US Pre-Market Brief | 2026-09-14 (ET) Monday

Coverage window: 2026-09-11 (Friday) 16:00 ET regular-session close → 2026-09-14 (Monday) 08:17 ET. Includes weekend news, the overnight Asia/Europe sessions and this morning's pre-market. Market snapshot uniformly taken at 09-14 08:17 ET (pre-market quotes are cumulative and will still change before the open). Single-stock pre-market price, previous close and pre-market volume source: Nasdaq official extended-trading API; futures/commodities/Treasuries: CNBC; Treasury closing yields: US Treasury daily yield curve CSV.

---

0. One-Sentence Summary of the Day

1. Strongest catalyst: Anthropic CEO Dario Amodei published "We Must Pace the Frontier" on 9/12 (Saturday), arguing for slowing the advance of frontier AI capability; Altman and Musk echoed him the same day, and the market read it as "AI capex is going to decelerate" — a broad selloff across the AI compute chain.
2. Strongest theme: the AI capex chain selling off (I measured 45 AI hardware/compute/power names — 0 advanced, mean −5.9%); the mirror image is money rotating into software (25 of a 26-name sample advanced, mean +2.0%). ⚠️ But the attribution "software up = AI disruption threat postponed" is not confirmed by the cross-section; see the self-correction in §2 — the evidence on the short side is stronger than on the long side.
3. Second independent main line: Saudi Arabia's East-West pipeline (Petroline) was shut after a drone strike on 9/10; Brent +3.43% to $108.20, WTI +3.09% to $103.14, energy equities broadly higher; diesel prices at an all-time high.
4. Driver type: primarily opinion/industry narrative (AI deceleration), layered on a geopolitical supply shock (oil) and macro (Wednesday's FOMC, CME FedWatch ~85.5% probability of a 25bp hike). Essentially no earnings-driven action today.
5. Pre-market state: S&P futures −0.69%, Nasdaq futures −1.69%, Dow futures −0.39%; VIX 17.61 (+11.17%); 10Y Treasury 4.977%, 30Y 5.359%, 2Y 4.634% (essentially flat vs Friday); Dollar index 99.49 (+0.37%); gold −1.84%, silver −2.7%. → 🔴 Characterization corrected: the first draft called this "an inflation/real-rate shock", but my own Treasury data does not support that claim. Today 10Y/2Y/30Y are all within ±0.5bp of Friday; the curve has barely moved. The real rate-hike repricing happened on Friday (2Y jumped from 4.56% on 9/10 to 4.63% on 9/11, +7bp), not today. Correct statement: today's combination is "equities down, yields unchanged, gold down, dollar up, oil up" — gold and the dollar are pricing rate-hike expectations, but the bond market has not confirmed it today; the two are inconsistent. What can be stated with confidence: this is not a classic geopolitical risk-off (no flight-to-quality bid in Treasuries).

---

1. News Overview

🔴 Item 10 must be discounted (found on second-pass review): these ratings all come from a recurring column of the "Monday's Top Analyst Calls" type; the column headline says "Monday" but the content often includes research published on the prior trading day. The second pass confirmed: Wells Fargo's ACN downgrade actually happened on Friday 9/11, and the $194 target was already $194 back in July this year (when the rating was still Overweight) — this time only the rating changed; the target did not move a cent. → Therefore: ACN's +2.96% today should not be read as "rising in the face of a downgrade" — that downgrade was already in the 9/11 close; the release timestamps for HPE, IREN and GEV could not be verified in this piece, and any reasoning that depended on "published this morning" has been flagged or withdrawn below. | 11 | 09-11 close | Multiple | US equities broadly higher Friday: S&P +0.86% to 7,656.98, Nasdaq +0.96% to 26,333.04, Dow +0.98% to 52,573.29; ending a four-day losing streak. The drivers were "oil pulling back + CPI broadly in line" | Market | Whole market | Background | B+ | CNBC¹¹ |

Key timeline: oil pulled back (Friday) → Friday's rebound; from Friday's close through the weekend, the pipeline shutdown gaining traction + the Amodei essay each knocked out one of the two premises of Friday's rebound. This morning's pre-market is a direct giveback of Friday's rebound, plus an entirely new AI narrative shock.

---

2. Strongest Themes, Descending

Theme 1's logic robustness has to be unpacked (the most important passage in this piece)

🔴 Major correction (rewritten after second-pass review): the first version of this section asserted that "the essay does not call for cutting compute" — that was wrong and has been overturned by the primary text. The first version relied on a summarized paraphrase of the essay and missed key paragraphs; a line-by-line comparison against the full text confirmed the omission. The original conclusion is void; what follows is the corrected version.

The essay contains two mutually balancing sides at once, and quoting only one of them produces a wrong conclusion:

Side A (the only side I saw in the first version): pacing "does not mean halting model training or technical progress", but rather ensuring companies leave enough time for alignment and safety verification. Step one is giving third-party evaluators such as METR resident, employee-level access (Anthropic committing unilaterally, OpenAI indicating it will follow) — this genuinely is a governance/transparency commitment.

Side B (which I missed in the first version, and which points directly at compute): the essay states verbatim —

"Frontier AI companies within democratic countries coordinate to establish common safety standards as well as limits on the rate of unchecked AI progress." "We should also consider pacing based on limiting the ingredients that go into frontier models, such as training compute, the nature of training runs, or internal use of AI to improve AI." "We must slow the pace at which we improve the capabilities of AI models."

So: the market reading this essay as "compute is going to be constrained" is not a misreading. "Limiting training compute" is explicitly listed by the author as one of the pacing levers worth considering. My first-version label of "catalyst and conclusion do not match" was wrong and has been withdrawn.

So how strong is this catalyst really? The constraint is real in direction but very weak in enforceability, for three reasons, all of them checkable:

1. Beyond step one, everything is coordination intent: steps two and three rely on "coordination among frontier companies within democratic countries" and "international coordination with authoritarian states" — no enforcement mechanism, no timetable, no signatories. The only commitment actually in place is third-party evaluator access (Anthropic unilaterally + OpenAI indicating it will follow), and that one consumes not a single GPU.
2. The compute limit is phrased as "should also consider" — i.e., listed as an option for discussion, not a commitment.
3. Politics is pushing the other way today: Trump explicitly rejected it ("Whoever wins with AI wins"), and White House AI czar David Sacks said Anthropic/OpenAI form a "duopoly" in frontier intelligence, warning against using regulation to squeeze out smaller labs and open source. The path to a policy-mandated slowdown actually got weaker today.

→ Corrected characterization: this is a catalyst that is real in direction but lacks an enforcement mechanism. It is enough to support a valuation-level repricing (higher tail risk in discounted forward cash flows), but it is not currently enough to support a cut to actual 2027 capex.

But you cannot conclude this is pure collateral damage either, because there is a second, real driver unrelated to this news:

ORCL delivered a double beat after the close on 9/10 (EPS $1.92 vs $1.74 expected, RPO $664 billion), gapped up to $164.43 on 9/11, and then sold off all session to a $150.28 close, −8.6% from the open, on 80.44 million shares (roughly 3–4× recent average volume) — while the broad market was up 0.86% that day. Before the "AI slowdown" news existed, the market was already voting with its feet against the "bigger backlog is better" narrative. What exactly was being rejected (verified against the 10-Q/10-K originals, harder than my first draft's speculation): ① new bookings collapsed 69% sequentially — the sequential RPO increment fell from +$85.4 billion in May to +$26 billion in August; ② operating cash flow includes $11.363 billion of customer prepayments, which Oracle itself notes in the earnings schedules contain a "significant financing component"; excluding them, the true FCF gap is −$16.759 billion, 3.1× the −$5.396 billion shown as reported; ③ FY27 gross capex guidance jumps to $90–95 billion, while consensus EPS ($8.14) sits exactly at company guidance ($8.10) — zero margin for error. See §5 ⑧.

Corrected conclusion: today's AI-chain selloff = one catalyst that is real in direction but has no enforcement mechanism (Amodei's essay does list limiting training compute as a pacing lever) + one independent driver grounded in financial statements (ORCL's bookings and cash-flow quality, already in motion since Friday). Both legs stand up; they just differ in strength. Therefore: ① The first version's "the news was misread, don't reflexively buy the dip" is void — the market's directional reading is correct, so "collateral damage" cannot be used as a reason to buy the dip; ② But linear extrapolation is also unwarranted: the essay carries no binding force, and Trump and Sacks are pushing the other way today — the policy path is actually weakening near term; ③ What is genuinely solid, and independent of any narrative, is still the financing/leverage dimension (ORCL bookings −69% QoQ, CRWV interest exceeding operating profit). This is why in §5 I give CRWV/NBIS/ORCL different labels from VRT/LRCX — the former have real problems in their financial statements; the latter are just high valuation percentiles meeting a long-dated narrative.

