US Pre-Market Brief | 2026-08-24 (ET) Monday

Coverage window: 2026-08-21 16:00 ET regular-session close → 2026-08-24 08:10 ET (includes the weekend, the overnight Asia/Europe session, and this morning's pre-market).

Quote conventions (this brief does not mix readings taken at different moments):

• Prior close, open, high, low, volume, pre-market price and pre-market volume for single stocks/ETFs come from the stockanalysis.com quote API; the prior close is the 2026-08-21 16:00 ET close, and each pre-market reading is tagged with its own timestamp (clustered at 08:00–08:10 ET).
• VIX, Treasury yields, the dollar index, commodity futures and crypto come from the CNBC quote API, with a uniform reading time of 08:10 ET (commodity futures 08:00 ET).
• Gold/silver/copper/crude are continuous futures contracts (@GC.1 / @SI.1 / @HG.1 / @CL.1 / @BZ.1) that carry a basis versus spot — do not mix them with spot figures.
• ⚠️ Pre-market volume and price are cumulative, not a snapshot. Every pre-market percentage change in this brief is a "cumulative value as of the tagged moment," and by the 09:30 open it may be completely different. A quote on thin pre-market volume is not a fact; this brief tags and down-weights every stock with pre-market volume < 20,000 shares.

---

⚠️ Read this one first: today's hardest catalyst is an SEC filing, not any news headline

This morning's media headlines framed the pre-market as "Canada trade talks collapse + Iran sanctions → Dow futures down." That framing does not hold today; both legs are dismantled by the pre-market data:

If the headline held, we should see	Actual pre-market readings (08:32–08:37 ET)
Iran sanctions day → oil up	WTI 85.36 (−1.95%), Brent 93.04 (−1.43%), USO −1.37% (442,000 shares), BNO −1.10%, XLE −0.28% — oil is down, and four independent lines agree
Canada tariff damage → US steel and aluminum benefit	NUE +4.05%, STLD +3.86%, CLF +2.93%, CENX +2.57% are indeed up, but the reason is not this round of tariffs — steel, aluminum and copper are explicitly excluded from 338 (see below)

Table scrolls horizontally

⚠️ But on "Canadian assets are not being sold," I got it wrong in the first draft — and wrong in exactly the way I myself banned in §8.

The first draft used BMO −0.81%, RY −0.06%, TD −0.24%, CP −0.20%, CNI −0.32% to assert that "the market is not pricing a Canada shock." On review: the pre-market volumes on those names were 226 shares, 1,221 shares, 718 shares, 213 shares and 689 shares — not one of them clears this brief's own 20,000-share threshold, and several of the values have already drifted (BMO is actually −0.25%, RY −0.46%, SU −1.10%, and ENB has flipped negative). Using three-digit-share quotes to refute a macro narrative is applying the thin-volume rule only in the direction that suits me.

The line that actually has volume is the one the first draft never checked: the core dispute in all three Section 338 proclamations is autos. STLA is −2.22% pre-market on 415,854 shares (real volume); HMC −1.14%. The auto chain is being priced; it just isn't showing up in Canadian bank stocks. This correction changes the conclusion from "the market is ignoring Canada" to: the market is pricing precisely who actually got hit by this tariff list, and Canadian banks/railroads were never on it.

⚠️ On "Canada tariffs are bullish for US steel and aluminum," the second round of verification produced something much stronger than "mechanism questionable": steel, aluminum and copper are explicitly excluded from Section 338, while Section 232's 50% on Canadian steel and aluminum has been in force since 2025-06-04 — 14 months of it already absorbed (Alcoa 10-Q original text), and Canadian steel imports into the US were already −32% YoY in 2025. There is no incremental protection left to capitalize. Three independent falsifications are in §5③; the sharpest is this: last Friday EWC (the Canada equity ETF) was +0.96% — on the last trading day before a 50% tariff took effect, the market bought Canadian stocks; and the day's leader was COPX (copper miners) +5.18%, with copper likewise excluded from 338.

What is actually rising is the spread between HRC and scrap: hot-rolled coil went from $797/short ton on 2025-10-13 to $1,147 (+43.9%) on 2026-08-10, while scrap rose only 8–11%, so the steel-scrap spread widened 61% in 10 months. NUE's gross margin went from 14.5% to 19.6% and EPS from $2.60 to $5.04 — all of it explained by that spread. This began in October 2025 and has no causal link to August 2026's 338.

What is actually moving today is something else: the AI hardware chain is being systematically de-risked, and the trigger is a prospectus filed after Friday's close.

Applied Optoelectronics (AAOI) filed a 424B5 on 2026-08-21, launching a $600 million ATM offering (agents Raymond James, Needham). This is not a rumor and not an analyst opinion — it is an SEC filing, with the size, the share count and the dilution all in black and white. AAOI is −12.59% pre-market (621,000 shares, 08:10 ET), and the whole optical chain is caving with it: COHR −6.23%, LITE −5.74%, CIEN −3.18%, CRDO −3.06%, FN −2.21%.

⚠️ But the causality has to be pulled apart — the second verification round corrected an over-attribution in this brief's first draft. Over the same window the entire chain was falling: COHR −5.71%, LITE −5.49%, CRDO −3.10%, FN −1.53% (CNBC, 08:06–08:28 ET). AAOI's −12.4% is roughly 2x or more the sector's decline, and only the excess half is what the ATM priced into the single stock. Charging the whole chain's decline to AAOI's offering is wrong — the sector was already falling; AAOI is simply the one falling hardest.

More importantly, the "cashing out at the top" intuition is also refuted by the numbers:

• AAOI is not at a record high — it is already 51.0% off its 2026-05-13 level of $223.10. This is issuing into a drawdown, not issuing at the top.
• $600 million is the right order of magnitude versus the already-disclosed capacity build: the company actually incurred $633.5M of capex in H1, while cash on the balance sheet is only $499.7M — it cannot afford the capex it already spent in the first half. The ATM equals 5.7 months of annualized capex and 8.1x the operating cash burn ($73.8M per half-year) — it cannot possibly be plugging operating losses; the company still has net cash of +$311.7M.

