US Pre-Market Brief | 2026-09-03 (ET) Thursday

Coverage window: 2026-09-02 16:00 ET regular-session close → 2026-09-03 09:00 ET. Includes 09-02 after-hours earnings, the overnight Asian and European sessions and geopolitical events, this morning's pre-market, and the initial jobless claims already released at 08:30 ET. Time of writing: 2026-09-03 08:00–09:00 ET (30–90 minutes before the open). Price snapshots are taken at two moments, 08:10 ET and 08:47 ET, and each instance below is labeled with which one is used; the 10:00 ET ISM non-manufacturing print had not yet been released as of this writing. Price sources: single stocks / ETFs come from the pre-market fields of the stockanalysis.com quote API (ep/ecp/epv); index futures, crude oil, gold, Treasury yields, FX and bitcoin come from CNBC real-time quotes (freshness verified item by item via last_time); earnings figures follow the original text of SEC 8-K EX-99.1 filings and company press releases.

This piece's volume rule: pre-market volume is always converted into a notional amount (shares × price) rather than a share count, with a uniform threshold of $10 million; readings below the threshold never support a directional conclusion, and that threshold applies symmetrically to both bullish and bearish items.

---

0. One-Sentence Summary of the Day

1. Today's macro backdrop is something many people have not yet re-wired: what the Fed is debating is "hike or not," not "cut or not." Ahead of the September 15–16 FOMC, the market is pricing roughly 60%–66% odds of a 25bp hike (CME FedWatch / Forbes 08-31¹). The drivers are the US-Iran conflict pushing energy higher, layered on core inflation stuck at 2.5%. This fully inverts the old reflex that "weak data = good for stocks" — in today's 10:00 ET ISM services print, what actually determines the pricing is the prices-paid sub-index (70.3 in July), not the headline value.
2. Geopolitics kept escalating overnight, but the pre-market reaction is not a textbook risk-off. US forces launched a round of strikes on Iran's Revolutionary Guard on Tuesday (09-01), and Iran says it has retaliated against US military positions in Jordan / Kuwait / Bahrain / Iraq / the UAE (CNBC²). Brent $97.16 (+1.60%), WTI $92.93 (+2.11%), gold futures $4,487.6 (+1.65%). But Treasury yields are falling (10Y 4.774%, −2bp vs yesterday's close), the dollar is down 0.50%, and VIX is only 15.34. Oil up, gold up, bonds up and the currency weaker all at once — this is not pure haven demand, it looks more like "stagflation-risk pricing."
3. At the index level there is almost no information; what actually happened is one clean sector swap: software up, semiconductors down — precisely yesterday's mirror image. In the sample with notional ≥ $10 million (measured at 08:30 ET), software is 8/8 up (median +2.23%) and semiconductors are 12/14 down (median −1.09%). Yesterday's close was IGV −2.60% / SMH +0.96%; this morning's pre-market is IGV +1.27% / SMH −1.09%. Taking the two days together, this is the same pool of money switching sides inside tech, not a long/short call on tech as a whole.
4. The strongest single catalyst is SNOW and the heaviest single drag is AVGO — and the "surprise amounts" of the two are almost identical in size, opposite in direction. SNOW raised its FY27 product revenue guidance from $5.840 billion to $6.070 billion (+$230 million), pre-market +23.6% (08:47 ET, $269 million notional); AVGO's Q4 total revenue guidance of $34.8 billion is $230 million below the $35.03 billion consensus, pre-market −3.20% (08:47 ET, $713 million notional, heaviest of the session). The same $230 million is a +3.9% raise in one case and a −0.66% shortfall in the other — the market buys the percentage relative to expectations, not the absolute dollar amount.
5. Pre-market state: S&P futures 7,673.75 (−0.04%), Dow futures +0.18%, Nasdaq futures −0.26% (Nasdaq weakest, consistent with semiconductor weakness); 10Y 4.774% (−2bp); VIX 15.34 (+0.92%); DXY 99.098 (−0.50%); USD/JPY 155.98 (−1.71%, yen sharply higher). Overnight Asia and Europe: Nikkei −0.17%, Hang Seng −0.39%, Taiwan −0.67% (same direction as semis), STOXX600 +0.15%, FTSE +0.39%.

