US Pre-Market Brief | 2026-09-02 (ET) Wednesday

Coverage window: from the 2026-09-01 16:00 ET regular-session close → 2026-09-02 10:13 ET. Read-timestamp note (important): all prices and percentage moves in this piece are read uniformly at 10:13 ET on 09-02, i.e. roughly 43 minutes after the regular session opened. This piece is therefore not a pure "pre-market snapshot" but "overnight + pre-market news + the actual verification from the first 43 minutes of the regular session." For every stock discussed, this piece gives both the opening gap and the open→current drift in two separate columns — the latter is the single most important piece of information today. Price sources: stockanalysis.com quote API (regular-session basis); Treasuries cross-checked from CNBC real-time plus the official US Treasury Daily Treasury Yield Curve CSV; crude oil / gold from CNBC futures.

---

0. One-Sentence Summary of the Day

1. Today's strongest catalyst is not any single piece of good news but a shared phenomenon: three companies (MDB, PANW, CRDO) all beat, and all were sold off. MongoDB beat on revenue by 5.0%, beat on EPS by 18%, and raised full-year guidance → −12.58%; Palo Alto beat and raised guidance → −8.24%; Credo beat on revenue, beat on EPS and guided next quarter above consensus → −17.57%. The same day, GitLab beat → +14.35%. All four beat, yet the outcomes were exactly opposite.
2. So today's main structure is neither "a beat means you go up" nor the "hardware up, software down" framing presupposed by yesterday's recap — Credo is AI networking hardware and was today's biggest large-cap decliner of the entire session. What actually sets direction is "whether the magnitude of the beat is heavy enough relative to the volatility priced in beforehand and the burden of proof the narrative requires," not the beat itself. CRDO delivered a revenue beat of 1.8% and an EPS beat of 2.6%, while its pre-earnings options-implied move was ±13.1% — the market paid for 13% of information content and received 1.8%.
3. Flow direction: out of mega-cap tech, into small caps, metals and healthcare. IWM +0.89% vs QQQ −0.03%; GDX +3.92%, XME +3.27%, XLV +1.42%, XLC +1.79% (META +3.56% contributing), while XLK −0.51%, XLU −0.92%, XLRE −0.84%.
4. Driver type: earnings·guidance dominant (largest weight), macro secondary, geopolitics receding. The macro item is ADP August private payrolls +38,000, far below the expected 47,000 and the weakest since January; on geopolitics, oil pulled back today (WTI $89.50, Brent $94.21) — yesterday's US-Iran supply-shock trade is cooling off today.
5. Pre-market / opening state: S&P 7,654.92 (+0.31%), Nasdaq Composite 26,142 (+0.16%), Dow 53,083 (+0.60%), VIX 15.64 (−0.70); 10Y 4.788% (−0.8bp), 2Y 4.383% (−1.1bp); dollar UUP −0.32%, gold GLD +1.35%. The indices are calm, but a violent structural reallocation is happening underneath them.

The single most important pending event today: Broadcom (AVGO) reports FY2026 Q3 after the close tonight, 09-02, with an options-implied move of about ±8.7%. It is the arbiter of the question "is AI hardware demand really decelerating," and Credo's crash today has just put that question on the table.

---

1. News Overview

---

2. Strongest Themes, Descending

---

3. Master Single-Stock Strength Board (split by direction, descending within each block)

How to read: "Gap %" = opening price vs previous close; "Open→now %" = current price vs today's opening price. The latter is today's most critical column — it separates "kept being bought after the gap" from "sold after the gap."

3A Bullish Direction

3B Bearish Direction

---

4. Single-Stock Scoring Model (100 points total)

Scoring note: GTLB has today's strongest news profile (the only large gainer driven by its own earnings), but it also carries the heaviest risk deduction — it has already retreated from +22.42% to +14.35%, and it has three flaws the market has not priced. "News strength" and "current investability" are two different things, and this table deliberately keeps them separate.

---

5. Detailed Analysis of Top Names

① AVGO (Broadcom) — tonight's arbiter, and today's only AI mega-cap whose expectations were not raised

• Related news: reports FY2026 Q3 after the close on 09-02. Consensus revenue about $29.43–29.47 billion (about +84.7% YoY), non-GAAP EPS about $3.21–3.22 (about +90.5% YoY). Last quarter the company guided Q3 AI semiconductor revenue to about $16 billion, growth of more than 200% YoY. Options-implied move ±8.7% (10-year average earnings-day move ±7.4%, median ±5.8%). Sources: Investing.com¹⁰, earnings-watcher¹¹
• Definition warning: another source gives "GAAP EPS $2.55". $3.22 is non-GAAP, $2.55 is GAAP; the two must not be mixed; when judging whether it "beat," you must align with the non-GAAP consensus.
• Catalyst logic: what the market wants is not "above consensus" but the slope of next quarter's AI revenue. Credo's lesson today is precisely this: beat the current quarter, but guide next quarter down sequentially, and you still fall −17.57%.
• Theme and stage: the re-pricing stage after the tide went out. AVGO trades at $367.68,52-week range $287.17–$495.00, only at the 40% position, about 25.7% below its 52-week high. In August it was sold off on BofA flagging a $370 billion AI-financing vehicle and on the VMware vCenter vulnerability (CVE-2026-59310).
• This is today's most important differentiating observation: MDB (85th percentile), PANW (86th percentile) and OKTA (92nd percentile) all entered earnings at highs and were killed; AVGO enters at the 40th percentile, so its pricing position is much closer to GTLB, which rose today, than to MDB, which was killed. That does not guarantee it goes up, but it does mean the "a beat doesn't work" mechanism applies markedly more weakly to AVGO.
• Fundamental verification: see the "sub-agent verification" subsection below.
• Final judgment: priority deep-dive, but no position before the event. An implied move of ±8.7% cannot be managed with intraday stops. If you participate, it can only be through position sizing or options-defined risk, not stop orders.

