start of the ramp. Counterpoint puts 2026 HBM4 share at SK Hynix 54%, Samsung 28%, Micron 18%; supply-chain checks give Hynix 60–70% of Vera Rubin volume. Samsung went from HBM4 mass production in February to HBM4E samples in May — three months.

Two consequences follow:

• Three qualified suppliers is a different pricing regime than one. The premium in HBM came from scarcity of qualified supply, not scarcity of wafers, and that scarcity is gone. Nvidia is reportedly willing to relax HBM4 specifications to widen supply further — which is what a buyer does when it wants competition, not when it is desperate.
• Micron is the third-place vehicle for the HBM story. If the argument is “HBM is structural,” this expresses it through the vendor with 18% share, where HBM is 23% of revenue. SK Hynix is the same trade at the same multiple with three times the share and a 76% operating margin.

The LTAs are the strongest genuinely-new argument in the structural case, and they are almost always cited one-sidedly. Both directions matter.

What they genuinely fix

Take-or-pay contracts with no cancellation clause, $18bn of cash already collected ($22bn with financial commitments), five-year terms running calendar 2026 to 2030. That is a real floor under the trough. Memory’s historic problem was that in a glut, price and volume collapse together; these contracts break that link for a slice of the book. Trough earnings in the next downcycle should be structurally higher than 2019 or 2023. That deserves a higher multiple, and it is part of why the whole complex re-rated.

What they also do

• The floor sits far below today. $100bn of minimum revenue across a five-year term is roughly $20bn a yearcalc — against TTM revenue of $90bn and a Q4 annualised run-rate near $203bn. The LTAs guarantee something like 10–22% of the current run-rate. They are a floor on catastrophe, not a lock on today’s pricing.
• Fixed prices cap the upside too. Customers moved from quarterly spot negotiation into three-to-five-year agreements at fixed prices. If the shortage worsens, Micron cannot reprice the contracted book. Part of what that 6.9× reflects may simply be that a chunk of FY27–28 is already sold at a known number, leaving the surprises asymmetric to the downside.
• Take-or-pay is only as good as the counterparty. These contracts sit against an AI buildout increasingly financed off balance sheet. Private credit lending to AI infrastructure has gone from near zero to over $200bn, Morgan Stanley expects another $800bn of data-centre financing over two years, and Meta’s $30bn Hyperion carries only about 20% on its own balance sheet. A take-or-pay contract with a levered special-purpose vehicle is a different asset than one with Microsoft.

Micron’s own management, asked when the shortage ends, said they have no idea. That is admirably honest. It is not a plan you can size a position against.

Where we actually are, without narrative. The upcycle is 36 months old against a historical range of 18–30 months — already the longest on record. Consensus peak timing has migrated from H1 2027 to Q3 2027 or later, with a confidence interval running into H1 2028.

The structural indicators are still green. The cyclical ones have started to yellow. This is the board worth holding a position against.

That board is the honest answer to “is it over?” No. Not yet. The single most important indicator — hyperscaler capex — is still accelerating, supplier inventories are still tight, and 2027 output is sold out. A long position here is defensible. A long position here without that board written down and checked is not.

Grant the thesis entirely for a moment. Even then, long-dated call options are the wrong structure, and this is the part of the question that matters most.

MU’s 30-day implied volatility sat near 67% in mid-August, with an IV rank around 32 — high in absolute terms, middling against its own year. Price long-dated calls off roughly 60% implied and the arithmetic is brutal.

Black-Scholes estimates at an assumed 60% long-dated implied vol, not live quotes. Pull the actual chain before acting — long-dated MU implied vol could plausibly run 50–70%, which moves these numbers materially.

You need MU up 29% to 40% just to break even, on a stock that has already risen 555% in a year. But the arithmetic is not the real objection. The failure mode is.

Sideways is not a tail scenario for a cyclical at a cycle high after a six-fold move. Sideways is what a top looks like while it is happening. An outright long call converts a correct-but-early view into a zero — and correct-but-early is the single most likely way this thesis plays out even if its logic is sound.

If you want the exposure, the ladder in order of preference:

1. Common stock, smaller size, written exit. No theta, survives sideways, and survives being early — the scenario that actually matters.
2. Deep-in-the-money LEAPS around 0.75–0.80 delta. A financed stock position rather than a volatility purchase: roughly 2× leverage with a manageable carry. It still decays if the stock goes nowhere for two years, just more slowly.
3. A call spread, never an outright call. Sell the $1,400–1,500 strike against it. You finance elevated implied vol with elevated implied vol instead of paying it outright, and you cap out around where the sell-side target already sits.

Outright at-the-money LEAPS is the one structure to rule out entirely.

Separate the two claims inside the pitch. AI infrastructure spending is structural — I believe that. Memory pricing is structural — I do not. The interesting question is whether you can own the first without underwriting the second.

Wafer fab equipment — Lam Research, Applied Materials, KLA

The most underrated expression here. Equipment makers are paid on the capex, and the capex is rising because memory is short. Crucially, they get paid more if the producers overbuild — which is precisely the risk that destroys the Micron thesis. DRAM equipment is up 29% to $37bn this year, NAND up 28% to $14bn, and memory is now 46% of Lam’s systems revenue. You are buying the cure rather than the disease. They are cyclical too and will be cut when the glut lands, but with a lag, from lower multiples, on less extreme margins.