⚠️ The single most important structural reminder in this piece: the only five categories with genuine pre-market price discovery today are "mega-cap tech / semiconductors, software, gold, oil ETFs, and earnings movers." Every other sector has almost no pre-market volume (08:10 ET readings, share counts are the most intuitive here): banks KRE just 772 shares (about $60,000), steel STLD just 100 shares, XME just 121 shares, defense LDOS just 508 shares, tankers FRO 17,000 shares (about $800,000), TNK just 4 shares. The three questions yesterday's recap left for today (KRE's direction after ISM, whether steel and precious metals move back in the same direction, and tankers' independent driver) cannot be answered at all from pre-market data — this piece refuses to answer them with noise and leaves all three to the regular session.

---

1. News Overview

---

2. Strongest Themes, Descending

---

3. Overall Single-Stock Strength Ranking

Reading note: the "pre-market %" and "notional" in the table below are read at 08:47 ET; where the overview section quotes the 08:10 ET readings, that is noted separately. The overall tape warmed up between the two moments (SPY from −0.02% → +0.19%, QQQ from −0.27% → +0.02%, SMH from −1.09% → −0.54%), which shows the pre-market is a cumulative process rather than a static fact; no single-point reading in the table below should be treated as "already settled for today".

3A Bullish Side

3B Bearish / Weakening Side

---

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

Deduction rationale in detail:

• SNOW −12: already +23.6% pre-market and 10.5% above the 52-week high; P/S (TTM) 24.0x is higher than any quarter-end reading in the past 3 years (range 9.4–23.0x); IV crush after earnings; still a huge GAAP loss.
• AVAV −1: the only deduction is the approaching 09-09 earnings date; the share price is in its 52-week low zone rather than at a high, so gap risk is actually minimal.
• TSLA −5: event-driven "sell the fact" risk.
• GDX −6: today's 10:00 ISM prices sub-index can reverse it in a single print; gold still has 20% of room to the January high but has already risen for several days running.
• NOW −5: no catalyst of its own, a pure sector rebound; a rate-hike environment is unfavorable for duration assets.
• XOM −9: XLE is already above its 52-week high and CVX is only 0.7% off its 52-week high, so this line is close to "already re-priced".

---

5. Detailed Analysis of Top Names

① SNOW — the only one whose earnings report is genuinely solid on its own, but the price has already run ahead of the fundamentals

• Related news (09-02 16:05 ET, SEC 8-K EX-99.1⁴, primary-verified): product revenue $1.4919 billion (+37%), total revenue $1.5468 billion (+35%), NRR 126%, RPO $9.00 billion (+30%), 828 customers over $1 million (+27%). The CFO's own words: "Q2 marks our third consecutive quarter of product revenue growth acceleration".
• Catalyst logic: this move is driven by "guidance," not by "a beat in the reported quarter." FY27 product revenue guidance went $5.840 billion → $6.070 billion (+3.9%), with the growth rate framing going from 31% → 36%; non-GAAP operating margin 13.5% → 14.5%; Q3 product revenue guidance $1.588–1.593 billion (+37~38%). Raising growth and margin at the same time is the scarcest combination this quarter.
• Theme and stage: this is the starting day for the software sector's shift from "valuation washout" to "rebound," and SNOW is the standard-bearer of that line.
• Fundamental verification (primary 8-K + third-party valuation):
  • Still a huge GAAP loss: Q2 GAAP operating loss of −$263.0 million (−17.0%) versus non-GAAP operating income of +$237.0 million (+15.3%), a 32.3pp difference. Per the reconciliation table in the 8-K itself, stock-based-compensation-related expense is 29% of revenue (= $456.4 million / $1.5468 billion), i.e. about 91% of the gap, with the remainder being 3pp of intangible amortization. TTM net loss $1.0908 billion.
  • The cash flow is real: TTM FCF $1.1953 billion, 22.0% of revenue, consistent with the full-year adjusted FCF margin guidance of 23.0% (the mere 5.4% in Q2 alone is seasonal; Q1 was 16.7% and the January quarter 59.6%, so a single quarter cannot be extrapolated).
  • Valuation: at the pre-market price of $377, market cap is about $130.6 billion, P/S (TTM) 24.0x; the TTM P/S range across the past 13 quarter-ends is 9.4–23.0x — today's reading is higher than any quarter-end in the past three years. Forward (FY27 product revenue $6.07 billion) is about 21.5x.
• Pre-market and technicals: +23.59%, notional $269 million (second heaviest of the session, with both price and volume real). The pre-market price of $378 is already about 10.5% above the 52-week high of $341.95 — meaning it opens in all-time-high territory.
• Final judgment: watch only (do not chase the opening price). The quality of the report is enough to support a re-rating, but yesterday's complete GTLB script (+22% after hours → the open was the high of the day → gave back 6.59%) is less than 48 hours old; and P/S is already at a three-year high. If you must participate, the only acceptable setup is to wait for it to hold the opening price for 30 minutes after the open before considering it, not to take it in the opening auction.