So the real red flag is not "why issue stock" but "what return is this money generating": capex/revenue went from 37% in H1-2025 to 185% in H1-2026, PP&E rose +85% in half a year, while over the same period gross margin fell from 30.3% to 27.7% (−255bp) and the GAAP operating loss widened from −$16.0M to −$24.7M. So far there is no financial evidence that this enormous capex is improving unit economics. (Sources: AAOI's 10-Q and 8-K EX-99.1 filed 2026-08-06, SEC primary.)

⚠️ And there is a thornier layer that only turned up on review: this group of stocks routinely falls this much together on days with no news at all.

At least 5 catalyst-free joint declines in the last two months: 6/23 (AAOI −13%, COHR −9%, LITE −8%), 7/02 (AAOI −17%, COHR/LITE −10%), 7/15 (AAOI −12%), 7/28 (COHR −11%, AAOI −10%, LITE −9%), 8/10 (COHR −12%, LITE −7%, with the contemporaneous report stating explicitly "classic profit-taking and pre-earnings de-risking… with no fresh negative headline behind it").

This morning COHR is −6.0% and LITE −5.7% — smaller than every one of those. In other words: using a 424B5 to explain a pattern this group has produced spontaneously 5 times in two months, at larger magnitudes, is mistaking the base rate for an event effect. The test I set in the first draft — "COHR/LITE closing down ≥3% supports this brief's conclusion" — cannot distinguish the two hypotheses at all, because they routinely fall 7–12% on no-news days. The test has been rewritten in §8.

Revised position: AAOI's ATM explains its own extra half of the decline; the chain-wide decline needs no news from today to explain it — it is an ebb that has been running for two months.

One more thing that must be said in the same breath: today is Day 0 of "this week's four big events," and every position taken today is really a bet on Wednesday. Wednesday (8/26) brings NVDA earnings after the close plus July core PCE that morning; Friday (8/28) brings Warsh's first Jackson Hole keynote as Fed Chair. This morning's pre-market shows QQQ −0.58%, SMH −1.23% against SPY at just −0.18% — that structure by itself is "cut AI exposure, don't cut gross exposure."

---

0. Today in One Sentence

① Strongest single catalyst: AAOI's $600 million ATM offering (SEC 424B5, filed 2026-08-21). Driver type = dilutive financing, not earnings, not guidance, not macro. It is today's only event with a first-hand filing + an exact amount + verifiable dilution, impact level A+. ⚠️ But it is not the cause of the whole optical chain's decline — COHR/LITE/CRDO/FN were all falling over the same window, the sector was already ebbing, and AAOI is merely the one falling about 2x the sector. Details in the box above.

② Strongest theme: the AI hardware chain is ebbing (optical → memory → AI power), while software and the broad index are barely touched. Optical comms −2.2% to −12.6%, SNDK −5.58%, MU −3.35%, MRVL −3.42%, semicap AMAT/KLAC/LRCX/TER −2.6% to −3.3%, BE −4.03%, NBIS −3.97%, CRWV −2.90%, IREN −2.65%; meanwhile MSFT −0.08%, META +0.20%, AMZN +0.30%, AAPL +0.32%, CRM +0.21%, SNOW +0.13%, NOW +0.12%. This is de-risking inside the chain, not "tech stocks are falling."

③ Likely flow direction: from AI hardware → hard assets (gold/crypto) + US steel and aluminum + defensives. What is up pre-market: GLD +0.86%, IBIT +1.37%, COPX +0.71%, XLB +0.52%, XLP +0.50%, XLU +0.37%, plus US steel and aluminum (NUE +4.22%, STLD +4.33%).

④ Driver-type distribution (very unusual today): financing/dilution 1 item (AAOI, S-grade), earnings 2 items (PDD delivers, XPEV does not), supply-side industry 1 item (Samsung's return program + YMTC IPO → memory), geopolitics 1 item (Iran sanctions day, but already priced), trade 1 item (Canada 338 tariffs take effect, but the mechanism is being widely misdescribed), macro 0 items (no heavyweight data today, only the 08:30 ET Chicago Fed National Activity Index, prior −0.02).

⑤ Pre-market state (08:10 ET): SPY −0.16%, QQQ −0.61%, DIA −0.02%, IWM −0.09% (08:32 ET); 10Y 4.71% (−3bp), 2Y 4.24% (flat), 30Y 5.238% (−4bp); VIX 15.92 (+5.22%); DXY 99.001 (+0.20%, the dollar is up); gold 4,714.40 (+0.72%), BTC 78,192.51 (+1.08%). Overnight Asia/Europe bias: Samsung Electronics closed −8.70% on the KRX today, SK Hynix −3.41%, KOSPI −3.1% (the program was released after Friday 8/21's close, so today is the first pricing day) (see §2 theme two); China ADRs broadly lower, with PDD the lone gainer on earnings.

⚠️ One methodological note that has to come first: gold and bitcoin are both up today, and so is the dollar (DXY +0.20%). So today's hard-asset strength cannot be attributed to "a weaker dollar" — that transmission chain does not hold today. What is actually cooperating is the long end: 30Y −4bp, and on 8/19 the Treasury just announced it would raise the per-operation cap on long-end (10–20 year and 20–30 year buckets) buybacks from $2 billion to at least $4 billion — but note it does not take effect until 2026-09-09, so that money is not in today's bid; what is moving now is expectation, not flow. The attribution belongs on "long-end real rates / policy credibility," not on the exchange rate.