SK Hynix — the actual HBM leader

If the thesis is specifically HBM, Hynix is the same multiple (6.79×) with 54–70% HBM4 share against Micron’s 18%, at a 76% operating margin. The caveats are real: Korean listing access, governance, and a pending US listing. Note what that last one means. A company whose stock is up 340% choosing to sell equity to the public is textbook cycle-top capital raising — a reason for care about Hynix specifically, and a signal about the complex generally.

Samsung — the laggard

Cheapest relative to its own history, most optionality on HBM4 share gains, least single-product exposure, and it did not participate in the melt-up. Against that: a conglomerate discount and a long record of execution disappointment in exactly this product.

Cash

A legitimate vehicle. “I believe the theme and I am not being paid enough to own it here” is a complete position, and it is the one the trigger board in §09 exists to let you exit into.

The position worth holding

• Own the theme, not the commodity price. A WFE name captures the AI infrastructure case without requiring DRAM contract prices to stay three to four times their 2025 level.
• If you own memory equity, own stock and own less of it. A 50–60% drawdown in memory is a normal outcome, not a tail — MU did it in 2018–19 and again in 2022, and mean reversion to even $500 from here is −52%. Size so that outcome is survivable and boring.
• No outright at-the-money LEAPS. See §10.
• Write the exit before the entry. The pitch hands you a falsification condition — supply — and then never operationalises it. Operationalise it.

What would make this note wrong

The argument here is that the cycle is late and the price already embeds the structural case. These would say it is underrated, and they are worth acting on rather than defending against:

• Customer inventories stay below 8 weeks through Q1 2027 while contract prices keep rising — meaning demand is genuinely absorbing supply rather than the channel filling up.
• Commodity DRAM contract prices hold flat-to-up through the 2027 fab ramps. That is the real test of “structural,” and it resolves on a known date.
• 2027 hyperscaler capex guides above $900bn and is funded from operating cash flow rather than private credit and SPVs.
• Micron’s HBM4E share moves toward 25% or better, changing the vehicle argument in §07.
• A large beat-and-raise on 30 September that lifts FY28 — not FY27 — consensus. That lands nine days after publication and could invalidate the deceleration read in §05 immediately.

What would take the position out

Any Tier-1 trigger from §09 firing, plus two thesis-specific ones: two consecutive quarters of negative quarter-on-quarter DRAM contract pricing, or FY28 consensus EPS revisions turning negative. Whichever comes first, no renegotiation. The point of writing it now is that it will feel wrong to act on when it happens.

Why this is not a YRT programme

YRT’s credibility rests on one thing: the acceptance bar. At least 150 trades, three or more non-overlapping windows including a hostile regime, real bid/ask costs, walk-forward, three months of forward paper, bootstrapped forward distributions shown beside a 50%-edge-decay scenario, and a risks section that does not soften. The graveyard notes — 0DTE and GEX, the overnight effect, the SPY 0DTE panel — are what make the published programmes worth reading.

A discretionary single-name macro position is n = 1 and cannot be falsified by that machinery. Putting it in the same book quietly tells a reader the bar is negotiable. Two further objections are specific to this book. The QQQ opening-range programme is already long AI beta, and MU is a three-to-four-times-beta expression of the same factor — the “adding SPY is not a second independent bet” warning from Note 01 applies harder here, not less. And a large single-name option position consumes buying power in the same account the systematic programmes trade, which has already gone wrong once when an assigned 0DTE position left stock on margin blocking the candle book.

The version that could be a programme

There is a real YRT note hiding inside this pitch, and it is not the story. DRAM contract pricing is a public monthly series running back to the 1990s, spanning four or five complete boom-bust cycles. That is a testable question: does memory-equity return respond to DRAM contract-price momentum, or does the equity front-run the print? Thirty years, walk-forward, real costs, proper out-of-sample. If the answer is “the stock leads the contract price by two quarters and there is no tradeable edge,” that is a graveyard note — and worth as much to the brand as a live one.

House rule: the limitations section does not soften.

• No backtest. No walk-forward. No out-of-sample. n = 1. This is a discretionary judgment about one security at one moment. It does not clear the acceptance bar and should not be read as though it did.
• Figures are compiled from public secondary sources on 21 September 2026, not from a maintained data pipeline. Unlike Note 01, where no figure on the page is typed by hand, several numbers here are typed by hand and several more are derived: every derived figure carries a calc tag, and they include the implied normalized EPS ladder, the FY27 implied revenue, HBM as a share of total revenue, the LTA annualisation, and all option premiums.
• Option premiums are Black-Scholes estimates at an assumed 60% implied vol, not live quotes. Long-dated MU implied vol could plausibly run 50–70%, which moves the breakevens materially in either direction.
• LTA terms come from press reporting, not the 10-K. The take-or-pay mechanics, the fixed-versus-floating price split, and what share of FY27–28 volume is actually contracted matter a great deal to §08, and none of it was verified against primary filings.
• Micron reports Q4 FY26 on 30 September 2026, nine days after this note. A large beat-and-raise could invalidate the momentum-deceleration read within the week.
• The 2018 comparison is an analogy, not a model. HBM’s wafer cannibalisation and the take-or-pay contract regime are genuinely new. “It rhymes” is weaker evidence than the confidence of §04’s prose might imply.
• No position is held in any name discussed, and no non-public information was used. This is analysis, not investment advice.