② AVGO — the heaviest single block of pre-market selling pressure in the market, but what is being sold is not "AI demand"

• Related news (09-02 16:05 ET, 8-K EX-99.1³, downloaded and verified by this piece itself): Q3 revenue $29.591 billion +86%; GAAP operating income $15.955 billion (+171%), non-GAAP $20.095 billion (+92%); non-GAAP EPS $3.32 (+96%); FCF $13.665 billion = 46% of revenue; semiconductor solutions $20.839 billion (+127%), infrastructure software $8.752 billion (+29%). Q4 guidance: revenue about $34.8 billion (+93%), non-GAAP operating margin held at 66%.
• Catalyst logic — this must be split into two layers:
  1. The 8-K layer (known): Q4 total revenue guidance is about $230 million (−0.66%) below the $35.03 billion consensus. The entire shortfall comes from the non-AI business: Q3 non-AI semiconductors were about $4.14 billion, only +4% year over year; whereas AI semiconductors were $16.7 billion +221% YoY / +54% QoQ, with Q4 guidance of $21.7 billion +236%.
  2. The call layer (what was actually digested overnight): Q4 blended gross margin guidance of about 73%, versus 78% a year ago; the CFO attributes it to "a rising XPU mix, whose higher memory content dilutes gross margin." This number appears only on the call, not in the 8-K. This piece independently reconstructed that trend from the 8-K income statement (adding back acquisition-related intangible amortization and restructuring charges in COGS; stock-based compensation in COGS was not added back, so the absolute level is slightly below the company's non-GAAP basis, but the direction and slope are comparable): Q3'25 76.8% → Q2'26 76.1% → Q3'26 74.2%. A continuous decline, with Q3 alone down about 190bp sequentially — consistent in direction with the company's "down again to about 73% in Q4."
• Sequence: 8-K released → −5~6% after hours → the call gave the big FY27 numbers → recovered to roughly flat (measured at 18:03 ET, 13.14 million shares) → overnight digestion of the gross margin and the non-AI stall → −3.20% pre-market. That is: the call's "rescue" lasted only about 14 hours.
• ⚠️ Two figures that must be corrected (this piece's verification overturns the numbers carried over from yesterday's recap):
  • FY27 AI revenue is not $100 billion. The $100 billion figure is the old number from the 2026-03-04 Q1 call; the 09-02 call revised it up to "secured the supply to again double AI revenue to approximately $115 billion"; the basis is "annual revenue for which supply has already been locked in," not orders and not a cumulative total (FY26 $58 billion / FY27 $115 billion / FY28 $230 billion).
  • Neither "$30 billion of AI orders / $11 billion already delivered" nor "Anthropic is the largest XPU customer" could be traced to a reliable source by this piece, so neither is used. The transcript lists six XPU customers — Google, Anthropic, OpenAI, Meta and two unnamed frontier-model developers — with no ranking language of "largest" anywhere.
• Pre-market: −3.20%, notional $713 million (heaviest of the session). It is still −28.2% from the 52-week high of $495.
• Final judgment: watch only — neither catch the knife nor short it. A company with FCF at 46% of revenue and AI revenue up 221% year over year falling 3% because of a 0.66% total-revenue shortfall and a 200bp gross margin guidance cut is understandable in direction, but the magnitude does not constitute a trend; at the same time, "gross margin down three quarters in a row" is a real problem, so this is not the level at which to go long. This piece found no overnight sell-side rating or price-target changes whatsoever — attributing the decline to "sell-side downgrades" has no evidence behind it.