---

1. News Overview

⚠️ Failures encountered while pulling and verifying data for this brief, for internal confidence assessment:

• Local yfinance was not used in the main pipeline; all single-stock quotes went through the stockanalysis.com API (correct pre-market convention), and macro went through the CNBC quote API. This is the established pipeline, not a failure.
• The CNBC futures symbols @ES.1 / @NQ.1 / @YM.1 / @RTY.1 returned entirely empty fields this morning (all four index futures unobtainable). Therefore §0/§8 of this brief cite no index-futures levels and use SPY/QQQ/DIA/IWM pre-market ETF quotes instead. This is a convention downgrade, not missing data — but readers comparing this brief against headlines like "S&P futures +0.4%" will find an inconsistency, and this is why.
• Multiple media sources returned 403/429: thestreet.com (403), specific cnbc.com article pages (403), coindesk.com (429), tradingkey.com (403), stocktitan.net (DNS timeout). Neither CNBC's Iran-sanctions original nor CoinDesk's bitcoin analysis piece was retrieved in full text, so the related statements rely only on search summaries + already cross-verified price data, and are tagged "no primary obtained" in the text.
• On Samsung's return program and the YMTC IPO, as of writing only English-language secondary relays were available; Samsung IR and the SSE primary announcements were not retrieved. A sub-agent is verifying. The text is tagged "no primary obtained" and downgraded accordingly. The "−9%" figure for Samsung's share price is not cited in this brief, because it cannot be traced to a Korean-close primary source.
• One error that nearly made it in: a search summary said "Morgan Stanley cuts MU's target price from $105 to $75" and implied it was today's event. MU trades at $966.78, so this is plainly years-old news re-indexed by a recurring-column page. Removed, not used. Likewise, this morning's entire "analyst rating changes" block (CARA target cut to $1, PASG cut to $5, etc.) could not have its publication dates confirmed, so the whole block was discarded and this brief carries no ratings-change table.
• The PDD and XPEV earnings figures come from secondary financial-media relays (Investing.com / PR Newswire reprints); the company IR sites investor.pddholdings.com and ir.xiaopeng.com both returned empty responses, so the primary press releases could not be retrieved. The figures have been cross-checked against two independent sources, but they are not grounded in the company originals, so confidence is slightly below the AAOI item.

---

2. Strongest Themes, Descending

⚠️ Themes 2 and 4 in this section have causal explanations completely different from the first draft after second-round verification. The corrected versions appear in the table below and in §5; the first draft's erroneous explanations have been deleted, but the "where I got it wrong" notes are kept for tomorrow's recap reconciliation.

---

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

⚠️ Note on pre-market reading times (tightened after review; no longer claiming a "uniform moment"): the body of this table reads at 08:08–08:10 ET, the names added in §5.5 read at 08:27–08:37 ET, and illiquid names may carry far older quote times — CMC's, for instance, is 07:17, OIH's is 07:40, KALU's is 07:18. stockanalysis only updates a pre-market quote when there is a new trade, so the "reading time" is effectively "the time of the last trade"; the smaller the volume, the staler that moment. Any name with pre-market volume < 20,000 shares is tagged and not used as fact. The final full refresh before the open is in §8.5.

3A. Bullish Side

3B. Bearish Side

---

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

Component definitions are in the workflow. Below is the line-by-line breakdown for five representative names.

⚠️ Two scoring conventions have to be stated up front, or this table will be misread: ① For bearish names the components are "mirrored." The model was originally designed to rank bullish ideas. For bearish names, "earnings elasticity" scores the degree of deterioration and "moat/fundamentals" scores fragility — a higher score means a stronger bearish case, not a better company. ② For bearish names the "risk deduction" scores "the risk that this brief's bearish call is falsified," not trading risk. Trade-level positioning (e.g., AAOI already down 12.6%, poor odds on chasing short) does not enter the score and appears only in the conclusion.

⚠️ Three self-imposed constraints on the scoring method: 1. The score measures event strength, not executability. AAOI still carries the market's highest score at 72, but the conclusion is "avoid / weak short watch," not "priority deep-dive" — it is already down 12.6% and the odds on chasing it short pre-market are poor. A high score ≠ tradable now. 2. This column was revised down sharply on review, and the process is left here for tomorrow's reconciliation. The first draft gave AAOI 86, with full marks on "catalyst directness 20 + expectation gap 10," resting on two premises later overturned: (a) "AAOI's offering crushed the whole optical chain" — in fact COHR/LITE/CRDO/FN were all falling over the same window, and this group has already fallen together with no news 5 times in two months, at larger magnitudes; (b) "disclosed only after Friday's close, completely unpriced" — in fact this is AAOI's 4th ATM of 2026, and the previous one was also $600 million. Those two cuts total 7 points, and the risk item was re-estimated on the "risk this call is falsified" basis. 3. The first draft's components did not add up to its stated total; this is fixed. In the prior version the components summed to only 51 while the total read 86 — the total was made up, not computed. In this version every column's "subtotal − risk deduction = total" can be checked cell by cell, and §3's totals match this table.

---

5. Detailed Analysis of Top Names

① AAOI — a $600 million ATM: today's only self-evidencing filing

Related news: 2026-08-21, SEC 424B5 filing (original¹).

Hard numbers copied from the original (all from the prospectus body, not relayed):

• Offering size: up to $600 million of common stock, at-the-market (ATM), sold at prevailing prices through Raymond James and Needham, commission cap 2.0%.
• Shares outstanding before the offering: 84,906,289 (as of 2026-08-20).
• Pricing reference cited in the prospectus: the 2026-08-20 close of $129.10.
• Net tangible book value (NTBV) as of 2026-06-30 was $1,664,391,529, or $19.72/share; rising to roughly $26.20/share after the offering.
• Immediate dilution to new investors of $102.90 per share.
• Use of proceeds: "general corporate purposes, which may include repayment of debt, working capital, capital expenditures and acquisitions" — the prospectus names no specific capacity project (but the earnings call does, see below).

⚠️ Here is the second and most important judgment this brief overturned on review: this is not the "second" one, it is 2026's fourth — and its signal points the opposite way from what I wrote.

I pulled every 424B5 original filed under EDGAR CIK 1158114 in 2026 and listed the sizes and pricing references:

The originals also state that as of 2026-03-12 the $250 million from February had been fully sold — "we have sold an aggregate of 2,476,307 shares … for gross proceeds of approximately $249,999,983" (implying an average price of about $100.96) — and it took less than two weeks. Shares outstanding went from 75,198,817 to 84,906,289 in 6 months, +12.9%.