⚠️ Mandatory self-correction: my first version overstated the attribution for Theme 2

After running a cross-sectional check with primary SEC data, I overturned the clean causal story in my first draft that "software up = AI disruption threat postponed". Three honest admissions:

① The cross-section is inconsistent with the narrative (the hardest piece of counter-evidence) If today were really pricing "disruption threat postponed", then the most damaged, cheapest names should bounce the most. The reality is exactly the opposite:

⚠️ Convention note: the pre-market moves in this table come from a separate, independent read timestamp and differ by 0.03–0.24pp from the 08:17 ET snapshot used everywhere else in this piece (ADBE +2.41%, ACN +2.96%, NOW +4.88%, PANW +4.96%, CRWD +5.20%). Pre-market quotes are cumulative; both sets are correct, just read at different moments, and the ranking conclusion is unaffected.

The order of gains is almost exactly the inverse of "degree of disruption victimhood", and is instead aligned with "valuation percentile". In other words: today's leaders are the most expensive, not the most wounded.

② But the alternative explanation "high multiple = high beta" also fails: expanding the sample from 5 to 18 software names and recomputing, the correlation between pre-market move and P/S is only Pearson +0.237, Spearman +0.252 (n=18) — the beautiful monotonicity in the 5-name sample is coincidence and cannot be used as a rule. (A reminder too: perfect monotonicity in a small sample is almost always coincidence.)

③ This rotation did not start today: over the past 3 months IGV +11.9%, SMH −8.3%, a 20pp spread already. Today is an acceleration day in a three-month-old trend, not a launch day. My first draft labeled it "launch phase", which was wrong; §7 has been changed to "acceleration phase".

While at it, one circulating figure got calibrated: the claim that "software has bounced ~40% off the SaaSpocalypse bottom" — IGV actually bounced from its 2026-04-10 low of 74.67 to 101.52, +36.0%, and is still 13.8% below its 52-week high.

So what is today, actually? The evidence only supports a weaker conclusion: this is a capital relocation with an identifiable funding source (sell AI compute, buy software); what software is receiving is flow-driven β, not fundamental α. "Disruption threat postponed" is a plausible but cross-sectionally unconfirmed candidate explanation. Key corroboration: none of these five had any company-level new information today — their earnings fell on 8/26 (CRWD), 7/22 (NOW), 9/1 (PANW) and 9/10 (ADBE), while ACN does not report until 10/1. And the verdicts on their own earnings went in completely different directions (the day after earnings CRWD +20.5%, ADBE +1.4%, PANW −9.3%), yet today they have all been smoothed into the same move. That by itself shows the driver is not single-stock fundamentals.

Implication for execution: since it is β rather than α, there is insufficient basis for ranking by "who benefits most from the narrative", and better basis for ranking by "valuation percentile + crowdedness" — which happens to push my ranking from CRWD/PANW toward ADBE/ACN, but also means the durability of this bounce is weaker than my first draft judged.

---

3. Single-Stock Strength Rankings

Pre-market price, percentage move and pre-market volume all come from the Nasdaq official API, snapshot at 09-14 08:17 ET. "Pre-market notional" = pre-market volume × pre-market price (in millions of dollars), used to gauge whether real money sits behind the price move.

3A. Long Side

3B. Short Side (details in §6)

---

4. Single-Stock Scoring Model (out of 100)

The table below is the revised version after verification against primary data. The three largest changes are flagged: ADBE 76→71 (catalyst directness cut: no company news today and the smallest gain in the field), NOW 73→66 (crowdedness recomputed on P/S percentile), PANW 65→52 (growth quality falsified).

---

5. Detailed Analysis of Top Names

1Definium TherapeuticsDFTX+16.38% · $45.26

• Related news: pre-market on 09-14, the company's IR primary disclosure that the Phase 3 Panorama study (GAD) met its primary endpoint: week-12 HAM-A total score improvement vs placebo of 5.1 points, p<0.0001, Cohen's d=0.64; all key secondary endpoints also met; 245 patients enrolled across roughly 32 sites. Source⁹

• 🔴 Three major corrections after second-pass review (the first version badly understated the risk):

  ① Panorama is the second Phase 3 in GAD, and it is weaker than the first. The first version wrote it up as "the third positive Phase 3 → the regulatory path is essentially cleared", a framing that inflated its marginal information content. In reality:

  	Voyage (August, first GAD Phase 3)	Panorama (today, GAD confirmatory replication)
  HAM-A improvement vs placebo	5.4 points	5.1 points
  Effect size Cohen's d	0.81 (large)	0.64 (moderate-to-large)
  p-value	p<0.0001	p<0.0001

  Table scrolls horizontally

  → The confirmatory study's efficacy is slightly below the first study's. This is a successful replication (important), but not incremental upside information — and that is the real reason it is up only 16% rather than more.

  ② DT120's active ingredient is lysergide (LSD), a US DEA Schedule I controlled substance. DEA rescheduling must be completed before marketing approval, an independent regulatory gate on top of the FDA. The first version did not mention this layer at all.

  ③ The dosing model is very heavy; this is not an ordinary oral anxiolytic. The Phase 3 protocol uses a single 100µg orally disintegrating tablet under full monitoring, with discharge assessment beginning only 8 hours after dosing and subjects required to stay at least 12 hours. That means commercialization requires an in-clinic dosing network + a reimbursement pathway, and the peak-sales model is not remotely comparable to an ordinary prescription drug.

• Theme and stage: an event-driven launch phase, and one entirely uncorrelated with today's two macro main lines — on a day when Nasdaq futures are −1.69%, this is one of the few catalysts that does not eat broad-market beta.

• Fundamental verification: market cap $5.94 billion (Nasdaq, on the pre-market price); a clinical-stage biotech with no revenue and no profit, so P/E and P/S do not apply. Cash runway and 2026Q2 financial details were not verified item by item in this piece and are left blank.

• Pre-market and technicals (independently rechecked): previous close $38.89; 52-week range $8.77–$49.70, the pre-market $45.26 sits at the 89th percentile of the range, −8.9% from the 52-week high; sell-side one-year target $66.00 (+46% vs the pre-market price). Liquidity is better than the nominal notional suggests: pre-market volume of 384,445 shares (up to 462,796 by 08:35), against a full-day average volume of only 1.93 million shares — about 24% of average daily volume has already traded before the open. For a small/mid cap, that ratio speaks to participation better than the "$17.4mm notional" does, and it constitutes genuine volume expansion.

• Final judgment: watch closely, but with materially less enthusiasm than the first version. Catalyst quality remains the highest in the field (primary company source, statistically significant, falsifiable), and it is the only name today that does not eat macro beta; but the three corrections together mean this is a "successful confirmatory replication with efficacy slightly below the first study", not an event that opens up new space — and standing between it and revenue are two structural gates: DEA rescheduling + a 12-hour in-clinic dosing model. Per the iron rules, do not chase a pre-market gap up — the 08:35 recheck shows it has already faded from +16.38% to +13.65%; gap-fill is under way (see §8).

• ⚠️ The verification point has to be replaced (the first version's was unfalsifiable): the first version said "watch the safety/tolerability details on this morning's 8:00 call", but safety topline was already in the same PR (generally well tolerated, no new safety signals, no drug-related SAEs) — the answer existed at publication, so it was unfalsifiable. Use these two genuinely falsifiable checkpoints instead: ① progress and timetable of DEA rescheduling — the only variable that could produce "all Phase 3s succeed and the drug still cannot be sold"; ② whether the company's guided pre-NDA meeting (4Q26) and NDA submission (1H27) happen on schedule. Intraday: whether it can hold the pre-market price range after the open, or keeps gap-filling back toward $38.89 (Friday's close).