② CRDO (Credo) — today's most important lesson: what was killed is not demand, it is "the shift in the burden of proof"

This section was fully re-checked by the fundamentals-analyst sub-agent against primary SEC filings, which overturned this piece's original core thesis; it has been rewritten against the primary source. The correction process is recorded in the internal note at the end.

• Related news (primary source: SEC 8-K EX-99.1²): 09-01 16:05 ET, FY2027 Q1 revenue $479.003 million (+114.7% YoY, +9.6% QoQ, about 1.8% above consensus); non-GAAP diluted EPS $1.20 (about 2.6% above); GAAP diluted EPS $0.67.
• ⚠️ A widely circulated error that must be corrected (this piece's first draft also believed it): multiple secondary reports said Credo "guided next quarter to $465–475 million" and used that to explain the crash as "a sequential guidance decline." That is wrong. The 8-K's actual Q2 FY2027 guidance is "between $525 million and $535 million" — midpoint $530 million, +10.6% QoQ, and about 2.6% above the consensus of $516.5 million. $465–475 million is the Q1 guidance Credo itself gave last quarter, i.e. the number it just exceeded, not next quarter's guidance. At the same time the full-year guidance moved from "growth >80%" to ">85%" (⚠️ the CFO's wording was "We continue to expect," which reads more like a reiteration than a raise; this remains open).
• So: the explanation "it crashed because guidance was cut" does not hold. The real reasons are the three below.

(a) The size of the beat cannot carry the volatility priced in beforehand. The pre-earnings options-implied move was about ±13.1%, and the delivered beat was 1.8% on revenue and 2.6% on EPS. For a stock whose TTM P/S once reached 33x, that is a "weak beat."

(b) The slope implied by the full-year guidance and the quarterly guidance are inconsistent — this is the hardest contradiction, and it is calculated, not read off the page:

That is: the company has loaded 59% of full-year revenue onto a second half that has not started, while the actual sequential slope in the first half (+9.6%, +10.6%) is only half the slope the target requires. And when management was asked directly in Q&A for the first-half/second-half split between AEC and optical, they gave only a qualitative answer with no numbers, and conceded that growth in the core AEC business will be slower than optical. The company also did not disclose Q1 optical revenue, so the ">$600 million" full-year optical target cannot be verified for progress.

(c) A margin step-down appears at the GAAP level, sourced from an acquisition that has just been booked. GAAP gross margin 68.2% → 64.5% (−3.7pp), GAAP operating profit −22.5% QoQ, GAAP EPS −23.9% QoQ, GAAP operating expenses +32.5% QoQ. The root cause is the DustPhotonics acquisition ($750 million cash + about 920,000 shares) completed in April: goodwill rose from $93 million to $986 million, intangibles from $29 million to $379 million, and acquisition intangible amortization booked into COGS dragged gross margin by 2.3pp. Non-GAAP gross margin was 68.0%, essentially flat — gross profit did not materially deteriorate, but goodwill + intangibles now account for 45.3% of total assets.

