Micron closed below $900. Our rule said sell. We didn’t.

The best week of numbers

Walk the two days carefully, because the pattern is the thesis. Wednesday pre-open, ASML — our newest position, and the subject of the initiation note we published on the print — delivered close to a perfect quarter: €9.3 billion of sales against a guide of €8.4–9.0 billion, a 54.0% gross margin against a guided 51–52%, a third-quarter guide of €11–12 billion that implies the steepest revenue quarter in company history, and a full-year outlook raised to €43–45 billion — the third raise in six months. That is the answer to the question we posed last Tuesday: the 2027 order book is real, the capex boom extends, the “peak-out” case lost. ASML rose 2.2% on the day, helped by a soft June CPI print that lifted the whole tape. Micron fell 8.0% — to $904.28, four dollars above the stop.

Thursday, TSMC did it again, bigger: a fifth consecutive record quarter, a raised full-year revenue outlook, and 2026 capital spending hiked to $60–64 billion from $52–56 billion. Micron fell another 5.65%, closing at $853.20 — through the line. SanDisk lost roughly a fifth of its value across the two sessions; by Friday’s $1,354.82 close it sat 42% below its June 25 record close, with NAND spot reports and rising-supply chatter doing the storage-specific damage. SK hynix, which had jumped to $193.92 the previous Tuesday on the leveraged-ETF launches, closed Monday at $151.16 — the entire pop, round-tripped, in four sessions.

Hold the two halves of the week side by side and the market’s message is legible. The companies that spend on the AI buildout printed records and raised — and their stocks held or rose. The companies whose product that spending will eventually supply got sold on those exact same numbers. A raised ASML guide means more EUV tools shipping into DRAM fabs in 2027. A $64 billion TSMC budget means the foundry side of the complex is answering demand at record scale. To the memory bull, those headlines are confirmation; to the memory holder, they are the supply curve being financed in real time. Two weeks ago we watched the market sell Samsung’s world-record profit; last week it sold the best toolmaker print and the best foundry print of the cycle. Three record earnings events, three memory selloffs. That is not an accident. It is a repricing of the forward, happening in public, one confirmation at a time.

Chart 2 — The week, indexed

From the July 14 close: ASML absorbed its own good news and finished green — memory broke double digits on the same headlines, then ripped 11–14% today on analyst defense.

Closing prices indexed to July 14, 2026 (=100), through July 21. The two dotted verticals mark the ASML beat-and-raise (Jul 15) and TSMC’s record-and-capex-hike (Jul 16) — the sessions on which memory broke. Source: bpleon price feed.

We owe the tape one correction on our own framing. Last Tuesday we wrote that ASML’s print would settle the argument: a full 2027 backlog and “every memory multiple — including the one we’re cautious on — has to be revisited higher.” The backlog came in full, and memory multiples went down. Our either/or missed the third branch: the same capex that proves the demand boom is also the purchase order for memory’s future supply. Confirmation of the boom re-rated the toolmaker and de-rated the commodity — because tools become wafers, and wafers become price competition. That is precisely the supply-crossover logic of our July 5 re-underwrite, arriving a year early through the order book. We got the instrument right — we own ASML, which got paid on the very report that hurt memory — and the framing incomplete. Both things are true, and only one of them cost money.

The rule fired. We didn’t.

The mechanics, so the record is exact. The rule — published June 23, repeated July 2, July 8 and July 14 — was a closing rule: intraday breaks don’t count, a close below $900 does. It survived one honest test: earlier this month Micron pierced $900 intraday, closed above it, and we correctly did nothing. On Wednesday July 15 the stock closed at $904.28 — the market walked up to the line and stopped four dollars short. On Thursday July 16 it closed at $853.20. The rule said: out in full, next session. Friday came and went. So did Monday. The runner — 15 shares, the ~40% of the position we’d carried since June — stayed in the account through three consecutive closes below the line. Tonight it sits at $970.82, and the honest label for what happened is not “we re-evaluated” or “we widened the band.” It is: the rule fired and we froze.

Why? The reasons are worth naming precisely because none of them survive contact with our own standards. The stock had gapped $50 through the line on the TSMC news into a storage complex down 40% in three sessions — selling there felt like selling the bottom of a flow cascade, the same flow-not-fundamental pattern we had correctly called twice. The hyperscaler capex prints — the single most important read for the whole thesis — were five sessions away. TrendForce’s pricing data never confirmed the panic. Every one of those is a reason, and on another week we might have written a note titled “why we’re overriding the stop” before the fact, with a new line and a dated re-underwrite. That would have been a process. What actually happened was silence: the trigger printed, and we did neither the selling nor the override note. A rule you only follow when it’s comfortable is not a rule — it is a prediction wearing a seatbelt costume.

Now the part that stings in the other direction. Tonight the freeze is ahead: the runner at $970.82 is 14.4% above the ~$849 the rule would have realized, which puts the blended call about seven points ahead of the disciplined counterfactual. If we were selling you a system, this is where we’d quietly re-label the violation “conviction.” We are not. The same freeze that is ahead tonight is the freeze that rides to $625 in our own bear case — which we still assign 30% — and the fuel for today’s rip was analyst commentary: the same fuel as the July 14 pop that round-tripped inside 48 hours. One favorable draw does not audit a process. We have spent ten weeks writing that the scale-out and the stop are why this call worked; it would be obscene to abandon that logic the first time abandoning it happened to pay.