2AdobeADBEverified item by item against the SEC 8-K original · +2.41% · $258.32

• Related news: theme capital rotation. ADBE already reported FY26Q3 (ended 2026-08-28) after the close on 09-10 (Thursday), so today is not pre-earnings; the next report is confirmed by Adobe's official IR as 2026-12-09, almost 3 months out.
• Actual reported numbers (left blank in the first draft, now filled in): revenue $6.760 billion (+12.9%), above its own guidance of $6.67–6.72 billion; non-GAAP EPS $6.13 (+15.4%), above guidance of $6.05–6.10; GAAP EPS $4.62; OCF $2.523 billion (a Q3 record); AI-first ARR +150% or more YoY. FY26 full-year guidance raised to revenue $26.576–26.626 billion and non-GAAP EPS $24.45–24.50. → This is a genuine beat & raise.
• ⚠️ But three deceleration data points must be stated alongside it (entirely absent from my first draft):
  1. RPO $22.16 billion, up only +8.4% YoY, below revenue growth of +12.9%, and down slightly QoQ from $22.27 billion in May. Backlog growing more slowly than current revenue means forward 12-month revenue growth most likely steps down further.
  2. Non-GAAP operating margin 44.0%, −232bp YoY (46.3% a year ago). The company says it is pushing a freemium customer-acquisition strategy — costs up front, revenue deferred.
  3. Non-GAAP net income up only +7.6% while EPS is +15.4% — roughly half the growth comes from buyback-driven share shrinkage (about 9.5 million shares repurchased in the quarter).
  4. 🔴 A fourth point added on second-pass review, and the only negative not yet digested: Q4 revenue guidance of $6.80–6.85 billion has a midpoint of $6.825 billion, below the ~$6.85 billion consensus. That, not weak Q3 numbers, is why the stock fell after the close on 9/10. It also means "earnings risk is cleared" is not entirely true. → Amodei's "pace the frontier" fixes none of these four — they are execution and competition problems, not "when will AI disrupt you" problems.
• Price evidence (OHLC I measured myself): 8/31 close $292.79 → 9/10 close $248.83. On 9/11 it opened at $242.17 with a gap down, low $241.51, then recovered all session to close at $252.23 (+1.37%) on 10.78 million shares. The gap down being bought back is a fact; but raising full-year guidance bought only +1.37%, which shows the market is already highly immune to incremental ADBE information. Also: consensus FY26 EPS $24.48 vs guidance midpoint $24.475 — essentially no expectation gap.
• Sell-side actions today (missed in the first draft): Citi maintained Hold and cut its target from $301 to $250 (below the current price); Goldman Sachs maintained Sell and raised its target from $190 to $200. Of 40 covering analysts, 23 hold, 1 sell, 4 strong sell — a neutral-to-bearish majority.
• Valuation and crowdedness (still its strongest attribute): P/S 3.86x, at its own 5-year 4% percentile (5-year median 10.30x); forward PE only 10.3x (on the FY26 guidance midpoint), P/FCF 9.47x, PEG 0.67. 52-week range $190.12–$370.86, currently the 38th percentile, −31.4% from the high, −28.1% over 12 months.
• ⚠️ One fact that contradicts the narrative: ADBE is +2.41% today, the smallest gain among the five software names. If today were really pricing "disruption threat postponed", the most typical victim should have bounced the most. This contradiction is one of the main reasons I cut my confidence in the whole Theme 2 attribution in §2.
• Final judgment: watch closely (downgraded from the first draft's "priority deep-dive"). Reason for the revision: of the first draft's "new catalyst + bad news exhausted", the "new catalyst" leg does not hold — it had no company-level news today and the smallest gain in the field. Only two legs actually hold: valuation at its 5-year 4% percentile, and no near-dated event risk (next earnings 12/9). So it is a cheap, uncrowded value name, not a catalyst-driven trading name — the reason to hold and the reason to buy are not the same thing, and the two must be kept separate.

3ServiceNowNOWverified · +4.88% · $139.00

• Related news: theme capital rotation. The earnings report is 2026Q2 (ended 6/30), released 7/22; no company-level news today.
• Fundamentals (SEC 8-K / 10-Q originals): revenue $3.987 billion (+24.0%), subscription revenue +24.5% and 150bp above the top of guidance; cRPO $13.20 billion (+21%), RPO $29 billion (+21%); non-GAAP operating margin 29.5% (flat YoY); non-GAAP EPS $0.90 (+11.1%).
• ⚠️ Three key items absent from the first draft:
  1. Growth is stepping down, and the company itself said so: Q3 subscription revenue guidance +20.5%, cRPO guidance +19.5% (Q2 actuals were +24.5% / +21% respectively).
  2. GAAP profit is flattered: of $298 million in GAAP net income, $273 million is strategic-investment revaluation gains plus $51 million of one-off tax release; excluding those, GAAP core net income is around breakeven. (The company's non-GAAP figure of $930 million already excludes that gain, so this time non-GAAP is clean and GAAP is dirty — the opposite of the usual direction.)
  3. Already swung from net cash to net debt of $1.746 billion: total debt jumped from $2.431 billion to $8.453 billion in one quarter (acquisitions such as Armis/Veza), and intangible amortization rose from 1% to 5.5% of revenue — the main reason GAAP gross margin is −680bp and GAAP operating margin fell from 11% to 4%.
• ⚠️ Crowdedness: two rulers disagree, and this is the easiest name in the piece to misjudge
  • By price position: 52-week range $81.24–$194.73, currently the 51st percentile, −31.1% from the high, −29.4% over 12 months → looks like "recovery from a low".
  • By valuation percentile: P/S 9.30x, at its own 5-year 89% percentile (5-year median 3.33x) → not low at all.
  • By recent momentum: from $91.94 on 7/23 to $132.53 on 9/11, +44.1% in 7 weeks; +29.7% over 3 months, outperforming IGV (+11.9%) by nearly 18pp. → Overall judgment: there is a significant chase component; this is not recovery from a low.
• Final judgment: watch closely (downgraded from the first draft's "priority deep-dive"). My first draft used only the single ruler of "51st percentile of the 52-week range" to call it uncrowded — that was the wrong ruler: the price position is low because it fell a lot 12 months ago, while the valuation percentile is already back at 89%.

4CrowdStrikeCRWDverified · +5.20% · $217.50

• Related news: theme capital rotation; CNBC attributed cybersecurity strength to "rising AI security concerns"; Wedbush initiated at Outperform on 9/11 (target $250).
• Fundamentals are the hardest of the five (this must be acknowledged): FY27Q2 (ended 2026-07-31) revenue $1.471 billion (+25.8%); ARR $5.84 billion (+25%), with record quarterly net new ARR of $332.8 million (+51%); RPO $10.7 billion (+48.6%), growing 23pp faster than revenue; non-GAAP operating margin 25.3% (+345bp YoY); FCF $377.4 million (company definition); net cash roughly +$4.19 billion.
• ⚠️ But three equally hard points on the other side:
  1. Valuation near historical extremes: P/S 39.3x, at its own 5-year 98% percentile — the 5-year median is only 6.05x. Forward PE 164.7x (on the FY27 guidance midpoint of $1.255), 173.4x at the pre-market price. This is not "high-growth stocks are always expensive"; it is the same company at its most expensive relative to its own history.
  2. No company-level support today: its own catalyst was fully spent — and partly given back — with the 8/26 earnings → +20.5% in a single day on 8/27; today's +5% has nothing to do with that print.
  3. Consensus is already pinned at the top of guidance: FY27 consensus revenue $6.01 billion and EPS $1.26, while the top of company guidance is $6.011 billion / $1.26 — the top of guidance has effectively become the pass mark (see "consensus is an echo of guidance").
• Position: +90.8% over 12 months, +136% off the 52-week low, and only −10.5% below the 52-week closing high. ⚠️ Note that CRWD did a 4-for-1 split after the close on 2026-07-01; all per-share figures above are on a post-split, retroactively adjusted basis.
• Correction to CNBC's attribution (keeping the first draft's view but softening it): "AI security → security spending" cannot explain Chegg, Coursera, Zoom, DocuSign, Paycom and Cognizant rising at the same time. But see §2 — my own "disruption threat postponed" explanation is equally unconfirmed by the cross-section. All that can be said now: this is flow-driven β, and both narratives are candidate explanations, neither proven.
• Final judgment: maintain watch only, and cut the score (68→62). In one sentence: the strongest fundamentals in the field, the most expensive valuation in the field, and today's reason for rising has nothing to do with its fundamentals. Within the same theme there are ADBE (4% percentile) and ACN (6% percentile) to choose from, and buying the one at the highest position is the worst execution.