This table overturns two things in the first draft:

1. "The expectation gap is very high, completely unpriced" does not hold. A company running its fourth ATM in 6 months, whose last one was also $600 million, doing another $600 million is not a surprise. The expectation-gap score in §4 has been cut from 10.
2. "The company is selling stock at the top → management thinks the price is attractive" reads the sign backwards. The previous ATM of the same size was struck at $223.10; this one at $129.10, 42% lower. By my own signal logic, this looks more like "having to keep selling at a worse price" — which is consistent with the magnitude test ("H1 capex $633.5M > period-end cash $499.7M") and inconsistent with "cashing out at the top." That reading has been deleted from the text.

⚠️ The true scale of dilution is far more than 5.5% (the following comes from 10-Q Note 3 and the 8-K, SEC primary):

Amazon warrant terms (granted 2025-03-13, 10-year term): 7,945,399 shares in total at an exercise price of $23.6956 — a 78.3% discount to this morning's price, with intrinsic value of roughly $680M. Of those, 6,621,166 shares vest on the condition that "Amazon and its affiliates purchase a cumulative $4 billion within 10 years" — a threshold that is 6.7x AAOI's current TTM revenue of $596M. In accounting terms, the grant-date fair value ($12.64/share) is charged against revenue from Amazon, with roughly $2.1M charged in H1-2026. This portion is explicitly excluded from the prospectus's dilution math.

One directly quotable corroborating point: when the company itself guided Q3 non-GAAP EPS on 8/6, it used a share count of about 92.8M, 9.3% above the 84.9M as of 8/20 — and that 92.8M does not include this new $600 million ATM.

Catalyst logic (revised per the verification results): $600M ÷ a $10.6 billion market cap = about 5.5%, and that arithmetic alone cannot explain −12.6%. But the "cashing out at the top" intuition does not hold either:

• The magnitude test supports "financing a capacity build": H1-2026 capex totaled $633.45M (PP&E purchases $335.13M + equipment prepayments $289.74M + land purchases $8.58M), while period-end cash was only $499.74M — the company cannot afford the capex it already spent in the first half — and on the 8/6 call the CFO said explicitly that capex intensity would be "higher in the second half." The ATM is 8.1x the operating cash burn, so it is not there to plug operating losses; the company still has net cash of +$311.7M. The capacity roadmap is also checkable in the 8-K original: currently "approaching 200,000 units per month" → roughly 650,000 units per month of 800G and 1.6T product by year-end (3.25x).
• But the issuance cadence really is opportunistic: the previous ATM was signed the day after the all-time high; shares outstanding are +164.3% over 3 years, with an offering in every one of the 14 quarters from 2023Q1 through 2026Q2, not one quarter flat; cumulative raises from FY2023 to date are about $1.81 billion, or roughly $2.41 billion including this authorization. The most recent 40 Form 4s show insiders selling a combined 520,215 shares / $89.5M at an average price of $172.07, heavily concentrated in the $166–205 range in May–June.

⚠️ One fact I got backwards in the first draft and must state plainly: AAOI is not at an all-time high. It is already 51.0% off the $223.10 of 2026-05-13, and it is +258.1% YTD (against the 2025-12-31 base of $34.86), not "up more than 4x." This is issuing into a drawdown, not issuing at the top.

Fundamental verification (SEC primary, 10-Q period ending 2026-06-30):

• Non-GAAP net income in Q2 was +$5.5M / +$0.06. The gap between GAAP and non-GAAP is a convention this brief must flag — the "profit" the media quotes is the latter.
• 14 consecutive quarters of GAAP operating losses. The reason the loss is widening is arithmetic: a 27.7% gross margin is below the 40.6% operating expense ratio (Q2 opex $77.9M), so every extra dollar of revenue widens the loss. Within that, G&A is $31.6M (+71.7%, 16.5% of revenue).
• TTM (25Q3–26Q2): revenue $595.97M (+61.8%), operating profit −$67.4M (−11.3%).
• The key line on cash-flow quality: H1 operating cash flow was −$73.78M, but a +$142.2M increase in accounts payable is holding it up; had payables been flat, OCF would have been about −$216M. This is financing a capacity ramp on supplier payment terms. H1 FCF (OCF − all capex) is −$707.2M.

Valuation (at this morning's $109.25 / $9.28 billion market cap): P/S (TTM) 15.6x, EV/Sales 15.0x, P/B 5.56x; forward P/S roughly 8.5x (annualizing the midpoint of Q3 guidance). Across peers: LITE 25.8x, CRDO 32.2x, COHR 8.0x, FN 3.4x — AAOI is not the most expensive among peers, but it is the only one with a negative operating margin, and its 28.9% gross margin is above only contract manufacturer FN. It carries a growth multiple while delivering losses and low margins.

⚠️ Two things I missed in the first draft that matter a lot today:

1. The Q3 guidance was in fact below expectations, and that happened back on 8/6. From the 8-K EX-99.1 original: Q3 revenue $255M–$290M (midpoint $272.5M, below the $278M consensus); non-GAAP gross margin 29%–30.5%; non-GAAP EPS $0.11–$0.26, with the ceiling below the $0.28 consensus. The top of guidance is below consensus — this is the overlooked carry-over negative on this chain today.
2. The widely circulated "Q4 revenue will exceed $500 million" is not guidance. A word-by-word search of the 8/6 8-K and EX-99.1 finds no revenue statement matching "fourth quarter / Q4 / 500," and the Business Outlook section covers only Q3. That figure comes from management's verbal outlook on the call (the CFO said "$330M is directionally right" for 800G + 1.6T). The main pillar of the current valuation is a call remark, not a guidance figure — that distinction has to be made explicit.

Pre-market and technicals: −12.59%, $109.10, pre-market volume 621,273 shares (08:10 ET). 52-week range $18.50–$233.67. Friday's regular session was already −3.32%, so roughly −15.5% cumulative over two days from Thursday's close.