③ VSXY — the name most worth writing up today: the headline, the guidance and the share price all disagree with each other

• Related news (09-03 pre-market, press release⁶ + 8-K EX-99.1¹³): net sales $1.611 billion (+10%), comps +9%, adjusted EPS $0.95 (consensus $0.774, a 22.7% beat), full-year guidance raised. The media headline is "beat + guidance raise." The stock is −13.9% pre-market (notional $60.3 million, 30% of yesterday's full-day turnover, an extremely strong reading).
• The real reason is not "weak Q3 guidance" but this:

The size of the raise collapsed from +$120 million to +$10 million, an order of magnitude smaller. And Q2 alone beat the company's own guidance by $29 million (actual $124.0 million vs guidance midpoint $95 million). The money added to guidance ($10 million) is not even enough to cover this quarter's own overage ($29 million) → the implied H2 guidance was quietly cut by about $19 million (−4.85%). Two routes back-compute to the same answer: old implied H2 = $565 − $80.083 − $95 = $389.9 million; new implied H2 = $575 − H1 actual $204.039 = $371.0 million. A difference of −$18.90 million ✓

• Why Q3 is the most conspicuous cross-section: Q3 guidance for operating income is only $10–20 million, implying EPS of roughly −$0.12 to −$0.03, versus consensus of +$0.06 to +$0.14. It is not "off by a multiple" but "roughly halved with the sign flipped" — Q3 is the seasonally weakest quarter with a profit base near zero, and a $20 million difference is enough to flip EPS from positive to negative.
• The deceleration reason the company gave itself (verbatim, not inferred): the CEO said "we are increasing our strategic marketing investment" (the fashion show, the Angels Among Us documentary, holiday activity); the CFO said the company is "making deliberate investments". The mechanism is "spending the beat" — a discretionary spending decision, not weakening demand.
• ⚠️ Three important framing reminders:
  1. Q2's beat has nothing to do with the tariff refund — VSXY is the only one in this batch that cleanly excluded the $140.3 million refund (GAAP operating income $256.6 million − refund $135.8 million + restructuring $3.17 million = adjusted $124.0 million ✓). Restated, gross margin was 38.79% (+3.22pp) and the SG&A ratio 31.29% (−1.47pp), so Q2 is a high-quality beat.
  2. But in the very same report, revenue actually came in slightly below consensus: $1.6107 billion vs consensus $1.6198 billion (−0.6%), and the company's "near the high-end" is relative to its own guidance. Once again, "beating guidance ≠ beating consensus."
  3. The full-year adjusted EPS guidance midpoint of $4.575 is 2.0% below the $4.667 consensus; broken out: H1 beat by $0.18, H2 guidance was cut by $0.27, for a full-year net −$0.09.
• The implied Q4 is not high: back-computed implied Q4 adjusted operating income is $336–376 million, +6.1%~+18.8% year over year (midpoint +12.5%), versus +150%/+125% in Q1/Q2. So this is not a case of "leaning on Q4 to bail it out" — management has deliberately pressed the entire H2 margin expansion down to near zero. The latter is harder to rebut: the former you can choose not to believe, the latter is a cut the company made itself.
• Final judgment: short watch. Over 52 weeks it went from $22.02 to $102.46 (about 4.65x), and what rose was precisely the "margin-expansion slope" line — which is exactly the line being re-priced today. Risk note: it is already down 13.9% pre-market, and the 08:30 ET call is in progress — if the Q&A frames the H2 deceleration as "purely marketing spend, reversible," the decline could be taken back. This is the name most in need of watching the call today.