5Valero EnergyVLO+1.54% · $396.44

• Related news: Saudi Petroline shutdown + diesel at an all-time high.
• Catalyst logic (the crack spread I computed myself): computing the 3-2-1 crack spread from CNBC futures prices:
  • Now: WTI $103.14, RBOB $144.28/bbl, ULSD (diesel) $212.72/bbl → 3-2-1 = $63.95/bbl
  • Friday: WTI $100.05, RBOB $138.90, ULSD $208.29 → 3-2-1 = $61.98/bbl
  • Widened +$1.97/bbl (+3.2%) today. Broken out: the gasoline crack $38.85→$41.14 (+5.9%), the diesel crack $108.24→$109.58 (+1.2%). Today it is the gasoline crack widening; the diesel crack is merely holding at an extremely high absolute level.
  • Absolute-level correction (the first version was too conservative): the 3-2-1 crack spread historically runs around $15–25/bbl, and peaked around $55–60/bbl in the 2022 energy crisis. $63.95 already exceeds the 2022 peak and sits in historically extreme territory — the first version's "far above historical norms" understated it. What refiners are earning right now is a crisis premium.
  • 🔴 But the second pass found a figure I failed to compute that weakens the ranking: the crack spread widened only 3.2% today while crude itself rose 3.1% — the "marginal" catalyst for refiners today is actually small. What is genuinely extreme is the absolute level that already existed on Friday, and that part is already reflected in VLO's 99th percentile. This directly weakens "VLO has the hardest logic" as a ranking rationale.
  • ⚠️ Another convention bias: computing the 3-2-1 off WTI overstates what VLO/MPC actually capture — the Brent-WTI spread is about $5 today, and a meaningful share of Gulf Coast refinery feedstock is Brent-linked.
• Why refiners rather than E&P: this is a pipeline/logistics disruption, not a loss of production, so the spread between crude and products benefits more directly than the oil price itself. (But the most mechanical transmission is actually tankers, see §7 — an entire line the first version missed.)
• Fundamentals: VLO's quarterly financials were not verified item by item in this piece and are left blank.
• Pre-market and technicals: 52-week range $155.29–$399.25, currently the 99th percentile, only −0.6% from the high — already near the 52-week high. Pre-market notional is only $7.7mm; participation is extremely thin.
• Final judgment: watch closely, not priority deep-dive. The hardest logic, but the most expensive position and the thinnest pre-market participation, and a single ceasefire headline could reverse it.

⑥ META (+2.15%) / ⑦ GOOGL (+1.77%) — beneficiaries of capex discipline

• Catalyst logic: this is the most easily overlooked line today. If the AI arms race decelerates, the people buying the shovels save money. META's and GOOGL's capex is discretionary, and their core business (advertising) does not make money selling AI compute, so lower capex directly improves free cash flow.

• Cross-sectional evidence (measured by me): META +2.15% and GOOGL +1.77%, while fellow mega caps MSFT +0.43%, AMZN −0.55%, AAPL +0.16% clearly lag.

• 🔴 Two QC corrections:

  ① "Largest and second-largest pre-market notional in the field" is a mechanical artifact of market cap, and has been withdrawn. GOOGL's $427mm and META's $396mm are indeed the largest in absolute terms, but divided by market cap, their participation intensity is actually the lowest:

  	Pre-market notional / market cap
  NOW	0.0556% (highest)
  CRWD	0.0453%
  ADBE	0.0397%
  META	0.0269%
  GOOGL	0.0127%
  MSFT	0.0072%
  AMZN	0.0057%
  AAPL	0.0048% (lowest)

  Table scrolls horizontally

  → Largest absolute notional ≠ most aggressive buying. Elsewhere I correctly used notional as a liquidity threshold, but here I treated it as participation intensity — those are two different things. → But one point still holds after normalization: within mega caps, META (0.0269%) and GOOGL (0.0127%) do show higher participation intensity than MSFT/AMZN/AAPL (0.0048–0.0072%), by a factor of 2–5. The divergence itself is real; it just should not be described as "largest and second largest in the field".

  ② One critique claims AAPL falsifies this logic; I think it only falsifies half of it. The critique's argument: AAPL is the mega cap with the lightest AI capex burden and sells no compute at all, so by my logic it should be the purest beneficiary, yet it is only +0.16%. My rebuttal: AAPL never had large AI capex to cut in the first place, so the "cut capex, release FCF" channel does not apply to it — its muted reaction does not conflict with the logic. But I accept the core of the critique: n=4 was cherry-picked after the fact, wording like "precisely matches the logic" is overconfident, and there is an equally self-consistent, simpler explanation — META/GOOGL are simply the two largest parking spots within the same sector after money exits semis (pure rotation) — which requires no capex narrative at all. The current data cannot distinguish between these two explanations.

• 🔴 Adding the 52-week percentiles the first version skipped (this cuts against me, but it has to be written): META's pre-market $661.08 within the $520.26–$790.80 range = 52nd percentile; GOOGL's pre-market $343.40 within the $235.84–$408.61 range = 63rd percentile. → GOOGL's 63% is above NOW's 51%, while in §7 I had just laid down the discipline that "with the same logic, pick the one at the lowest position". The first version gave no percentile at all for these two recommended slots while giving them for every other name — precisely the pattern of "the checklist gets skipped on the branch you most want to recommend", corrected on the second pass. Under that discipline, GOOGL should rank below META (52%).

• Final judgment: both are watch closely, but the attribution is labeled "unconfirmed", and GOOGL ranks after META because of its higher position. "Precisely matches the logic" has been deleted. The fragile point in this logic: the moment the AI race re-accelerates (Trump's stance points that way), the logic flips instantly.

• A test that can distinguish the two explanations (for the session): if it is "capex discipline", the gains should correlate positively with each company's AI capex as a share of revenue (a full-sample regression, not 4 cherry-picked names); if it is "rotation parking", the gains should correlate only with market cap/liquidity and not with capex.

8OracleORCLverified item by item · −4.08% · $144.15

• Related news: FY27Q1 (ended 2026-08-31) results after the close on 9/10. Revenue $19.345 billion (+29.6%), GAAP EPS $1.56, non-GAAP EPS $1.92 (vs $1.74 expected); Cloud total $11.607 billion (+62%), of which IaaS $7.405 billion (+121%) and SaaS $4.2 billion (+10%); the Software business −3%.

• Price action (you have to look at OHLC; the close alone misses it entirely): on 9/11 it opened at $164.43, hit a high of $166.00, and sold off all the way to a $150.28 close.

  • Close-to-close only −1.74% (9/10 close $152.94) — looking at the daily percentage alone, you would think "nothing happened";
  • Open-to-close −8.6%, on 80.44 million shares (3–4× the recent daily average), on a day the S&P was +0.86%.
  • This is a textbook "sell the news", and it happened before the Amodei essay (9/12).

• What exactly is the market rejecting — three reasons verified against primary sources:

  ① The sequential increment in RPO collapsed (this is the real explanatory variable for the 9/11 breakdown)

  Quarter end	RPO	Sequential increment
  2025-11-30	$523.3 billion	+$29.3 billion
  2026-02-28	$552.6 billion	+$29.3 billion
  2026-05-31	$638.0 billion	+$85.4 billion
  2026-08-31	$664.0 billion	+$26.0 billion

  Table scrolls horizontally

  The headline +$209 billion YoY looks great, but new bookings collapsed sequentially from +$85.4 billion to +$26 billion (−69%). The company says it signed "over $30 billion of AI cloud contracts" in the quarter, while net RPO grew only $26 billion.

  ② Operating cash flow is padded, and Oracle itself disclosed it in the earnings schedules

  	FY27 Q1
  Operating cash flow (as reported)	$23.103 billion
  of which: customer prepayments containing a "significant financing component"	$11.363 billion
  Capex	$28.499 billion
  FCF as reported	−$5.396 billion
  FCF excluding prepayments	−$16.759 billion (3.1× the reported figure)

  Table scrolls horizontally

  ③ The balance sheet: interest-bearing debt $125.337 billion, cash and securities $37.077 billion → net debt $88.260 billion; including operating leases of $34.621 billion → $122.881 billion. TTM interest expense $5.104 billion, with $1.428 billion in Q1 alone (+55% YoY). Share count rose from 2.880 billion to 3.024 billion in a single quarter, dilution of +5.0% (a $20 billion ATM offering completed in Q1, net proceeds $19.909 billion); there is also $4.954 billion of 6.50% Series D mandatory convertible preferred. ROE is 39.1% but ROA is only 7.4% — the high ROE comes from leverage, not asset efficiency.