• Two hypotheses the market can rule out (with evidence):
  • It is not "order timing at a single large customer": customer concentration is in fact improving systematically — the largest customer's share went FY25 67% → FY26 49% → this quarter 33%; customers above 10% went from 2 to 4 (33%/28%/13%/10%); nowhere in the full call transcript is there any mention of a customer pause, program transition or destocking.
  • It is not "AI networking demand has peaked": next-quarter guidance is above consensus, full-year guidance was not cut, all three optical product-line targets were reiterated, and the company is actively building inventory for the second half (inventory +24.8% QoQ, the CFO's own words: "we are leaning in from a supply chain standpoint").
• Cross-sectional evidence (same moment, 10:12 ET, decisive): SMH +0.32%, NVDA +2.27%, ALAB −2.09%, MRVL −2.41%, ANET −1.67%, AVGO −0.57%. If the market read CRDO as "AI networking demand has peaked," the semiconductor ETF could not be green. CRDO's excess decline vs SMH (−18.2pp) is 7.5x that of the closest comparable, ALAB (−2.4pp) — this is a textbook idiosyncratic single-stock event.
• The decline did not start today: 8/17 close $282.82 → 9/1 close $206.63 (−26.9%, of which the single-day −13.03% on 8/18 was a rate-driven broad sector selloff); the initial after-hours read post-earnings was only about −4.8%, and the rest of the decline happened during the conference call and after the next day's open — the timestamps themselves point to "no second-half split could be given in Q&A," not to any single number in the press release. Cumulatively −40.0% since 8/17.
• Valuation (computed by the sub-agent, at about $170): TTM P/S 19.9x, non-GAAP P/E 41.4x, forward P/S 12.8x; net cash about $743.5 million with no interest-bearing debt. ⚠️ The percentile must be disclosed on both bases: on the full sample since IPO it is the 47th percentile, but on a trailing 1-year window it is the 6th percentile. And for a company whose TTM revenue is still doubling, the rolling P/S percentile declines mechanically (the denominator inflates automatically every quarter), so a low percentile is a mathematical artifact, not evidence of cheapness — this is the mirror image of the "low PE trap."
• Final judgment: avoid, but the reasoning must be rewritten. It is not "demand has peaked"; it is that the entire valuation margin of safety is staked on a second-half ramp that has not started, has not been quantified, and is carried by optical assets just bought for $1.24 billion. It is still accelerating downward after the open (−9.71%), with no sign of a bottom. At the same time it must be made explicit: it should not be used as a bearish case for the AI networking sector — the cross-section has already vetoed that reading.

③ MDB (MongoDB) — beat + raised guidance + still −12.58%

• Related news: after the close on 09-01, Q2 revenue $771.8 million (+30% YoY, consensus $735.16 million), EPS $1.90 (consensus $1.61), full-year revenue raised to >$2.96 billion, full-year EPS raised to >$6.14. Source: StockTitan¹
• The real reason it was sold (must be stated): Atlas (the core growth engine) grew about 29% YoY, and has now been stuck at that level for 5 consecutive quarters — the company beat on "total volume," it did not beat on "acceleration." The Seeking Alpha analyst note¹² is literally headlined "high-quality Q2, but 29% Atlas growth disappoints investors."
• Pricing position: previous close $434.21, 52-week range $215.68–$473.10, entered earnings at the 85th percentile.
• Pre-market and technicals: opened at $379.25 (gap −12.66%), currently $379.57, open→now +0.08%, the only earnings stock today that has stopped falling and stabilized.
• A public data error that must be corrected: several news summaries state MDB's previous close as $453.37 and compute a "16.9% decline" from it. Reverse-computing from the pre-market price falsifies it immediately: $381.43 ÷ (1−12.16%) = $434.2, consistent with the previous close of $434.21 read independently in this piece. $453.37 is the wrong baseline, and every decline calculation citing it is overstated.
• Final judgment: avoid, but downgrade it to a candidate for "watch only" — it is the only one that stopped falling after the open, which suggests the selling pressure may have been fully released in the after-hours session. See §8 for the verification point.

④ GTLB (GitLab) — today's only winner, but none of the three flaws are priced

• Related news: after the close on 09-01, Q2 revenue $286.25 million (+21% YoY), non-GAAP diluted EPS $0.24 (consensus $0.18, a 33% beat), NRR 117%, 1,571 customers with $100k+ ARR (+17%), full-year guidance raised to $1.129–1.133 billion, full-year non-GAAP EPS $0.85–0.87. Primary source: GitLab IR press release³
• Definition correction: some summaries put EPS at $0.25; the company's press release says $0.24, and this piece follows the primary source.
• Three flaws today's rally has not priced (this piece's core increment):
  1. Q3 guidance implies a sequential decline. Q3 guidance is $281–283 million, versus Q2 actual of $286.25 million — the midpoint is −1.5% QoQ. Checking its last 6 quarters of revenue: $214.51 → $235.96 → $244.35 → $260.4 → $264.16 → $286.25 million (unit: USD millions), every single quarter grew sequentially, without exception. If Q3 guidance is met, it will be the first sequential decline in at least 6 quarters.
     • One thing that needs saying: GitLab's self-managed subscriptions recognize license revenue upfront in a one-time manner under ASC 606, so this quarter's "record total bookings" may have pulled some license revenue forward into Q2, inflating the Q2 base. This is my most reasonable explanation for the phenomenon, but it is an inference, not confirmed by the company; the full-year guidance implies Q4 recovers to about $295 million, which is consistent with that explanation.
  2. Gross margin has declined for 5 consecutive quarters: 88.33% → 87.92% → 86.81% → 86.63% → 85.81% → 84.06%, a cumulative −4.27 percentage points. Over the same period the GAAP operating margin deteriorated to −19.89%. For a SaaS company this is the classic shape of rising AI inference costs.
  3. The sell side's reaction is restrained: William Blair only upgraded from Underperform to Market Perform (i.e. to "neutral"), not to buy.
• Pre-market and technicals: previous close $45.09, traded as high as $55.20 (+22.42%) pre-market, opened right at $55.20, currently $51.56 — it has given back 6.59% after the open. 52-week range $18.73–$55.55, that pre-market spike nearly touched the 52-week high before being sold.
• Final judgment: watch only (do not chase). The news is A+, but the opening price was the day's high, and at least two of the three flaws (gross margin, sell side only at neutral) are confirmed facts.