④ AVAV — today's only clean piece of news that "nobody knew before yesterday's close"

• Related news: first disclosed at about 17:40 ET on 09-02 (after the close), the US Army awarded a $464.8 million E-HEL (Extended-Range High-Energy Laser) contract, LOCUST X3 system (30kW), with "dozens of units" to be delivered. This is the Army's first production contract for a directed-energy weapon (DefenseScoop¹⁴).
• Why this item is worth more than the others: AVAV was only +0.85% in the 09-02 regular session, pricing in none of this; the +6.33% pre-market is first-time pricing. This piece specifically verified the first disclosure date and confirms it is news, not old news.
• The fundamentals must be written alongside it (weak): TTM losses (P/E n/a, ROE −10.03%, FCF −$141 million); forward P/E 45.15; −39.76% over 52 weeks, with the current $145 against the 52-week high of $417.86 putting it in the low zone rather than a high one; short interest 7.27%. There is also the legacy of the Space Force BADGER/SCAR stop-work order and accounting issues.
• Pre-market: +6.33%, notional $16.2 million (above the threshold, so the reading is usable).
• Final judgment: priority deep-dive. The reason is that the combination of catalyst directness and expectation gap is the best in the market today: its own contract, entirely unpriced yesterday, and a share price in its 52-week low zone — gap risk is actually the lowest on today's list. But it reports FY27Q1 on 09-09 (consensus EPS $0.32 / revenue $474.6 million), so the risk of holding through earnings must be assessed by each reader.

⑤ The gold complex (GDX / GLD / AEM / NEM) — four macro variables all pointing the same way; this is the most internally consistent line today

• Related news: September FOMC hike probability 60–66%; escalation of the US-Iran conflict; gold futures $4,487.6 (+1.65%); the dollar −0.50%; 10Y −2bp.
• Catalyst logic: oil up + gold up + bonds up + the dollar down all at once — those four are only internally consistent under "stagflation-risk pricing" — if it were pure haven demand the dollar should be rising; if it were pure risk-on gold should not be rising.
• Pre-market: GDX +3.03% / $110 million, GLD +1.85% / $258 million, AEM +3.16%, NEM +2.83%. GLD's and GDX's notionals are the 4th and 8th heaviest of the session, making them the heaviest money flows today outside the earnings movers.
• Positioning: gold at $4,428–4,488 is still far below the 2026-01-29 record high of $5,595 (about −21%), so this is not a chase-the-high level. GDX is −14.2% from its 52-week high.
• Final judgment: watch closely. The falsification condition is very clear: if today's 10:00 ET ISM prices-paid sub-index (70.3 in July) drops noticeably, rate-hike pricing loosens and this line weakens on the spot.

⑥–⑩ The Other Five (in brief)

---

6. Bearish / Avoid List

Two names not included in this list: ETD (the −11.21% is the $3.00 special dividend going ex, +0.28% once restated, not a decline); AVGO (direction is bearish but the fundamentals do not support shorting; it is listed 3rd in §3B and flagged "do not catch the knife and do not short").

---

7. Intra-Theme Ranking

Theme 1: Software rebound (today's strongest, but mostly names with no company-specific catalyst)

Theme 2: AI hardware / semiconductors (direction bearish, but the spillover is decaying)

Theme 3: Energy / Strait of Hormuz (the hardest logic, but the worst tradability)

---

8. Open-Verification Signals

① Pre-market (already observable)

• Whether the software vs semiconductor swap holds: in the sample with notional ≥ $10 million, software is 8/8 up (median +2.23%) and semiconductors are 12/14 down (median −1.09%). This is today's only structural fact with breadth in both directions, not a single point. Robustness check (mandatory, otherwise this statistic is being held up by SNOW alone): excluding the earnings mover SNOW, software is still 7/7 up with a median of +1.76%; excluding HOOD as well for its brokerage character, it is still 6/6 up (NOW +3.24%, CRWD +2.23%, MDB +1.76%, CRM +1.51%, PANW +1.28%, ORCL +0.52%), median +1.64%. The conclusion does not depend on SNOW.
• The spillover is decaying: between 08:10 → 08:47, SMH converged from −1.09% to −0.54%, NVDA flipped from −0.29% to +0.33%, MU converged from −0.81% to −0.01%, and SPY went from −0.02% to +0.19%. The pre-market is a cumulative process, not a static snapshot — any conclusion written from a single 08:10 point is already partly invalid at 08:47.
• Cross-market corroboration: Taiwan −0.67% (same direction as semis), Nikkei −0.17%, Hang Seng −0.39%; Europe is actually up (STOXX600 +0.15%, FTSE +0.39%).