• Rumor check (three publicly reported figures I cited in the first draft, each traced to a primary source): FY26 capex $55.663 billion ✅ true; "raised $43 billion of debt" ⚠️ right order of magnitude but inaccurate — FY26 gross senior long-term debt issuance was $46.093 billion; FY27 "net capex ~$70 billion" ✅ true, but that is net; gross guidance is $90–95 billion.

• Guidance and consensus: FY27 full-year revenue "at least $90 billion", non-GAAP EPS "$8.10"; consensus (33 analysts) $90.46 billion / $8.14 — consensus sits exactly on guidance, with no buffer at all. Also note Q2 guidance of non-GAAP EPS $1.85–1.93: excluding the year-ago one-off Ampere investment gain that is +21%~25%, but including that gain it is −18%~−14% (a YoY decline).

• Valuation: TTM PE 23.56x, at the lowest percentile (~5%) of its own 5-year range (22.93–51.05); on the company's own FY27 non-GAAP EPS guidance of $8.10, the pre-market price is only 17.7x. ⚠️ But the TTM PE is likewise flattered by a one-off gain (the 2025-11 quarter includes the Ampere disposal gain).

• Position: 52-week range $114.50–$329.50, currently the 14th percentile, −56.3% from the high.

• ⚠️ A new counter-signal today (missed in my first draft, found on second-pass review): an 8-K (filed 2026-09-14, event date 9/12) discloses that Larry Ellison has cancelled his 10b5-1 selling plan, under which no shares were sold. This is an insider signal that runs counter to today's decline, and a short has to price it in.

• Final judgment: avoid (short-watch conviction has been cut). The wording reduces the short bias versus the first draft: it is cheap and fragile at the same time — PE at its lowest 5-year percentile, but bookings down 69% sequentially, a true FCF gap of −$16.759 billion, and FY27 gross capex jumping to $90–95 billion with consensus offering zero margin for error. Cheap is not a buy signal, but Ellison cancelling his selling plan plus the lowest valuation percentile mean the risk/reward on a naked short is not good.

9MarvellMRVLverified, and correcting my first draft's judgment · −7.92% · $217.40

• Current fundamentals have not deteriorated at all: FY27Q2 (ended 2026-08-01) revenue $2.7393 billion (+36.5%), with data center at $2.1715 billion, 79% of the total, +46% YoY; non-GAAP EPS $0.94 (+40.3%); GAAP operating margin 16.78% (+232bp); TTM FCF $1.730 billion; net debt of only $1.030 billion, ND/EBITDA ≈0.36x. On Friday (9/11) MRVL was still +4.03%, rising against the tape.
• ⚠️ But the TTM PE is contaminated: XBRL shows GainLossOnSaleOfBusiness = $1.8304 billion, all falling in the single quarter ended 2025-11-01 (the automotive Ethernet business sold to Infineon). Excluding it, TTM core GAAP net income is only about $809.5 million → core GAAP PE ≈ 256x, and ROE drops from 16.2% to 5.0%. The meaningful measure is FY27 consensus EPS of $4.21 → 51.9x at the pre-market price.
• The real risk point: P/S 21.91, at the all-time high of its own 3-year range (5.84–17.38) — even though the share price is still −34% from the 52-week high. The denominator (revenue) cannot keep up with the past multiple expansion.
• A significant new item (8-K 2026-08-19, not elaborated in most coverage): a custom silicon agreement was signed with Google on 2026-07-29, and on 08-18 warrants for up to 58,970,907 shares at an exercise price of $206.58 (below the current price) were issued to Google, of which roughly 57.61 million shares vest against purchase volume — one tranche vesting per $500 million of custom product revenue, across 240 tranches, i.e. full vesting requires a cumulative $120 billion of custom revenue — running through FY2033. Potential dilution of 6.7%. The implication: the Custom acceleration is real and contractually backed, but the company bought that revenue curve with equity.
• Position and context: 52-week range $66.14–$329.88, currently the 58th percentile, −33.9% from the high. After the 8/27 earnings it fell from $253 to $241 (41.01 million shares traded), then based out in $205–215, only recovering to $236.10 by 9/11 (+4.03% that day, rising against the tape), and today gives back the entire recovery of the past two weeks in a single session. Pre-market notional $205.1mm — the selling pressure is real.
• Final judgment: maintain avoid, but with the reasoning revised to "extreme valuation percentile (P/S at a 3-year high) + the cost of equity dilution (the Google warrants, 6.7%)", and not "fundamentals deteriorating". The reason not to short it: unlike CRWV/ORCL it has no financing problem, and the current data-center business is still accelerating.

⑩ CRWV (−8.90%) / NBIS (−9.06%) — the most leveraged link | CRWV verified item by item

• CRWV's situation in one sentence (the cleanest formulation after verification): the company's own guidance for a single quarter of Q3 interest expense ($860–940 million) exceeds its guided full-year adjusted operating profit ($960 million–$1.150 billion).
• Q2 2026 (ended 2026-06-30): revenue $2.575 billion (+112.5%), but GAAP gross margin 65.86% (−831bp YoY), GAAP operating profit −$49 million (swung from positive to negative), adjusted operating profit $128 million (down 36% YoY in absolute terms); interest expense (net) $640 million (+140%); GAAP net loss −$626 million. → Note this combination: revenue doubled, yet operating profit went negative. "Scale brings operating leverage" holds at the adjusted-EBITDA level (+100%) but does not hold at the operating-profit level (single-quarter D&A has reached $1.393 billion).
• Debt: interest-bearing debt $35.551 billion (versus $24.859 billion at the end of March, +43% in one quarter), $52.091 billion including leases; net debt $46.552 billion; net debt/EBITDA = 8.0x (interest-bearing only) / 12.4x (including leases); shareholders' equity is only 6.5% of total assets.
• Backlog and customer concentration: RPO $103.7 billion (+246%), but the 10-Q's own recognition schedule is "41% within the first 24 months, 39% in months 25–48" — only about $42.5 billion maps to the next two years. Customer concentration: A 36% / B 26% / C 10%, top three totaling 72%; receivables from A+B total 64%. (A year ago customer A alone was 71%, so de-concentration is under way — this one is an improvement.)
• ⚠️ But most of the valuation bubble is already gone: P/S 6.47, at the lowest percentile (~10%) of its since-IPO range (6.35–23.59).
• Final judgment: short watch, with an explicit caveat that it is unsuitable for a naked short. The leverage risk is real (that is what the thesis targets), the valuation bubble is largely worked off (so the downside is not as large as the decline implies), and names like this have crowded shorts and violent squeezes.

⑪ VRT / LRCX — the two whose declines least match current evidence

• VRT (−7.63%) has the strongest current fundamentals of the five: Q2 2026 net sales $3.2743 billion (+24.1%), adjusted operating profit $738 million (+51%), margin 22.6% (+410bp), adjusted EPS $1.52 (+60%), operating cash flow +241%; net cash position, ND/EBITDA ≈0.05x, ROE 41.9% and operational in nature. FY26 guidance raised across the board: revenue $13.8–14.2 billion, adjusted EPS $6.65–6.75 (+58%~61%).
  • ⚠️ Two unfavorable details almost nobody mentions in the coverage: ① the earnings text admits Q2 had "slight revenue timing dislocation, primarily due to temporary supply chain congestion"; ② full-text searches of all three Q2 primary documents — the 8-K, the 10-Q and the official results presentation — turn up no backlog figure and no order figure, leaving only qualitative phrases like "strong backlog"; the widely circulated "$15 billion backlog" traces back, on checking, to an earlier period. Ceasing to disclose order figures in the quarter when the AI capex debate is most intense is itself information.
  • Valuation: P/S ~93rd percentile, EV/EBITDA ~90th percentile, while the PE percentile is only ~55% — the PE looks undemanding only because it assumes a 22.6% operating margin stays permanently parked at an all-time peak.
• LRCX (−7.78%) has the cleanest accounting of the five: FY26Q4 revenue $6.7222 billion (+30.0%), GAAP EPS $1.81 vs non-GAAP $1.82, a difference of one cent (no room for manipulable adjustments); net cash +$1.857 billion, ROE 64.1%; next-quarter guidance of $8.10 billion ±$400 million, +20.5% QoQ and +52.1% YoY, EPS $2.15±0.15 — what it is guiding to is acceleration.
  • ⚠️ Two things easily read backwards: ① equipment revenue is supply side, not demand side — LRCX's revenue equals chipmakers' capex, and it speaks to capacity supply 12–18 months out, not current AI demand; ② in the FY26 end-market split, foundry rose to 54% (+900bp), which the 10-K explicitly attributes to "mature-node investment", while memory's share actually fell 300bp — using LRCX as a pure proxy for AI cyclical strength is biased.
  • Valuation: PE 51.77x / P/S 16.06 / EV-EBITDA 43.02, all three simultaneously at roughly their own 5-year 93rd percentile, with PE about 2.4× the 5-year median (~21x).
• Final judgment: both maintain avoid (trend and theme are against them), but it must be labeled honestly: today's declines lack any current-fundamentals basis; this is "a high valuation percentile meeting a long-dated narrative". If the theme is falsified, these two (especially VRT, with net cash and raised guidance) have the greatest rebound elasticity. This is a completely different category from ORCL/CRWV's "real problems in the financial statements".