So here is the accounting, and the repair. The violation gets an F on process — recorded on the track record alongside the calls, because process failures are results that haven’t happened yet. The repair is a line with no discretion left in it: any close below $848.95 — the lowest close of this violation, the level the market has already shown it can print — and the runner is sold at the next open, in full, with no note, no debate, and no second override. And the hold itself now has a deadline: the hyperscaler capex prints land July 22–31, Microsoft on the 28th. Either that window produces a re-underwritten case for the runner that stands on its own — published, with fresh kill-shots — or we exit into those prints regardless of price. We said it above and we mean it operationally: if we break this line too, you should stop trusting anything this site says about risk, because we will have proven the risk talk decorative.

The ledger, marked — not closed

With the runner still open, there is no final scorecard — only an honest mark. Coverage began May 6 at ~$668 with a $1,100 target the Street then considered aggressive. The stock ran to $1,213.56 by June 25. We booked roughly 60% of the position through the pre-planned $950–$1,200 scale-out (average ~$1,100): +65%, realized, banked — that leg is beyond the reach of anything the runner does. The ~40% runner, marked at tonight’s $970.82, shows +45% on paper. Blend the realized and the open and the call stands at ~+57% marked to tonight, against ~+50% had the stop been honored at ~$849, and against +45% for never having sold a share. The scale-out remains the single best decision of the ten weeks: it converted a paper rally into banked gains while KeyBanc printed $1,750 and Cantor $2,000, and it is the only reason a stop violation is a footnote rather than a catastrophe risk.

Chart 3 — The ledger

The banked leg (+65%), the open runner (+45% on paper), the marked blend (~+57%) — and the ~+50% the rule would have locked. Ahead by luck is still ahead; it’s just not credit.

All legs vs. the ~$668 May 6 coverage basis. The runner bar is open (unrealized) and moves with the tape; the counterfactual assumes the published rule’s exit at the July 17 close. Source: bpleon research.

The call ledger beneath the mark, both columns, because a scorecard with one column is an advertisement:

1. Right: the boom, early. Long at $668 in May on AI-memory pricing, before the FQ3 blowout, the $100B contract framework, and the melt-up. The core thesis paid for everything else.
2. Right: the scale-out. Selling 60% into strength between $950 and $1,200 looked timid while the Street printed $1,750–$2,000 targets. It is the entire reason tonight’s note is about process, not survival.
3. Right: flows, not fundamentals. The July 1 crash (Korean ETF unwind), the July 14 pop (US ETF launches), and this week’s double-digit whipsaws — each called as money that has to move, not information. The tape keeps proving it.
4. Wrong: the June direction call. We said the FQ3 print was priced and would fade; it ripped 16%. Owned then, still owned.
5. Wrong: the ASML either/or. We framed a full backlog as bullish for memory multiples; the market read it as the supply answer being financed and de-rated the commodity instead. The toolmaker position covered the miss; the framing was incomplete.
6. Wrong, and the subject of this note: the execution. The stop fired and we froze. F on process, graded above, new line installed. The other five entries earn no offset against this one.

The position, the line, the referee

So the book tonight, stated plainly: still long the 15-share Micron runner — in violation, under the new line — plus the two positions the re-underwrite preferred all along. The DRAM basket (Samsung, SK hynix, Micron and the storage names together) took a 15% drawdown to $52.34 mid-week and has recovered most of it at $58.85; it remains the cleaner way to own memory pricing, because it holds the un-capped names that keep their upside. And ASML — initiated at ACCUMULATE with a ~$2,000 probability-weighted target — closed at $1,801.51, green on the week that broke the commodity, which is the entire point of owning the toolmaker: it got paid on the exact report that hurt memory. The barbell did its job in its first week; the single name is the part of the book that failed — or rather, the part where we failed.

On Micron the analytical map is unchanged, and it matters that we say so: base ~$1,100, bull ~$1,550 (the number BofA’s new target now sits on exactly), bear ~$625 at 30%. Nothing on the published thesis-break list — confirmed HBM4 share loss, contract prints below forecast, a 200-day breakdown — has triggered. Morgan Stanley’s claim today that Q3 memory prices could rise ~25% sits above the TrendForce +13–18% band we track; if that’s where contracts settle, it is genuinely bullish for the industry — and it accrues most to the un-capped names, which is why the basket, not Micron, is our expression of it. The two findings that pushed our target to the Street’s low end stand un-refuted: the SCA ceilings that cap Micron’s share of any price boom, and an HBM4 allocation reading 5–10% against the 20–25% it held in HBM3E.

Which frames next week honestly: the hyperscaler capex prints, July 22–31, Microsoft on the 28th, are the referee — for the complex, and specifically for the runner. If the buyers of all this compute raise again, the supply-wave fear gets a demand answer, and a re-underwritten case for holding Micron may clear our bar; we would publish it with fresh kill-shots before keeping a single share. If they blink — or if the case doesn’t clear — the runner goes, into whatever price the prints leave behind. And beneath all of it sits the mechanical floor: one close below $848.95 and the decision makes itself. Three ways out, all of them dated, none of them discretionary. That is what rebuilding a discipline in public looks like: not a promise to do better, but a structure that doesn’t need us to.