Final call: Avoid / weak short watch (downgraded). The corrected conclusion is weaker than the first draft's: this is not a story about "management cashing out at the top," it is "a company genuinely pouring money into capacity, whose returns on that capacity have yet to appear, issuing another $600 million of stock after a 51% drawdown." The former is a tradable short thesis; the latter is not — it is more of a valuation-anchor reset, while the ATM's persistence means rallies will be sold repeatedly. The 12.6% decline has already played out pre-market, so the odds on chasing short are poor; what really deserves watching is whether that $500M verbal Q4 outlook becomes formal guidance.

---

② PDD — a revenue miss yet up 4%; what is rising is "expectations were too low"

Related news: Q2 unaudited results released before the US open on 2026-08-24, management call at 07:30 ET.

The figures (cross-checked across secondary sources; the company IR original was not obtained):

• Revenue RMB 112.4B (about $16.6B), +8% YoY (RMB 104.0B in the year-ago period), below the RMB 113.9B market estimate → this is a miss.
• Adjusted earnings per ADS RMB 19.33 ($2.85), above the RMB 18.35 estimate → this is a beat.
• Adjusted net income attributable to ordinary shareholders RMB 28.5B (about $4.2B), −13% YoY (RMB 32.7B a year ago).
• Adjusted operating profit RMB 29.1B (+5% YoY).
• Transaction services revenue RMB 54.7B (+13%) is the main source of growth.

Catalyst logic: this is a "profit is falling but not as badly as feared" quarter. Revenue growth slowed to +8% and adjusted net income fell 13% YoY, but operating profit is still +5% and EPS beat, meaning Temu's overseas spending and domestic subsidies are eating profit more slowly than the sell-side models assumed.

⚠️ A convention trap that must be named: "EPS beat" uses the adjusted (non-GAAP) basis, and "net profit −13%" is also on the adjusted basis. The two are not contradictory — the beat is against slashed estimates, not against last year. Reading the +4% as "a fundamental turnaround" is wrong.

Theme and stage: the opening phase of a low-expectations repair, but occurring against a broadly weakening China ADR complex (BABA −2.30%, NTES −2.63%, BIDU −1.46%, JD −1.09%), so the lone-gainer character is pronounced.

Pre-market and technicals: +4.16%, $92.06, pre-market volume 920,252 shares (08:10 ET), enough volume to constitute a fact. 52-week range $71.94–$139.41, putting the current price at roughly the 30th percentile of that range — not a chase-the-high spot, which is the best thing about this name.

Final call: Watch closely. A +4.16% gap is moderate and does not amount to a "do not touch" gap; the real issue is that it is the only China ADR that is up. §8 sets the verification point.

---

③ US steel and aluminum — the tariff rationale is thoroughly falsified, but the reason it is rising is real (just a different one)

⚠️ This is the most heavily revised section in the brief. The first draft said "the mechanism is questionable, this may be sentiment"; after second-round verification the conclusion is both stronger and more specific: Section 338's direct contribution is close to zero, and the real driver is a variable entirely independent of tariffs that has been running for 10 months.

Three mutually independent falsifications first:

Falsification one (statutory): steel and aluminum are not only excluded from 338, they were already hit with 50% under 232, and have been taking it for 14 months.

The original sentence in Alcoa's 10-Q (filed 2026-07-30): "...subject to a 25 percent tariff beginning March 12, 2025 until increasing to 50 percent on June 4, 2025 under Section 232..."

So 338 excluding steel and aluminum is not an "oversight" — it is because they are already fully taxed under a different statute. And the substitution has already fully played out: Canadian steel imports into the US were −32% YoY in 2025, with H2'25 down 44% YoY (MEPS). There is no incremental protection left to capitalize.

Falsification two (Friday's jump was not a steel move at all). Sector ETFs on 8/21:

COPX (copper miners) +5.18% > XME +4.05% > PICK +3.37% > SLX (steel) +2.29% > XLB +2.14%, against SPY at just +0.41%, XLI +0.27%, TLT −0.35%, UUP −0.04%.

The leader was copper — and copper is likewise explicitly excluded from 338. More damning still: EWC (the Canada equity ETF) was +0.96% that day. On the last trading day before a 50% tariff on Canada took effect, the market bought Canadian stocks. This was a broad rotation into metals and mining, not a "protected America" trade.

Falsification three (this morning's pre-market volume cannot support any conclusion at all): NUE 4,026 shares, STLD 8,563 shares, CMC 892 shares, RS 1,829 shares, KALU just 90 shares (it flipped pre-market from +2.56% to −0.50% entirely on 6 shares traded). Meanwhile the only one of the three aluminum names with volume, AA (30,878 shares), is exactly +0.00%.

---

So what is actually rising? — the spread between HRC and scrap, and that has been running since October 2025.

Source: SteelBenchmarker™ Report #488, price date 2026-08-12 (US hot-rolled coil, FOB mill, $/short ton):

That is +36.9% YoY and +43.9% off the October 2025 low. On a metric-ton basis it is $1,264/t, already above the prior record peak of $1,203 set on 2008-07-28.

The key point: this is a spread story, not a price story. Over the same period scrap rose only 8–11% (#1 Busheling 427→461, shredded 378→421, #1 HMS 322→356). The steel-scrap spread widened from 2025-08's $457/short ton to 2026-08's $735,10 months, +61.0%. That is the entire mechanism behind three EAF mills' gross margins jumping simultaneously.

Three pieces of hard evidence "independent of tariffs":

1. The 8/10 +3.7% (+$41) print came 12 days before 338 took effect, and 338 had already been announced back in July — the price rise came before the news.
2. The world is falling and only the US and Western Europe are rising: over the same period world export HRB fell −8.8% to $500/t and China −1.0% to $415/t. SteelBenchmarker's own report is titled USA and Western European HRC Prices Diverge from Global Trend. The US at $1,264 ÷ the world export price of $500 = 2.53x, a $764/t spread — while a 50% tariff on $500 is only $250. The spread is 3x the tariff arithmetic.
3. Scrap did not follow (+8%), showing this is not cost-push but end-demand pricing power.