---

6. Bearish / Avoid List

---

7. Intra-Theme Rankings

Theme 1: money rotating from AI compute into software (long side) | Stage: acceleration phase, not launch phase

Important: crowdedness must be judged by "valuation percentile", not by "distance from the high" — the two rulers give completely opposite answers on NOW (price −31% looks low, P/S at the 89th percentile looks high). The table below ranks primarily on my own computed P/S own-5-year percentile.

The most counterintuitive row in this table: CRWD has the strongest fundamentals in the field and also the most expensive valuation in the field; PANW has the weakest fundamentals in the field and the second most expensive valuation. Yet their gains today are nearly identical (+5.20% / +4.96%) — this is the most direct evidence that "today is flow β, not fundamental α".

Theme 2: crude supply shock (long side)

🔴 Convention correction: the first version of this table showed percentile values such as COP 102% and CVX 100% — a percentile cannot mathematically exceed 100%. The cause was using the pre-market price as the numerator while the denominator was the 52-week range as of the 9/11 close; once the pre-market price exceeds the old high, it overflows. This has been rewritten as "≈100% (pre-market has exceeded the 52-week high)", which is not distorting. The other percentiles (ADBE 38%, NOW 51%, CRWD 89%, ORCL 14%, MRVL 58%, VLO 99%) were rechecked, the algorithm is consistent and the conclusions unchanged, but readers should note they too use the pre-market price as the numerator, i.e. an "including pre-market" convention.

Note the divergence in oil services: crude is +3.4% while SLB is only +0.27% and HAL +0.75%. Oil services not following shows the market characterizes this as a "short-term logistics disruption" rather than "the start of a capex cycle". That is a counter-signal on the durability of the energy theme.

🔴 Addendum: tankers/shipping — an entire line the first version missed, and its transmission is more mechanical than the crack spread

Added on second-pass review. The logic: Petroline is precisely the onshore pipeline that bypasses Hormuz; with it shut, that crude has to be rerouted by sea → ton-mile demand rises mechanically → tanker rates go up. This transmission chain does not depend on any narrative and is more direct than the refinery crack spread I led with.

Two things that must be said together:

1. The logic is the most mechanical, and it is not vulnerable to a ceasefire headline — even with a ceasefire, rerouting demand takes weeks to normalize.
2. But they are the group in this whole list that was "found latest, sits highest and has the thinnest participation": they are already up 6~14% over the past week, DHT/FRO/TNK are above their 52-week highs pre-market, and TNK has traded only 1,082 shares pre-market, INSW 776 — essentially no quote at all. → Conclusion: this is a real main line that my first version missed, but it was found too late. Characterized as watch closely; a good logic is not a reason to upgrade it into a chase. This is also one lesson from this piece: in §1 item 5 I wrote "Themes: energy/shipping" myself, yet gave not a single shipping name in §2/§3/§7 — the checklist was written but not executed.

Theme 3: the AI capex chain (short/avoid side, descending by fragility)

---

8. Opening Verification Signals

Pre-market (measured by me, usable directly as an opening benchmark)

1. The breadth is overwhelming, not idiosyncratic: of the 45 AI hardware/compute/power names I pulled individually, 0 advanced, with a mean of −5.88%; while of a 26-name "disrupted by AI" software sample, 25 advanced, with a mean of +2.04%. That symmetry of 0/45 against 25/26 shows this is theme-level repositioning, not collateral damage from a headline.
2. But the money on the two sides is not equal — this is the thing most worth watching today:
   • Sell-side notional: NVDA $1,093mm, MU $969mm, AMD $307mm, INTC $268mm, ORCL $265mm, AVGO $229mm, MRVL $205mm, DELL $201mm, NBIS $194mm
   • Buy-side notional: GOOGL $427mm, META $396mm, CRWD $101mm, PANW $89mm, NOW $80mm, ADBE $41mm
   • Energy names' pre-market notional is small enough to ignore: XOM $13.9mm, CVX $10.9mm, VLO $7.7mm, MPC $3.2mm, COP $2.5mm, EOG $0.5mm. → Implication: the selling in the AI chain is heavy-position repositioning; the buying in software is a light-position probe; energy's +1.5~1.8% has almost no pre-market participation and will very likely be repriced after the open (in either direction). Do not treat the pre-market energy move as a confirmed sector consensus.
3. Pre-market timing structure: the decliners (MRVL/LRCX/VRT/INTC/AMD/ORCL/CRWV) almost all made their pre-market lows at 08:03–08:06 ET, while the advancers (CRWD/PANW/NOW) made their highs at 07:07–07:53 ET. That is, selling pressure is strengthening into the open while long momentum is fading. If that structure reverses within the first 30 minutes, it is the first sign of short exhaustion.
4. NVDA's relative strength is the thermometer for the whole chain: NVDA is −2.48%, notably more resilient than SMH (−4.77%) and the chain-wide mean (−5.88%), and it absorbed the largest pre-market notional in the field. If NVDA gives way and drifts toward −5%, the selling has escalated from "sell the periphery" to "sell the core" and the sector has a second leg; if NVDA recovers to within −1%, the oversold peripheral names will bounce violently.

08:35 ET recheck (pre-market quotes are cumulative; a single snapshot is not a fact)

18 minutes after the 08:17 snapshot I pulled the data again, to test the "selling pressure strengthening into the open" judgment above:

Three readings usable directly at the open:

1. DFTX has already faded from +16.4% to +13.7% — validating my "do not chase the pre-market high" judgment; gap-fill is in progress.
2. Every advancer faded slightly, with energy fading most (XOM +1.61%→+1.27%), consistent with the direction of my note in §8 that "energy's pre-market participation is extremely thin and it will very likely be repriced after the open".
3. NVDA's pre-market volume doubled in 18 minutes while the price barely moved — that means large-size absorption. This is the single most important thermometer reading for the whole chain: someone is taking delivery at this price.

Intraday

• First 30 minutes: watch the relative move of SMH vs IGV. Pre-market today SMH is −4.77% and IGV +2.00%; if that scissors gap keeps widening after the open, the theme trade is adding; if it converges quickly, it was an emotional open.
• Volume confirmation: if the AI-chain selloff is not accompanied by volume expansion, it is most likely market-maker pre-market pricing rather than genuine turnover, and prone to an afternoon bounce.
• Energy: watch the divergence between XLE and WTI. If WTI holds +3% while XLE cannot manage +1%, the equity market does not believe $108 is sustainable.

Options Sentiment

• VIX 17.61 (+11.17%); the absolute level is still low (it has not broken 20), which says this is repositioning within sectors rather than systemic panic.
• There are no major earnings today, so there is no IV crush risk; but IV will stay elevated into Wednesday's FOMC, which is unfavorable for buying options to express direction.

🔴 This week's event calendar (the first version wrote only the FOMC; completed after QC flagged the omissions)

The first version flagged only Wednesday's FOMC (mentioned in four places, all accurate) but left the entire macro calendar blank — the stated reason for leaving things blank was "conflicting sources on the earnings calendar", and the macro calendar should not have been left blank along with it.

Risks

• Reversal after the gap: DFTX (+16.4%), CRWD (+5.2%) and PANW (+5.0%) have large pre-market gains, so gap-fill risk exceeds gap-and-go.
• Lone-name risk: the energy sector's gains are uniform but pre-market participation is extremely thin — this is "quotes rising", not "money rising".
• Divergence between index futures and Treasuries: equities are falling, yet Treasuries are not rallying and gold is actually down. That means the market is not pricing today as a "risk-off day" but as a "rate hike + growth narrative double kill day". If Wednesday's FOMC is unexpectedly dovish, the long-duration growth names hit hardest today will also bounce hardest.
• Headline risk (two-way, and both are near): ① a restart of the Iran–Gulf talks or a Petroline restart → oil and energy equities give it all back instantly; ② further escalation of the Trump/Sacks opposition → the AI slowdown narrative is falsified, and today's biggest losers become tomorrow's biggest gainers.