② Intraday (first 30 minutes)

• SNOW: whether it can hold the opening price. Yesterday's GTLB script (+22% after hours → the open was the high of the day → gave back 6.59%) is the most recent control case.
• AVGO: if it recovers to within −2% in the first 30 minutes, it means "73% gross margin" has been fully digested; if the decline widens to −5%, the spillover will press down on all of SMH again.
• IGV vs SMH relative strength: −3.56pp yesterday, +2.21pp pre-market this morning. Watch the sign of that spread at the close — it is the most direct test of whether "the software washout is over."
• VSXY: the 08:30 ET call is in progress, and the intraday direction depends on whether the Q&A frames the H2 deceleration as "discretionary marketing investment."
• Banks / steel / defense / tankers: there is no pre-market volume, so real volume after the open is required to judge them — do not substitute pre-market noise.

③ Macro (the only moment today that can change the whole picture)

• 08:30 ET initial jobless claims (released): 206,000, slightly above the 205,000 expected, prior revised up to 204,000; continuing claims +8,000 to 1.779 million. The magnitude is tiny and not enough to change September rate-hike pricing; today's macro variable is still entirely riding on 10:00.
• 10:00 ET ISM non-manufacturing: the headline is expected at 54.3–54.5 (prior 54.1). But today's real variable is the "prices paid" sub-index (70.3 in July) — with September hike odds priced at 60–66%, the prices sub-index determines the rate path, and the headline value does not. Also note the employment sub-index prior of 47.4 (contraction territory).
• Transmission chain: strong ISM prices → higher hike odds → 10Y up → the software rebound (a duration asset) stops on the spot and gold weakens at the same time. This single data point can knock out both of today's strongest lines at once.
• 09-04 (tomorrow) payrolls is this week's actual verdict.

④ Options and Risk

• IV crush: SNOW / AVGO / HPE / NTAP / VSXY are all on the day after earnings, so implied volatility will collapse rapidly after the open, and the risk-reward of buying options to express direction deteriorates sharply after today's open.
• Post-gap reversal: SNOW (+23.6%), TLYS (+32.0%) and CHPT (+24.1%) are large gaps, so gap-fill risk exceeds gap-and-go.
• Lone-survivor risk: AVAV is the only big gainer in the defense sector (LMT/RTX/NOC are all <+1% pre-market and were all down yesterday), so if there is no sector follow-through after the open, this is a single-stock event rather than a sector move — which is exactly the correct characterization of it, but it also means it cannot be used to infer the sector.
• Divergence monitoring: Nasdaq futures −0.26% while Dow futures are +0.18%; the divergence inside the indices is itself today's main information.

---

9. Final Conclusions

① The 5 Stocks Most Worth Watching Today

② Today's Top 3 Themes

③ What to Avoid Today

1. Avoid reading AVGO as "peak AI demand." In the same 8-K, AI semiconductors are +221% YoY / +54% QoQ and Q4 guidance is +236%. The real negative is gross margin (Q4 about 73% vs 78% a year ago), and it appears only on the call, not in the 8-K.
2. Avoid chasing any retail stock with a "headline beat" unless you first restate the tariff refund. Of the 7 reports this morning, 4 received IEEPA refunds and the four treatments all differ from each other; DLTH's "adjusted EPS of $0.50" does not exclude the refund, and restated it is only about $0.06.
3. Avoid the tankers (FRO/TNK/STNG). This is the most counterintuitive item in this piece: a collapse in Hormuz transits looks bullish for tankers, but Frontline's actual Q2 VLCC TCE was only $152,700/day, about 29% of the Baltic TD3C index ($520,000+/day); and Q3 is already 86% locked in at $156,900/day, only +2.8% sequentially; TNK has only 44% of Q3 Suezmax locked at $105,000/day, below Q2's $109,171. That is: the spot index surge has not flowed into these companies' actual revenue, and FRO's pre-market $45.69 is already above its 52-week high of $45.59. Layered on near-zero pre-market volume, this line is "right logic, wrong instruments."
4. Avoid using pre-market data to judge banks / steel / defense. The pre-market notionals of KRE, STLD, XME and LDOS are all on the order of $100,000 or less, so the three questions left by yesterday's recap cannot be answered in today's pre-market.
5. Avoid adding to both the software rebound and gold before the 10:00 ISM. Both lines share the same premise (that rate-hike pricing stops strengthening), so a single data point can knock out both.
6. Avoid treating ETD's −11.21% as a negative. That is the $3.00 special dividend going ex, +0.28% once restated.
7. Avoid carrying over the "AVGO FY27 AI revenue of $100 billion" figure. It was revised up on the 09-02 call to about $115 billion; $100 billion is the old number from 2026-03.