⑤ PANW (Palo Alto) — "beat + raised guidance" sold for the second day running; yesterday's open question is settled today

• Related news: after the close on 09-01, FY26Q4 EPS $1.02 (consensus $0.98), revenue $3.41 billion (consensus $3.35 billion); Q1 guidance $3.30–3.31 billion (consensus $3.22 billion), full year $14.1–14.2 billion, adjusted EPS $4.16–4.19. Source: CNBC⁴
• Settling yesterday's open question (carrying over from the 09-01 recap): yesterday's recap recorded PANW's after-hours reads as −0.02%/+0.44%, while the media basis was −2%, and it was flagged at the time as "pending next-day open to verify which is closer to true pricing." Today the answer is clear: opened at $346 (gap −4.44%), currently $332.27 (−8.24%). The media had the direction right, and those two after-hours reads captured an illusion produced by thin volume. This is the same shape as the memory item "pre-market quotes accumulate, a single snapshot is not a fact" — the same holds after the close.
• Pricing position: previous close $362.09,52-week range $139.57–$398.88, entered earnings at the 86th percentile, and it had already fallen 5.25% in the 09-01 regular session before the print. Cumulatively about −14% over two days measured from the 8/31 close.
• Final judgment: avoid. But one must acknowledge: this is a company that beat and raised guidance, and its decline is a valuation event, not a fundamental one. If it gets killed down below the 60th percentile, fundamental buyers will reappear — that is not now.

⑥ PCG / EIX (California utilities) — the verification point set yesterday resolves today in the worst possible way

• Related news (same-day primary): 09-02, PG&E announces it will defer about $2 billion of next year's investment. CEO Patti Poppe's own words: PG&E must be redesigned so it is "not held back by a below-investment-grade credit rating." Affected projects include new generation interconnection, large-load customers (including data centers), and new residential connections; the company still retains its first announced 1.8GW data-center project, but explicitly says "the pipeline will definitely slow." The company maintains its 2026 earnings guidance, but will re-evaluate its long-term EPS growth rate and its 2028–2030 capex outlook. Source: Insurance Journal, 09-02⁶
• Why this is new news rather than a repeat of old news: the 8/31 crash (PCG −18%, EIX −23%, EIX's worst single day since 2001) stemmed from the legislature not including a liability cap; on 9/1 the two rebounded +5.95%/+8.93%; today the company itself confirmed the consequence through action — moving from "legislative risk" to "an actual capex cut." That is a conversion from expectation to fact, not the third trade on the same piece of news.
• Carrying over from yesterday's recap: the judgment given yesterday for EIX/PCG was "neutral (no longer bearish, but don't chase)," and the verification point written was "whether any special-session-related statement emerges — this is a new binary risk source with unknown direction." That binary risk resolved to the downside today. The judgment "direction unknown" was right; the half-sentence "no longer bearish" was wrong — wrong in equating the closing of the legislative window with the risk being flushed out, and in ignoring that the company can still take action on its own.
• Transmission note: what PG&E is cutting is the data-center interconnection pipeline. If large-load interconnection slows broadly across California, more than the utilities themselves will be affected — there is no evidence for this today, so it is listed as an observation, not a conclusion.
• Final judgment: avoid. Currently PCG $12.705 (−9.64%), EIX $54.33 (−7.60%), and both are still falling after the open (−6.31%/−4.31%), with no sign of stabilization whatsoever.

⑦ DELL (Dell) — last night's big positive, and today the open was the high

• Related news: raised full-year guidance after the close on 09-01, backlog $95 billion, up as much as +8.57% after hours.
• Today's reality: previous close $425, opened $462.05 (gap +8.72%), currently $436.25 (+2.65%) — it has given back 5.58% since the open.
• Final judgment: watch only. Anyone who bought at the open is down 5.6% right now, and anyone who saw "+8.57% after hours" last night and decided to chase today got the worst price of the day. The unanswered question raised for DELL in yesterday's recap (AI server gross margin / operating margin, disclosed only on the call and not in the press release) still has not been obtained, and this piece cannot confirm it.

⑧ META — first-tier gainer, but the same-day catalyst could not be verified

• Today's action: $599.11 (+3.56%), opened at $578.79 and then rose steadily +3.51%, the only mega-cap today that was "continuously bought after the open" with a meaningful magnitude, on volume of 3.74 million shares.
• Catalyst verification result: the search found Morgan Stanley maintaining overweight, listing it as its internet top pick, price target $775, and saying "META has priced in all of the capex but none of the option value of any future product," estimating Meta AI search could add $2.89 per share, subscriptions about $1.88, and API about $1.22. But this piece could not confirm that the report was published on 09-02 (the related page returned 403, so the publication date could not be read).
• Honest conclusion: handled under "can't find news ≠ there is no news" — this piece records only price and flow behavior, and does not stitch a possibly-old research report into "today's positive." META's strength is a fact; its same-day attribution is unverified.
• Final judgment: watch closely. It has today's healthiest flow behavior (no gap, continuously bought after the open), but the reason for buying is unclear, which is itself a reason to cap the position.