Fundamental verification (SEC XBRL / 10-Q primary):

Valuation (TTM and forward differ by a factor of two — that is where all the tension lies):

⚠️ A 2x gap between TTM and forward is not a valuation issue, it is a price-assumption issue. The TTM figures still contain the Q4'25 trough quarter at HRC $797. Which basis you use depends entirely on whether HRC $1,147 holds.

Three single-stock traps that must be named:

• CENX's "7.6x TTM P/E" is an accounting artifact: of the $601.7M of TTM net income, $287.9M is a one-off gain from the 2026-02-02 sale of the Hawesville site to a TeraWulf affiliate (a site TeraWulf subsequently leased, about 401MW, to Anthropic on 2026-07-06, for 20 years). Ex-item the P/E is ≈15x. Q2'26 also includes $40.1M of insurance proceeds and $26.3M of 45X credits offsetting COGS. This is this brief's textbook "low-PE trap" sample.
• RS's profit is not turning into cash: receivables rose from $1,539.9M to $2,210.4M (+43.5%) while H1 revenue grew only 21%; Q1'26 OCF/net income was just 0.57x. A distributor in an HRC upcycle earns FIFO inventory gains while being drained by working capital.
• CLF's problem is not valuation, it is the repayment clock: 8 consecutive quarters of losses; net debt/TTM EBITDA 26.9x; Q1'26 operating cash flow −$325M; net assets shrank from $6,254M to $5,604M over 5 quarters. It is an option on HRC, not a fundamentals name.

Today's sell-side actions (8/24 pre-market): Wells Fargo (Timna Tanners) covered all four at once, and the direction is not uniform — NUE $285/Buy, STLD $255/Buy, RS $391/Hold (only +0.8% of room left), CLF $12/Hold; Wolfe the same day at NUE $297/STLD $275/RS $398/CLF $11; Barclays the same day reiterated Underweight on CLF with a $10 target. If this were a "tariff windfall" report, it would not leave RS only 0.8% of upside, and there would certainly not be a same-day Underweight reiteration. This looks much more like a spread-sensitivity report.

⚠️ A dirty-data warning on the ratings figures (not for the client version): in this feed the pt_old field shows a perfect mirror swap across all four names — NUE (297↔285), STLD (275↔255), CLF (11↔12), RS (398↔391). Four exact transpositions cannot be coincidence; the feed is mistaking "the previous record's PT" for "that firm's own prior value." So pt_now is usable but the "raise/cut" direction is not, and the body of this brief carries only target prices and ratings, no direction. Also, the feed timestamps the item 2026-08-24 10:35:43 while it was only 08:20 ET when the data was pulled — most likely UTC (= 06:35 ET, consistent with a pre-market publication), but this could not be confirmed.

The only possibly real 338 transmission is second-order and slow: 50% on Canadian finished vehicles → if North American vehicle assembly reshores to the US → US flat-rolled demand rises. CLF is the most auto-sheet-dependent of the group, which explains its +4.93% on Friday and its +3.11% this morning on the most real volume in the group (94,000 shares). But the flip side of that chain holds equally: if North American vehicle output simply falls net, CLF loses volume — Barclays' Underweight reiteration this morning is standing on exactly that side. This is an unverified (not obtained by this brief) narrative, not a delivered fundamental.

Final call: Watch closely, but the rationale has to be swapped out. Section 338 contributes zero to this move; what the market is paying for is the HRC spread. The top pick within the sector is NUE, which has the cleanest fundamentals and cash-flow quality (H1 OCF/net income 1.20x, net debt/EBITDA just 0.74x, H1 free cash flow about $1,054M), rather than STLD, which is up the most and is the most expensive (P/B 3.48 and EV/EBITDA 13.6, both first in the group). ⚠️ But NUE's +4.22% this morning rests on 4,026 shares; that quote does not itself constitute a fact and must be confirmed by real trading after the open.

⚠️ The risk in one sentence: anyone buying on 338 will not find a sell signal when steel prices peak, because they were never watching that signal. What actually deserves watching is the HRC print in SteelBenchmarker's next issue (#489, out around 8/26, price date 8/24) — the $1,147 I cite is already two weeks old.

---

④ Memory — every primary source is now in hand, and two of the first draft's three premises were wrong

⚠️ This section was heavily rewritten after second-round verification. Of the first draft's three claims — "Samsung published its return program today / the market read it as a capex peak / no primary source obtained" — two were wrong, and on the third my own suspicion pointed the wrong way. All corrections follow.

Related news (primaries obtained):

1. Samsung Electronics' program was released on 2026-08-21 (Friday) after the Korean close, not today. The source is the Samsung Global Newsroom official release, Samsung Electronics To Implement Largest-Ever Shareholder Return in 2026, Estimated at KRW 90 to 110 Trillion, whose text states the board approved it "today." Today (8/24) is the first and only pricing session.

• Size: 2026 shareholder returns of KRW 90–110 trillion (about $65.0–79.4 billion), roughly 5x the 2020 record of KRW 20.3 trillion; a Q3 cash dividend of about KRW 30 trillion, with the remainder deferred to a board decision in January 2027; plus approval of roughly KRW 15 trillion of buybacks, but with the stated purpose being "employee compensation," not cancellation.

2. YMTC's IPO can be traced to a Chinese-language primary: the entity is 长江存储控股股份有限公司 (YMTC Holding) (a holding platform, not the operating company), on the SSE STAR Market, accepted 2026-08-21, sponsored by 中信证券 (CITIC Securities) and 中信建投 (CSC Financial). The filed prospectus discloses RMB 33.000 billion of proceeds earmarked in total for the investment projects (RMB 20.800 billion for production-line technology upgrades + RMB 12.200 billion for R&D), ÷ 6.7291 = $4.90 billion, which reconciles with the "$4.9 billion" figure. The total raise is still blank in the filed draft (no bookbuilding yet).

• The genuinely new information in this item is not the IPO itself, it is that the prospectus cites TrendForce for the first official-document confirmation that in January–March 2026, YMTC ranked third globally and first in China among NAND makers on both a revenue and a shipment basis.

⚠️ Why the market read "record returns" as bearish — the first draft's explanation was wrong.