---

9. Final Conclusions

Note: not a single name in this piece is labeled "priority deep-dive"; that is a deliberate outcome of verification, not an oversight. The first draft gave that label to ADBE and NOW, and after verification against primary SEC data both were downgraded (reasons in §5). On a day with Nasdaq futures −1.69%, the main-line attribution itself in doubt, and a dot-plot FOMC on Wednesday, no single name simultaneously satisfies all three of "hard catalyst + low position + no near-dated event risk". Better to downgrade everything by one notch than to force a top rating.

① The 5 Names Most Worth Watching Today

② The 3 Strongest Themes Today

③ What to Avoid Today, and Why

1. The leveraged AI data-center chain (CRWV / NBIS / ORCL) — what drives them is financing cost, bookings quality and customer concentration; none of those repairs itself just because the news flow reverses, they are problems at the financial-statement level.
2. Semicap (LRCX / AMAT / KLAC) — already −37%~−46% from 52-week highs and still down 7%+; this is continued liquidation in an already broken sector, not the location for an oversold bounce.
3. The highest valuation percentiles among today's gainers (CRWD P/S 98th percentile / PANW 96th percentile / CVX at a fresh 52-week high) — a special note: PANW's revenue +34% comes entirely from the CyberArk acquisition, non-GAAP EPS is only +7.4%, and the market already voted on that print with −9.3% on 9/2; today's +5% is sector β repairing that decline. Within the same theme there are ADBE (4% percentile) and ACN (6% percentile) to choose from.
4. Airlines (AAL/DAL/UAL) — directly pressured on the jet fuel cost side, and the oil move is not over.
5. Anything gapping more than 5% pre-market — iron rule: do not chase a pre-market gap up; gap-fill is more likely than gap-and-go.

④ Final One-Sentence Judgment

Today is a large-scale relocation of capital, not a systemic risk event (VIX only 17.6, Treasuries not rallying, gold actually down): money is moving from "the sellers of AI shovels" into "software" and "energy". But two things have to be stated clearly: First, the case on the short side is more solid than on the long side. Amodei's essay does list "limiting training compute" as a pacing lever worth considering (my first version's claim that it "did not ask anyone to buy fewer shovels" was a misreading and has been corrected), so the market's directional reading is right — it is just that the constraint has no binding force and no timetable, and Trump and the White House's Sacks are pushing the other way today. The negative that genuinely depends on no narrative is ORCL's own financial statements: new bookings collapsing 69% sequentially (+$85.4 billion → +$26 billion), and a true FCF gap of −$16.759 billion excluding customer prepayments, plus CRWV's "single-quarter guided interest expense exceeding full-year guided operating profit". So the risk is concentrated in the most leveraged builders (ORCL/CRWV/NBIS), not in chip demand itself; names like VRT and LRCX, with net cash and raised guidance, have declines today that lack any current-fundamentals basis. Second, I have cut my confidence on the long side. The cross-sectional check overturned the clean story of "software up = disruption threat postponed": the biggest gainer is the most expensive (CRWD at the 98th percentile), not the most wounded (ADBE at the 4th percentile has the smallest gain), and this rotation has been running for 3 months. So on the software side stock selection should be done on valuation percentile and crowdedness (ADBE/ACN), not on narrative (CRWD/PANW), and one has to accept it may be β rather than α. The only independent opportunity that depends on none of the above macro judgments is DFTX's Phase 3 data. Ahead of Wednesday's FOMC (~85.5% hike probability, with the dot plot) and Friday's quadruple witching, event risk on positions in either direction is significantly higher than usual.

---

⚠️ Data-sourcing and failure log for this piece (not for client display):

Data pipeline

• yfinance was not called on this machine; single-stock pre-market price/previous close/pre-market volume all went through the Nasdaq official api.nasdaq.com/api/quote/{sym}/extended-trading (verified working, and the only channel that also provides pre-market volume); futures/commodities/Treasuries/VIX went through the CNBC quote API; Treasury closing yields went through the Treasury CSV (measured 9/11 closes of 10Y 4.96 / 2Y 4.63 / 30Y 5.35, cross-validated as consistent with CNBC's real-time 4.977/4.634/5.359).
• CNBC's Treasury change_pct field is, as usual, unusable (returning meaningless values like +0.0402%); this piece uses only absolute yield levels and does not cite its percentage changes.

Pitfalls encountered

1. CNBC article body 403 + lazy loading: WebFetch gets a straight 403 on cnbc.com; switching to curl with a browser UA returned HTTP 200 (773KB), but window.__s_data does not contain the article body (articleBody is lazy-loaded), so only the headline was obtained. The single-stock list in CNBC's pre-market movers piece was therefore obtained via search snippets, and all specific percentage moves were re-measured through the Nasdaq API; the article's numbers were not adopted.
2. The Nasdaq info and extended-trading endpoints contradict each other: info at one point returned −1.20% for ACN, while extended-trading at the same moment showed +2.85% (previous close $183.9 / pre-market $189.15 / volume 55,516). extended-trading was confirmed authoritative and used uniformly throughout. This is a silent trap: using info alone would produce a conclusion pointing in exactly the opposite direction.
3. Inconsistent capitalization in the 52-week range field: the info endpoint uses keyStats.fiftyTwoWeekHighLow (lowercase f) and returns null for most tickers; the summary endpoint uses summaryData.FiftTwoWeekHighLow (note FiftTwo, missing a y). The first version of the script therefore returned ERR across the whole table; it was switched to the summary endpoint.
4. stockanalysis's pre-market leaderboard is usable but not trustworthy: /markets/premarket/gainers|losers returns a parseable HTTP 200, but its percentage baseline has known problems; this piece used it only for breadth discovery (finding names outside my own universe such as NOK, DFTX, AEHR, SLS), and every number that made it into the piece was rechecked against the Nasdaq API. NOK rechecked at −9.61% (previous close $11.13), close to that site's −9.43% and directionally consistent.
5. stocktitan.net DNS resolution timed out (ETIMEOUT), so DFTX's Phase 3 data was taken from Definium IR and BioSpace instead; the primary PR link was obtained.
6. Search engines are broadly lagging on same-day 9/14 single-stock news: queries for "today's" MRVL/NOK news returned only content from 9/13 and earlier, and one MRVL snippet returned the "Sunday, September 13" intraday high/low as if it were trading-day data; this was identified as quote-page noise and discarded.

Two fundamentals-analyst subagents have returned, and this piece was substantively revised on that basis