⑨ The metals complex (NEM / GDX / FCX / CLF / STLD / NUE)

• Today's action: GDX +3.92%, NEM +3.51%, FCX +4.17%, XME +3.27%, CLF +6.50%, STLD +4.40%, NUE +3.94%, SLV +2.04%, GLD +1.35%, COPX +2.16%.
• Macro logic (with evidence): ADP badly below expectations → front-end rates ease slightly (2Y −1.1bp) → dollar weakens (UUP −0.32%) → non-yielding assets and commodities benefit. The evidence chain for precious metals is sufficient.
• The evidence chain for steel is insufficient: the moves in CLF/NUE/STLD (+3.9%~+6.5%) are markedly larger than macro can explain, but this piece could not find a sector-level independent catalyst dated 09-02. Everything the search returned was old content: July earnings and August US-Canada trade talks. No stitching, per the rules: this only records the facts, and the same-day catalyst is marked "unverified."
• A definitional conflict that must be flagged: the gold ETF GLD is +1.35%, while the CNBC gold futures read is $4,422.30 (+25.90, about +0.59%). The two differ by 0.76 percentage points. Judging by the internal consistency of GDX +3.92% and SLV +2.04%, the ETF complex's reading is more credible, and that futures print may be a non-synchronous quote. This piece adopts the GLD basis and discloses the discrepancy honestly.
• Final judgment: precious metals = watch closely; steel = watch only (no bet when the reason is unclear).

⑩ Airlines / cruise lines (UAL / LUV / CCL / NCLH)

• Today's action: UAL +3.79%, LUV +2.44%, NCLH +2.33%, CCL +2.30%.
• Logic: WTI −0.72 to $89.50, Brent −0.44 to $94.21, USO −1.67%, BNO −1.30% — four independent reads all confirm oil pulled back; fuel is about 20–30% of airline operating costs, so transmission is direct.
• Cross-sectional consistency check: the same day XLE −0.34%, XOM −0.79%, WMB −1.65% — energy stocks and airlines moved in opposite directions, exactly as an oil pullback implies, with no internal contradiction.
• Final judgment: watch closely. But it must be written plainly: this entire line rests on oil, and oil rests entirely on the US-Iran situation — it can reverse overnight and cannot be managed with intraday stops.

---

Sub-agent verification (internal)

fundamentals-analyst (CRDO) has returned, and overturned this piece's original core thesis; §5② has been fully rewritten against its primary sources (see "the most important correction of this run" at the end). bull-bear-debater (AVGO pre-earnings bull/bear) had still not returned at the time of finalization, so the fundamental details of AVGO in §5① (customer concentration, VMware renewal pricing, the single-quarter vs cumulative basis breakdown) have not received independent cross-verification; the current content is based on public disclosures and price data gathered by the main thread itself. If it returns before the email is sent, §5① should be backfilled.

---

6. Bearish / Avoid List

---

7. Within-Theme Ranking

Theme 1: Post-earnings valuation flush-out (software / AI networking)

Theme 2: Metals

Theme 3: AI hardware (AVGO adjudicates tonight)

---

8. Opening Verification Signals

① Today's strongest signal: every earnings gap is being eroded in reverse, while non-earnings themes are being continuously bought

This is the most operationally valuable table in this piece (read at 10:13 ET, 43 minutes after the open):

Conclusion: today is a "earnings stocks gap-fill, theme stocks gap-and-go" day. The two groups behave in exactly opposite ways, and the dividing line is not up-versus-down, it is "whether it is earnings-driven." Operational implication: do not chase gaps in earnings stocks today (in either direction), while post-open strength in theme stocks is credible.

② Still to be observed (rest of today)

• Whether MDB can hold its stabilization. The verification point set for it in yesterday's recap was "whether the next day's open can recover half of the after-hours decline" — the answer is no (after hours −11.56%, today −12.58%; not only was nothing recovered, the decline widened). But it is the only one that stopped falling after the open, so keep watching whether it turns green at midday.
• Whether GTLB can recover the $55.20 opening price. If it cannot, "the open was the top" is established, and everyone who chased today is trapped.
• Whether CRDO shows a high-volume bottoming. Volume is currently 6.05 million shares, and it is still accelerating downward with no bottoming signal.
• Whether the same-day catalyst for steel can be found. If there is still no news-based explanation by midday, then CLF +6.5% should be treated as "flow behavior" rather than a "fundamental event," and must not be extrapolated to tomorrow.