The first draft said "the market read it as a capex-cycle peak." That attribution does not hold, and the direction may be backwards — returning 50% of free cash flow to shareholders instead of investing it in capacity is a constraint on memory supply and therefore bullish for memory prices. Reuters' actual attribution is that both the composition and the total came in below expectations:

• Eugene Securities analyst Sohn In-joon: "Unlike SK Hynix, Samsung neither mentioned the possibility of raising its existing return policy nor announced a plan to cancel treasury shares — and cancellation is the action that would more directly support the share price."
• Morgan Stanley: "Big capital returns, slightly below expectations."
• For comparison: SK Hynix is buying back and cancelling KRW 40 trillion.

One quantitative piece of evidence that separates "company-specific" from "industry-wide": at today's KRX close, Samsung −8.70% (₩257,000) vs SK Hynix −3.41% (₩1,671,000), a 5.29pp gap, with KOSPI −3.1%. If this were industry news the two should move in the same direction and by similar magnitudes. That 5.29pp is the company-specific discount the market is assigning Samsung.

The pre-market NAND-purity gradient (stockanalysis, 08:18–08:24 ET) — this is today's cleanest piece of structural evidence:

The decline gradient correlates strictly with NAND purity. The market is not cutting indiscriminately; it is pricing structurally against the NAND supply information from YMTC. ⚠️ And SNDK was this brief's first-draft coverage gap — it fell more than MU on 519,000 shares of pre-market volume, and the first draft did not mention it once.

MU's actual numbers (SEC 8-K EX-99.1, released 2026-06-24, period ending 2026-05-28):

FQ4 guidance (same 8-K, not updated or withdrawn to date): revenue $50.0B ± $1.0B, gross margin about 86% (another ~1.4pp of expansion QoQ), diluted EPS $30.73 ± $1.00.

⚠️ One of my own suspicions was falsified and has to be written down. When assigning the verification task I specifically flagged "an 84.6% GAAP gross margin is suspiciously high, please check whether it is wrong." The result is that I was wrong; 84.6% is correct: the 8-K says in black and white "Percent of revenue 84.6%," and 35,056 ÷ 41,456 = 84.56% is internally consistent, while the four business units' gross margins of 79%–87% must weight into that range. Independent cross-check: YMTC's prospectus discloses a 2026 Q1 net margin attributable to the parent of 70.9% and a single-quarter ROE of 28.03%, and SK Hynix's TTM net margin is 85.6% — the whole industry is at this order of magnitude; 84.6% is not an outlier. (Skepticism itself also has to clear a primary source — this time I nearly threw out a correct number as a red flag.)

⚠️ But there is one line in the 10-Q far more important than the gross margin:

"DRAM products increased 343%, primarily due to a low-260% range increase in average selling prices and a low-20% percentage range increase in bit shipments."

That is: of DRAM revenue's +343% YoY, ASP contributed roughly 260pp and bit shipments only about 20pp. The overwhelming majority of current profitability rests on a price variable that can reverse within a single quarter. This explains both why an 84.6% gross margin is possible and why any supply-side news can knock this stock down 5–10% in a day.

Valuation — the textbook "low PE percentile + high PB percentile" profile, but you can only see it on forward PE:

The market will pay 10.84x for each dollar of book equity (an all-time high) but only 6.72x for each dollar of expected earnings (an all-time low) — that is the pricing expression of "the market explicitly does not believe this earnings level is sustainable." ⚠️ Convention trap: on TTM P/E (71.4th percentile) this profile is completely invisible, because the TTM denominator still contains low-profit quarters from partway up the cycle. On a cyclical whose profits are exploding higher, the TTM basis is distorted. ⚠️ Window limitation: the percentiles only go back to 2021Q4, so the 2018 memory cycle is not in the sample — read it as "the past roughly 5 years," not "ever."

⚠️ The most overlooked fact: the crack did not appear today, it appeared two months ago.

The three memory majors peaked collectively in the same week (6/18–6/25) and have since drawn down 20–43%. Today is not the day the crack appeared; it is the day MU — the shallowest drawdown of the three (22.5pp shallower than SK Hynix's) — catches down. That gap has not converged, and it matters far more than today's −3%.

Magnitude test (to stop "big news" from being mistaken for "big impact"): YMTC's RMB 33 billion of project proceeds = $4.9 billion = 0.69x one quarter of MU's capex ($7.1 billion) and 18% of MU's full-year FY26 capex guidance ($27.0 billion); and the projects are "production-line technology upgrades," not new fabs; and it is NAND, not DRAM/HBM; and from acceptance to inquiry to registration to listing to shipping output from the funded capacity there are 1–2 years. It has no direct effect whatsoever on MU's FQ4 guidance ($50B of revenue, 86% gross margin).

⚠️ A directional reminder (consistent with past lessons): IPO proceeds, equipment orders and capacity plans all speak to "the financing and preparation of future supply" — they are neither current supply nor, still less, current demand. Reading YMTC's acceptance as "NAND prices are about to fall now" compresses the timeline by 1–2 years.

Final call: Watch only. The corrected characterization is: this morning's MU −3.16% is mostly technical de-risking of a crowded trade, but it is not pure noise — the crack in the supply-side narrative is real, it just appeared in late June rather than today; all today did was upgrade a two-month-old crack from "shadow-stock rumor" to "a formal filing accepted by the SSE." The quantitative evidence for "technical": MU's 30-day annualized volatility is 94.3% (daily σ ≈ 5.94%), so this morning's −3.16% is only 0.53σ, and 10 of the last 30 sessions (33%) had single-day moves ≥5%; pre-market volume of 1.31 million shares is just 3.3% of the 20-day average (39.17 million shares), a normal level of pre-market activity, not news-driven volume.

⚠️ The decisive data this brief did not obtain: DRAM/NAND spot and contract quotes for the 3rd–4th weeks of August (TrendForce). Every pillar of the analysis above is price — MU's 84.6%, the 86% guidance, YMTC's 70.9% net margin all rest on ASP. The single decisive input for judging "technical de-risking vs. a substantive crack" is whether prices have turned, and this brief did not obtain it; the blank is left blank and not filled with an estimate.