④ Final One-Sentence Judgment

Today is not a "geopolitical shock day" but a "switching-sides-inside-tech day + a rate-hike-pricing day" — what is actually moving is that pool of money between software and semiconductors (software 8/8 up, semiconductors 12/14 down, exactly the reverse of yesterday), plus a macro backdrop many people have read backwards: the Fed is debating hiking, not cutting, so today's 10:00 ISM prices-paid sub-index will do more to determine today's close than the headline value or any single piece of news out of the Middle East.

---

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

Internal Notes (not sent to clients)

Data channels

yfinance failed across the board again: none of the 7 tickers in the retail batch succeeded, all returning YFRateLimitError. Consistent with recent runs, this piece did not use yfinance at any point.

The stockanalysis pre-market fields were fully usable today, and epv had values: all five fields ep/ecp/epv/es/eu returned real-time values during the pre-market session (read between 08:00–08:47 ET), with es="Pre-market". ⚠️ Correcting one record from yesterday's recap: yesterday noted that "epv is always null, so after-hours volume can only be taken from CNBC" — that was the after-hours behavior; during the pre-market session epv does have values, and all of this piece's volume criteria are built on it. Recommend changing the memory entry to "epv is available pre-market, null after hours."

CNBC futures were fresh today, the opposite of yesterday: the last_time for @CL.1/@BZ.1/@GC.1 was all 07:56 ET and incremented across two polls, so yesterday's silent lag did not recur today. Verified last_time item by item per the memory entry before relying on it.

⚠️ CNBC Treasury change_pct broke again today as a single field: US2Y showed last=4.367% and change=−0.019, while change_pct displayed +0.0391% (the opposite sign). Relying on it would have turned "yields falling" into rising, and in turn inverted today's entire macro characterization. This piece uniformly back-computed from last − previous_day_closing (10Y 4.774−4.794 = −2bp). Second direct hit for that memory entry.

EDGAR: two traps hit and worked around:

1. The "Q3 8-K" for AVGO returned by search was actually the old Q1 FY26 filing (avgo-02012026x8kxex99.htm, 02012026 = 2026-02-01 = Q1 period end). The file itself was HTTP 200, complete in content and correctly formatted; its only tell was that it says "first quarter" inside. Switching to data.sec.gov/submissions/CIK0001730168.json located the 09-02 8-K (accession 0001730168-26-000076), which finally yielded the correct avgo-08022026x8kxex99.htm. Same type as the memory entry "recurring column headlines hide the date," but this time the carrier is the SEC filing itself rather than a media headline — that is a new category.
2. The Archives path still requires the CIK without leading zeros; the submissions JSON side requires them.

WebFetch was 403'd by several sites (CNBC, TheStreet, Benzinga). CNBC was successfully retrieved using curl + a desktop UA + a regex to extract ``; recommend making this a standing fallback channel.

Tickers where the stockanalysis quote 404'd: HES, CMA, X (all three are already-merged / renamed tickers, not an API failure).

Sub-agent stability (the biggest operational problem this run)

3 of the 4 fundamentals-analyst agents failed midway with API Error: Connection closed mid-response. How they were handled:

• The SNOW / software one: a narrower one was launched again (asking only about valuation and earnings quality, not letting it write a long report) → succeeded, returned in 105 seconds.
• The AVGO / semiconductor and energy / Hormuz ones: recovered with SendMessage, with the recovery message explicitly requiring "output what you have already found immediately, do not run any new searches" → both returned successfully, and at high quality.
• Conclusion: after an agent fails on an API error, SendMessage recovery works, and on recovery the scope should be narrowed to "hand over the conclusions you already have." Making this recommendation standing.