③ Tonight's binary event (most important)

• AVGO reports after the close, implied ±8.7%. What matters is not whether it beats for the quarter but: ① the single-quarter number in next quarter's AI revenue guidance (not cumulative, not annualized, not backlog); ② whether the $16 billion AI revenue path is reiterated or raised; ③ whether customer concentration rises further.
• Three branches must be pre-specified, including the "nothing happens" branch: ① beats across the board and raises guidance → the AI hardware line repairs, and it simultaneously confirms in reverse that CRDO was an idiosyncratic single-stock event; ② beats but guides blandly → a thin beat meets a ±8.7% implied move, replicating today's MDB/PANW script, and SMH comes under pressure; ③ in line with expectations and no clear market reaction — yesterday's recap explicitly required this branch to be enumerated, and today it is done.
• A cross-verification point that can retrospectively test today's judgment (proposed by the sub-agent, adopted here): if AVGO's AI/networking guidance tonight shows YoY acceleration, and ALAB/MRVL tomorrow recover today's excess decline relative to SMH, then "CRDO is idiosyncratic" is confirmed; conversely, if AVGO's networking/custom ASIC guidance falls short and ALAB/MRVL continue to underperform SMH by more than 2pp tomorrow, then today's conclusion that "a sector-level demand problem is ruled out" must be overturned.
• Also reporting tonight: HPE (consensus revenue about $12.1 billion, EPS $0.94, company's own guidance $11.5–12.1 billion / $0.88–0.93) plus SNOW and NTAP. Note that HPE is −1.47% today and has given back 3.60% since the open — that +4.09% of DELL spillover in yesterday's after-hours session has been entirely erased.

④ Risk signals

• Divergence between the index and its internals: SPY +0.28% but QQQ −0.03%, IWM +0.89%. The calm of the index masks a systematic valuation kill in the software sector.
• Hike pricing is the thing hanging over every high-valuation asset: the market prices the 9/16 FOMC as a hike (48%–66% across venues). Against that backdrop, the protection a "beat" offers high-valuation growth stocks is systematically weakened — this may be the true common cause behind today's four samples.
• Friday 9/4 nonfarm payrolls is this week's real arbiter: today's ADP print of 38,000 has already lowered expectations, and if payrolls are equally weak, hike pricing will loosen and today's metals/small caps will continue; if payrolls are strong, today's rotation will be taken back wholesale.

---

9. Final Conclusions

① The 5 stocks most worth watching today

② Today's 3 strongest themes

③ What to avoid today + why

1. Avoid chasing any earnings gap — in either direction. Of today's 7 earnings-gap stocks, 6 moved in the opposite direction of the gap after the open; for GTLB (+22.42% → gave back 6.59%) and DELL (+8.72% → gave back 5.58%), the opening price was the day's high.
2. Avoid the "a beat means it goes up" game. Today has already supplied three counterexamples, among which PANW beat + raised guidance + still fell 8.24%. This lesson was already written in yesterday's recap, and today it is re-verified on a larger sample.
3. Avoid software names at high 52-week percentiles. OKTA (92%), CRWD (87%), PANW (86%), MDB (85%) — the tendency for a higher percentile to mean a bigger drop today is clear. But CRDO is an explicit counterexample: 54th percentile, and the largest decline of the entire session. So "only the high ones get killed" is a wrong simplification — the percentile is one risk variable, not a sufficient condition, and certainly not the only variable. (This piece's first draft nearly wrote this up as a rule; see §5②)
4. Avoid California utilities. The decline in PCG/EIX has today shifted from "legislative expectation" to "the company actually cutting capex," the nature has changed, and there is still no stabilization after the open.
5. Avoid heavy positions in names with no same-day catalyst. CLF (+6.50%) and META (+3.56%) are both very strong, but this piece could not verify a same-day catalyst for either. They can be watched, they should not be held heavily.
6. Do not short AVGO or its chain. Tonight's earnings event risk is asymmetric.

④ Final one-sentence judgment

The market today is doing one thing: repricing "certainty," not pricing "good news." All four companies beat, and the market bought only the one with the lowest expectations (GTLab) and slammed the other three — against the backdrop of September hike pricing hanging overhead, good results are no longer a talisman; "how much was priced in on entering earnings" is. The only reliable operating discipline today is: do not chase earnings gaps, follow only the non-earnings-driven themes (metals, the oil-pullback beneficiary chain), and manage tonight's AVGO as a binary event rather than a trend continuation.

---

Internal Record (not sent to clients)

The scheduling failure on this run (must be recorded)

This piece's us premarket slot failed four consecutive times at 20:00/21:00/21:20/21:40 CST; in reports/run.log all four show ALERT SENT immediately after START, and the cause of failure was You've hit your session limit · resets 10pm (Asia/Shanghai) — i.e. the LLM session quota was exhausted, claude -p exited non-zero and triggered set -e. All four failures sent alert emails to the owner group (sk@sfg.xyz / msx@sfg.xyz). This run finally succeeded on the fifth retry after the quota reset at 22:00 CST.

Consequence: this piece was actually generated at 10:13 ET, 43 minutes after the regular session opened, so strictly speaking it is no longer a "pre-market" report. The approach taken here is not to pretend it is a pre-market draft — the opening of the body explicitly marks the read timestamp and presents the "open→now" column as core information, in fact exploiting the informational advantage of this timing (it can directly verify whether the gaps were eroded). But note: clients will expect a pre-market piece when they receive it.

Recommendation: the retry window in bin/cron_slot.sh should add a branch for "if the open has already passed, switch template/title," or simply skip retries on quota-type failures (retrying will not succeed, it will only send four alert emails again). The current */20 21-22 retry policy is pure alert bombardment when it hits a quota-type error.