• Both subagents reported that yfinance / Yahoo query1 and query2 returned 429 all day today (IP-level rate limiting, not a leftover process), and both switched to SEC EDGAR XBRL companyfacts + the 8-K/10-Q/10-K originals, with valuation and consensus from stockanalysis. Every quarterly financial figure in this piece ultimately comes from primary SEC filings, not from yfinance.
• First-draft conclusions overturned or corrected by the subagents (this must be recorded; it is the largest set of changes in this piece):
  1. The software theme's attribution was overturned: my first draft said "AI disruption threat postponed" was the main cause and believed it had cross-sectional support. The subagent pointed out that the ranking of gains is the inverse of "degree of disruption victimhood" (the most victimized and cheapest, ADBE, rose least; CRWD, which is not in that narrative at all, rose most), and that this rotation has been running 3 months (IGV +11.9% vs SMH −8.3%). A full "mandatory self-correction" section has been added to §2, the theme strength was cut from A+ to A, the logic robustness from "moderate-to-strong" to "weak-to-moderate", and the stage from "launch phase" to "acceleration phase".
  2. ADBE cut from "priority deep-dive" to "watch closely": the first draft said "new catalyst + bad news exhausted", when in fact there was no company news today and it had the smallest gain in the field; three deceleration data points were added — RPO +8.4% < revenue +12.9%, operating margin −232bp, half of EPS growth from buybacks.
  3. NOW cut from "priority deep-dive" to "watch closely": the first draft used the single ruler of "51st percentile of the 52-week range" to call it uncrowded; the subagent used a self-computed P/S percentile (89%) to show that was the wrong ruler.
  4. PANW score 65→52: revenue +34% comes entirely from the CyberArk acquisition, non-GAAP EPS is only +7.4%, and RPO guidance is down sequentially. The first draft did not find this layer at all.
  5. ORCL's true cause was replaced: the first draft attributed it to "capex + OpenAI customer concentration"; the subagent found a harder one — the sequential RPO increment collapsing from +$85.4 billion to +$26 billion (−69%) — plus operating cash flow containing $11.363 billion of customer prepayments (which Oracle itself discloses contain a "significant financing component"), leaving a true FCF gap 3.1× the reported figure once excluded. It also found today's 8-K: Ellison cancelling his 10b5-1 selling plan, a signal running counter to the decline that the first draft missed, and short conviction was cut accordingly.
  6. VRT/LRCX changed from plain "avoid" to "avoid, but flagged that the decline lacks a current basis": both have net cash and raised guidance; VRT has stopped disclosing backlog/order figures as of this quarter (confirmed by the subagent via full-text searches of the 8-K, 10-Q and presentation), a "silence-type signal" entirely absent from the first draft.
  7. MRVL's reason for avoidance was replaced: from "fundamentals deteriorating" to "extreme valuation percentile + equity dilution" — its TTM PE of 79x is contaminated by a $1.83 billion business-disposal gain (core PE about 256x), its P/S is at a 3-year high, and there is 6.7% potential dilution from the Google warrants.
• Intermediate conclusions the subagents overturned themselves (evidence of their reliability): the software agent's first pass computed valuation percentiles using "pre-split share count × post-split price"; after correction, the five names' current P/S figures differ from stockanalysis's independent numbers by <1.5%; it also proactively falsified the hypothesis that "pre-market gains rank monotonically with P/S = high multiple, high β" (expanding to 18 names gives a Pearson of only +0.237), and that hypothesis was not written into the body. → A convention note while we are here: CRWD did a 4-for-1 split after the close on 2026-07-01, and NOW did a 5-for-1 on 2025-12-17; all per-share figures in this piece are on a post-split, retroactively adjusted basis.

Still not obtained / deliberately left blank

• DFTX's financials (cash runway, 2026Q2 detail) were not verified item by item, and the body explicitly says so. Neither fundamentals agent had it on their target list (I only found it during the breadth scan), and there was no time for another round.
• This week's earnings calendar has conflicting sources (one says Lennar is on "Tuesday 9/16", but 9/16 is a Wednesday; another says FedEx/Lennar/Darden are on Thursday 9/17); this piece adopted neither version and the body gives no specific earnings dates, saying only "no major earnings today".
• The release timestamps of today's two ACN downgrades (Wells Fargo / Morgan Stanley) come from stockanalysis's secondary annotation; the subagent flagged that "recurring columns' date labels have repeatedly recorded prior-trading-day content as same-day", and the brokers' original notes were not used for verification. They are retained in the body but labeled as sourced from an aggregator. risk-auditor QC has returned with 14 red flags; this piece accepted 13 and partially rejected 1

Accepted and already rewritten (by severity):

1. R1 [fatal, rewritten] Misreading of the Amodei original. My first version asserted "the essay does not call for cutting compute"; QC overturned it with the primary text. Line-by-line rechecking confirmed I had missed it: the essay contains both "limits on the rate of unchecked AI progress" and "limiting the ingredients that go into frontier models, such as training compute". The first time I read the original I used a summarizing fetch (WebFetch's small-model summary), which dropped both sentences, and I did not go back to the original to verify keywords. The core argument in §2 and the concluding sentence in §9 were rewritten wholesale. This is the most serious error in this piece, and a textbook instance of "self-correction becomes an exemption badge" — I thought I was safe because I had read a primary source, when what I actually read was a secondary summary.
2. R8 NOK attribution error: pulling daily bars confirms NOK was +12.08% from 9/4 to 9/11 (with Google's EUR 13 billion Finland investment on 9/9), so today's −9.3% is mainly a giveback, not the theme receding. It has been removed from the theme's representative names.
3. R12 the entire tanker/shipping line was missing: the pre-market moves, 52-week positions and last-week gains for FRO/TNK/DHT/STNG/INSW have been added. In §1 I wrote "Themes: energy/shipping" myself and then did not execute on it — "the checklist was written but not executed".
4. R4 rating release dates: QC confirmed that Wells Fargo's ACN downgrade was actually on Friday 9/11, and that the $194 target already existed in July (only the rating changed this time). The news grade has been cut from B+ to C with a warning box added; the release timestamps for IREN/HPE/GEV are still unverified and flagged as such.
5. R9 unfalsifiable verification points: ADBE's test required a 2.4% decline before triggering (while gap-fill, the money-losing case, would count as "passing"); DFTX's safety answer was already in the same PR; NOW's test did not test NOW at all. All three were replaced.
6. R7 percentile >100%: COP 102% / CVX 100% exposed the convention error of "pre-market price as numerator, closing range as denominator"; the phrasing has been changed, and the META (52%) / GOOGL (63%) percentiles the first version had skipped were added — GOOGL's 63% is above NOW's 51%, conflicting with my own "pick the lowest position" discipline.
7. R6 notional not normalized: divided by market cap, GOOGL's participation intensity of 0.0127% is actually on the low side among mega caps; "largest and second largest in the field" is a market-cap artifact, withdrawn, with a normalized table added.
8. R11 macro characterization: Treasuries moved ±0.5bp today, and the rate-hike repricing happened on Friday (2Y +7bp). "An inflation/real-rate shock" has been changed to "gold and the dollar are pricing the hike; the bond market has not confirmed it today".
9. R13 calendar omissions: 9/15 retail sales and the 9/18 quarterly quadruple witching have been added.
10. R3 ADBE Q4 guidance: the Q4 revenue guidance midpoint being below consensus — the only undigested negative — has been added.
11. R2 ORCL baseline: both figures, "close-to-close −1.74% / open-to-close −8.6%", are now given side by side in the body.
12. R10 sample selection bias: acknowledging that the 25/26 basket was picked after the fact on the basis of "having fallen", so on a morning of broad software gains it is guaranteed to be all green and cannot distinguish between the two explanations. Logic robustness has been downgraded.
13. Three major DFTX omissions (found by me separately after QC flagged them): the active ingredient is lysergide (LSD), DEA Schedule I, with rescheduling required before marketing; subjects must stay at least 12 hours; and Panorama is the second Phase 3 in GAD, with efficacy of 5.1 points/d=0.64 below the first study Voyage's 5.4 points/d=0.81. The first version's "third positive Phase 3 → the path is cleared" inflated the marginal information content; corrected.

Partially rejected, 1 item:

• R5 claims AAPL falsifies the META/GOOGL capex logic — I accept only half. AAPL never had large AI capex to cut, so the channel does not apply to it and its muted reaction does not conflict with the logic. But QC's criticism that n=4 was cherry-picked after the fact and that "precisely matches the logic" is overconfident wording is valid, and "rotation parking" is an equally self-consistent, simpler explanation; the attribution has been downgraded to "unconfirmed" and a test that can distinguish the two has been given.

Parts QC recomputed, passed and confirmed correct (unchanged): the entire 3-2-1 crack spread chain (the ×42 conversion, $63.95, $61.98, +3.2%, the gasoline/diesel components), ORCL's OHLC and the 80.44 million shares and −8.605% arithmetic, the futures symbols (avoiding the @ES.1 null trap), VIX 17.60/15.84, the .SPX/.IXIC closes, the Treasury CSV cross-validation, the 52-week ranges for CRWD/NOW/ADBE, and the FOMC on 9/16 with 85.5%.

• QC also supplied two corrections that strengthen this piece, both adopted: ① the 3-2-1 at $63.95 already exceeds the 2022 peak ($55–60), and the first version's "far above the norm" understated it; ② the crack spread widened only 3.2% today while crude rose 3.1%, so refiners' "marginal" catalyst is small, which weakens the ranking rationale of "VLO has the hardest logic".

Still unverified (left for the next shift): the original release timestamps of the IREN/HPE/GEV ratings; NOKIA.HE's (the Helsinki primary listing) move today, to cross-validate the NOK ADR decline; ADBE's Q4 EPS guidance and Digital Media net-new ARR.

---

⚠️ Risk disclaimer: this list is pre-market information gathering and observation only and does not constitute investment advice. US equities carry high volatility and pre-market gap risk, and post-earnings IV crush and guidance reversals are possible; automatically generated content may contain stale information or factual errors — rely on company disclosures/SEC filings, and do not use this directly as a basis for trading.