Search-side traps caught this run (five)

1. "Vessel transits fell from 23 to 10" is Bab el-Mandeb, not the Strait of Hormuz. It was very nearly written into the report as the Hormuz transit count. The real Kpler readings for Hormuz are 08-30 = 10 vessels, 08-31 = 5 vessels, 09-01 = 4 vessels, 10-day average 13 vessels, against a pre-war baseline of about 100/day (a decline of roughly 95%). Both numbers are real, both are about "strait transit volume," and they are even close in magnitude — the only difference is which strait. This kind of error does not expose itself.
2. All three articles on the TSMC "Asia chip selloff" were old news from 2026-07 (the Bloomberg URL says 2026-07-17, CNBC 2026-07-29). Relying on them would have attributed today's −0.67% in Taiwan to an event two months ago. Caught via the date in the URL.
3. The search summary for VSXY said "shares surge on earnings beat, guidance raise," while the measured pre-market was −16.87% ($58.71 million notional, 30% of yesterday's full-day turnover). Another item, Simply Wall St's "Up 5.7% ... And Proxy Win," actually points to the Q1 report back in 2026-06. It was the measured price that falsified the search summary, not the other way around.
4. AVGO's "FY27 AI revenue of $100 billion" is the old number from the 2026-03-04 Q1 call, revised up on 09-02 to about $115 billion. Yesterday's recap carried over the $100 billion figure, which this piece corrects. The same-source items "$30 billion of orders / $11 billion already delivered" and "Anthropic is the largest XPU customer" could not be traced to a reliable source, so this piece refuses to use them (the former contradicts Q3 AI revenue of $16.7 billion).
5. ETD's −11.86% is the $3.00 special dividend going ex, not an earnings-driven decline. An EDGAR check shows its most recent earnings 8-K was 07-29, with no 8-K at all on 09-02 or 09-03. Arithmetic check: $24.71 − $3.00 = $21.71, measured $21.78 → +0.28% once restated. And it was verified that $3.00/$24.71 = 12.1% < the 25% threshold, so the ordinary ex-dividend rule applies. Second hit for the memory entry "a special-distribution ex-date disguises itself as a pre-market crash."

⭐ The most important methodological takeaway this run: for the verification point set yesterday on ALAB/MRVL, the answer is "the premise itself does not hold"

Yesterday's recap listed "whether ALAB/MRVL can recover their excess decline relative to SMH" as today's top priority. This morning's answer is neither "they recovered" nor "they did not," but:

• ALAB's pre-market decline disappeared steadily over the course of observation: −2.06% (yesterday's close) → −1.65% (08:00) → −1.39% (08:02) → −1.02% (08:47), outperforming SMH along the way; and its pre-market notional was only $4.40 million, which should never have been used to draw a conclusion in the first place.
• MRVL is essentially in line with SMH (−0.90% vs −0.54%), and MRVL reported back on 8/27, so today is not the day after its earnings at all — pairing it with ALAB yesterday already conflated two different kinds of names.
• CRDO is +0.67% this morning, so the spillover did not carry over; and its −20.04% on 09-02, checked against the 8-K, was "a sequential stall against high expectations" (Q1 was only +9.6% sequentially, and Q2 guidance of +9.6~11.7% sequentially does not accelerate), not an absolute-level miss.

Lesson: a verification point must be set together with "whether the premise holds." Yesterday's was "can they recover the excess decline," whose implicit premise is "there exists a meaningful excess decline that needs recovering." Today proves that premise does not hold even at the liquidity level (a reading on $4.40 million of notional does not constitute "underperformance"). Complementary to the memory entry "the verification point beats the headline conclusion": that one says the verification point is more reliable than the conclusion, this one says the verification point itself can carry a wrong premise.

Another one worth recording: one tariff refund, four companies, four treatments

VSXY excluded it, DLTH included it in "adjusted", WOOF included it but also gave a normalized figure, CHPT included it in both GAAP and non-GAAP. The same economic item, disclosed the same night, handled four ways. This means "who beat by more tonight" is an invalid question before normalization. Same type as the memory entry "whether non-GAAP excludes investment gains is the company's own option," but this time it is a four-way comparison on the same item, so the evidentiary strength is far higher than a single case.