Data channels

This piece did not use yfinance (consistent with recent runs). All single stocks/ETFs went through api.stockanalysis.com/api/quotes/s/<ticker>, which beyond p/cl/o/v/cp also provides h52/l52, and all "52-week percentiles" in this piece were computed from that — this field has not been used before and is worth locking in: it is the only quantitative tool today that separates MDB (85%)/PANW (86%)/OKTA (92%) from AVGO (40%), and it is the data foundation of this piece's core thesis.

Treasuries: switched to the official US Treasury Daily Treasury Yield Curve CSV as the primary source (only through 09-01; same-day data is not published until after 15:30 ET), overlaid with CNBC real-time reads. Reconfirmed that CNBC's change_pct field is unusable; this piece takes only last and previous_day_closing and computes the change itself.

Concurrency: single-threaded throughout with 0.22–0.35s intervals, 203/209 succeeded (the 6 failures include the ticker-format issue with BRK-B). No rate limiting occurred, unlike the previous piece. Conclusion: a 0.22s interval is safe in a single task, provided you do not run fetches in the foreground and background at the same time.

Search-side traps caught this run (four, all adjudicated with our own data)

1. MDB's previous close was widely written wrong. Several summaries put the previous close at $453.37 and the decline at 16.9%. Reverse-computing from the pre-market price falsifies it instantly: $381.43 ÷ (1−12.16%) = $434.2, consistent with the independent read of $434.21. This is another direct hit for the memory item "price-triple-must-be-reverse-computed," and this time the search summary was wrong and the API was right.
2. Search summaries conjured up two earnings reports that had "already happened." The summaries said "SNOW's Q2 report is out today, up 18.1% after hours to $188.4" and "NTAP up 5.5% to $76.83." In reality: SNOW is at $313.50 and NTAP at $181.19, and both report after the close tonight, not yet released. $188.4 and $76.83 are prices from years ago. Exactly the same shape as the memory item "phantom-earnings-from-search-summary," and it was caught the same way, with our own prices — off by 40% and 137%, visible at a glance.
3. GTLB's EPS was written as $0.25 in summaries; the company press release says $0.24. Corrected against the primary source. A one-cent difference looks harmless, but it also changes the "size of the beat" from 33% to 39%, and the size of the beat is precisely this piece's core argumentative variable.
4. A same-day market headline pointed in the opposite direction from reality. TheStreet's headline was "Dow edges higher as oil prices rise and bond yields climb," while the actual measurements were: WTI −0.72, Brent −0.44, USO −1.67%, BNO −1.30%; 10Y −0.8bp, 2Y −1.1bp. Both oil and yields were backwards. That headline must have been written before the open. Same family as the memory item "opening-snapshot-sold-as-close-recap," but the reverse direction (this time a pre-market piece was taken for an intraday one). Copying it would have inverted the entire macro framework in §0, and the airlines line would have disappeared.

⭐ The most important correction of this run: I took "the company's old guidance from last quarter" for "next quarter's new guidance," and built the whole piece's central thesis on it

This is the most serious factual error in this piece, corrected by the fundamentals-analyst sub-agent pulling the primary SEC 8-K EX-99.1. It must be committed to memory.

• What I wrote: CRDO "guided next quarter to $465–475 million, with a midpoint below this quarter's actual $479 million, i.e. a sequential decline" — and used that to explain the crash as "the second derivative of growth flipping," even elevating it to the central thesis of the whole piece ("a break in the growth slope"), and using it to derive AVGO's risk tonight.
• What the primary source actually says: the 8-K reads "Revenue is expected to be between $525 million and $535 million", +10.6% QoQ and about 2.6% above the consensus of $516.5 million.
• The true identity of $465–475 million: it is the Q1 guidance Credo gave last quarter (in June), i.e. the number it has just exceeded. Secondary reports wrote "the old guidance that was beaten" as "the new guidance," and I took it at face value.

Why I failed to catch it myself: this error was internally consistent throughout — a sequentially declining guide perfectly explained the −18%, and echoed GTLB's sequential decline beautifully, forming the elegant symmetry of "both guided down sequentially, yet the market bought one and slammed the other." The more elegant, the more dangerous: it was exactly that symmetry that made me stop asking questions. Same family as the memory items "self-correction-becomes-exemption" and "skepticism-needs-primary-source-too," but this instance is more basic — I simply never looked at the 8-K, I used a single sentence from Benzinga.

Lessons (recommend committing to memory, working title "a guidance number must be confirmed as to which period it belongs to"):

1. For any "guidance" number, three things must be confirmed at once: which period (which Q / which FY), when it was published, and whether it is newly given or the old one that was beaten. Secondary reports very easily conflate "last quarter's guidance" with "next quarter's guidance," especially in a "beat" context where the two numbers appear side by side.
2. Test: if a guidance number is below the just-reported actual for the quarter, first assume you have the period wrong, rather than first assuming the company is issuing a warning. For a high-growth company, sequentially declining guidance is an extremely rare event, and when a rare event appears the burden of proof is on me.
3. This piece's §5④ conclusion that GTLB's "Q3 guidance implies a sequential decline" stands independently — it comes from a two-way cross-check between GitLab's IR press release ($281–283M) and stockanalysis's 6-quarter revenue series ($286.25M), it is not the same error committed a second time. But precisely because the CRDO one was wrong, the GTLB one needs its evidence source labeled all the more clearly.

The sub-agent also corrected/added three things I did not have at all: ① the DustPhotonics acquisition ($750 million cash, goodwill +$894 million) — it explains almost all of the GAAP gross margin −3.7pp and the operating profit −22.5%; ② customer concentration is improving systematically (largest customer 67%→49%→33%, 10%+ customers 2→4), which directly rules out the "single-customer order timing" hypothesis; ③ the cross-section vetoes the sector explanation (SMH +0.32%, NVDA +2.27%, ALAB only −2.09%), and this one matters most: I was about to write CRDO into the AVGO risk derivation as evidence that AI hardware demand had peaked, which would have been a directional error.

Another methodological takeaway

The simplified narrative "only the expensive get killed" was falsified on the spot today by CRDO. The original formulation was "the higher the 52-week percentile, the worse the drop," and MDB (85%), PANW (86%) and OKTA (92%) supported it perfectly. But CRDO's percentile is 54% — the biggest decliner of the whole session sits in the middle of the range. Because all three samples were self-consistent, no internal contradiction would ever have flagged it to me. Lesson: when a rule has three self-consistent samples, you should actively go looking for a fourth counterexample rather than stopping at three. Same family as "narrative-first-screening-blind-spot." The corrected mechanism is the ratio of "size of beat vs implied move priced beforehand": CRDO implied ±13.1% and delivered a 1.8% revenue beat; GTLB delivered a 33% EPS beat. This is the same as the memory item "iv-crush-comes-from-valuation-not-preannouncement," but today gives it a quantifiable form: the numerator of the beat must be divided by the implied move as the denominator.

Open gaps requiring next-day follow-up

• The same-day catalyst for CLF/NUE/STLD was not obtained. Moves of +3.9%~+6.5% clearly exceed what macro can explain, and the search returned only July earnings and old August US-Canada trade news. The body marks it "unverified" and downgrades to "watch only," with no stitching of old news. Tomorrow, check whether there was any Section 232-related action on 09-02 or a jump in steel price indices.
• META's same-day catalyst was not obtained. The publication date of the Morgan Stanley report could not be confirmed (the TheStreet page returned 403). The body records price behavior only.
• DELL's AI server gross margin still has not been obtained — this was a gap already listed in yesterday's recap, and it remains unclosed today (still only in the call transcript).
• The cause of GTLB's sequential Q3 guidance decline has not been confirmed by the company. The ASC 606 license-revenue-pull-forward explanation offered here is an inference, and the body says so explicitly. Tomorrow, check the CFO's explanation of the Q3 guide in the call transcript.
• Direct WebFetch to SEC EDGAR returned 403 (GitLab 8-K EX-99.1); the same set of numbers was obtained from the company IR page instead. But the sub-agent, using Bash + curl against the same batch of SEC addresses (data.sec.gov submission index, www.sec.gov/Archives/... full 8-K text, full 10-K text at 380,000 characters), got 200 on all of them. → Important operational conclusion: SEC is not refusing us, it is refusing the WebFetch channel. From now on, whenever a primary SEC filing is needed, use Bash+curl (with a User-Agent) directly, not WebFetch. This piece's major CRDO correction happened precisely because the sub-agent went through curl to get the original 8-K — if I had used curl earlier, the first draft would not have been wrong.
• Three open items on CRDO (flagged by the sub-agent, must be followed up tomorrow): ① Q1 FY27 optical revenue was not disclosed — that is the progress measure for the ">$600 million" full-year target and the fulcrum of the entire −18%, and the 10-Q (expected to be filed 9/3–9/4) may or may not break it out; ② Q1 FY27 operating cash flow / free cash flow was not obtained (the 8-K contains no cash flow statement), while this quarter inventory +24.8% and receivables +23.7% are both 2.5x the revenue growth rate, so if OCF is far below non-GAAP net income, "building inventory for the second half" must be priced as ramp risk rather than foresight; ③ whether the ">85%" full-year guidance is a "raise" or a "reiteration" is in doubt — the CFO's wording was "We continue to expect," which does not fully square with the judgment that it was "raised from >80%," and the June Q4 FY26 8-K needs to be checked.

Sub-agent status

Two sub-agents were dispatched, fundamentals-analyst (CRDO) and bull-bear-debater (AVGO), and neither had returned at the time this draft was written. The corresponding places in the body are marked as lacking cross-verification. If they return before the email is sent, the fundamental details in §5① and §5② should be backfilled (especially CRDO's customer concentration — that is the key to judging "single-customer order timing vs demand peaking," and the only substantive question left unresolved in this piece).

---

⚠️ Risk disclaimer: this list is pre-market information organization and observation only and does not constitute investment advice. US equities are volatile and pre-market gap risk is high; after earnings there can be IV crush and guidance reversals. 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.