---

⑤ MRVL — −8.5% over two days, three days before earnings; the direction cannot be extrapolated

Pre-market: −3.14%, $229.60, pre-market volume 485,502 shares (08:10 ET); Friday's regular session was −5.57%, so about −8.5% cumulative over two days. 52-week range $61.44–$329.88.

Key calendar item: FY2027 Q2 results on Thursday 2026-08-27 after the close.

Catalyst logic: MRVL is a core name in the AI ASIC/interconnect chain and is being pressed by two legs at once — the optical chain's re-rating plus position cleanup ahead of earnings. But historically the direction of a multi-day decline in the 3 days before earnings cannot be extrapolated: it could be informed money leaving early, or it could be shorts building into the print and getting squeezed after it.

Final call: Avoid (do not catch the knife into earnings, and do not chase it short either). This is not a position for "calling the direction"; it is a position for not participating.

---

⑥–⑩ Key Points on the Rest of the Top List

📌 Sub-agent status (at time of writing)

• The AAOI thread is back, and it overturned the first draft's core causality. §0①, the header box, §2 theme 1, §3B, §4, §5①, §6, §7 and §9 have all been rewritten per the verification results. Three substantive corrections: (a) "AAOI's offering crushed the whole chain" is over-attribution — over the same window COHR −5.71%/LITE −5.49%/CRDO −3.10%/FN −1.53%, the sector was already falling; (b) "cashing out at the top" does not hold, AAOI is already 51% off the $223.10 of 5/13, and H1 capex $633.5M > cash $499.7M, so $600 million does match the capacity build in magnitude; (c) the first draft missed two carry-over negatives — the Q3 guidance EPS ceiling of $0.26 is below the $0.28 consensus (present since 8/6), and "Q4 above $500 million" is not guidance but a verbal outlook on the call. The score was cut from 86 to 78 and the level from S to A+.
• That agent reports that yfinance was IP rate-limited throughout (YFRateLimitError) and unusable, and switched to SEC EDGAR XBRL companyfacts + 10-Q/8-K originals + the Nasdaq historical API + the CNBC quote API. The result is more authoritative than yfinance and is not a quality loss, but it does show the local yfinance path is still not restored (consistent with the past rate-limit record).
• The memory thread is back and overturned two of the first draft's three premises: (a) Samsung's program was released after the Korean close on 8/21, not today, and today is the first pricing day (Samsung closed −8.70% on the KRX, SK Hynix −3.41%, KOSPI −3.1%); (b) "the market read it as a capex peak" is wrong — Reuters' actual attribution, citing Eugene Securities and Morgan Stanley, is "the total and composition of returns fell below expectations, and there is no treasury-share cancellation," and the 5.29pp gap between Samsung's −8.70% and Hynix's −3.41% is the quantitative evidence of a company-specific discount; (c) the YMTC primary was obtained (SSE acceptance 8/21, RMB 33.000 billion of project proceeds), and the genuinely new information is the prospectus's first official confirmation of its global No. 3 position in NAND. Also, one of my own suspicions was falsified: I asked the agent to check "an 84.6% gross margin is suspiciously high," and 84.6% turned out to be correct (8-K original + internally consistent arithmetic + industry comparison, triple-confirmed) — skepticism itself also has to clear a primary source; this time I nearly threw out a correct number as a red flag.
• ⚠️ The memory thread also exposed a coverage gap: SNDK was −5.38% pre-market on 519,000 shares, falling more than MU, and the first draft did not mention it at all. It has been added to §2, §3B, §5④, §6 and §9.
• The steel and aluminum thread is back, and the conclusion is far stronger than the first draft's: not "mechanism questionable" but "338's direct contribution ≈ 0," and it found the real driver (the HRC-scrap spread, +61% in 10 months, SteelBenchmarker #488 primary). Three independent falsifications + AA's 10-Q capacity table (Canada at 36.3%) + Friday's EWC +0.96% are all written into §5③. That agent likewise reports yfinance returning 429 throughout, and specifically notes it was "occupied by another us-premarket process" — which suggests more than one generation process was running on this machine at once, and it is worth checking the scheduler for overlap.
• One piece of dirty data the agent found is now flagged in the text: the ratings feed's pt_old field is mirror-swapped across four names, so the "raise/cut" direction is unusable.
• risk-auditor is back and was the highest-value of the three agents. It confirmed this brief's most important counterintuitive judgment (338 excludes steel and aluminum — traced word for word to the operative provisions of Proclamation 11048 and 91 FR 54789), and also confirmed every number in AAOI's 424B5, the Jackson Hole calendar, the NVDA/MRVL dates, the 9-3 July FOMC vote with three dissents favoring a hike, and every PDD/XPEV earnings figure. But it caught 8 hard errors, all of which have been fixed in the text:

1. [High] "MU has the largest pre-market semi volume this morning" is wrong — NVDA at 2.08 million and INTC at 1.82 million are ahead; MU's 1.31 million is third. The claim was repeated in three places and was the sole reason MU made the Top 5. Fixed, with an added note on why INTC is not a signal.
2. [High] This is AAOI's 4th ATM of 2026, not its 2nd — 2/26 $250M → 3/12 raised to $500M (fully sold, 2,476,307 shares/$250.0M) → 5/14 $600M @ $223.10 → 8/21 $600M @ $129.10. "Completely unpriced" does not hold (expectation gap cut from 10 to 5), and the "management thinks the price is attractive" signal points the wrong way (the previous same-size ATM was struck 42% higher). Every filing was pulled from EDGAR originals, verified, and written into a table.
3. [High] "AAOI crushed the whole chain" lacks a base-rate test — this group has fallen together with no catalyst at least 5 times in two months (6/23, 7/2, 7/15, 7/28, 8/10), all at larger magnitudes than this morning. Test ① in §8 has been rewritten (the original test, "COHR/LITE down ≥3%," could not distinguish the two hypotheses at all).
4. [High] The header box violated this brief's own two-way thin-volume rule — it used three-digit-share quotes on BMO (226 shares), CP (213 shares) and CNI (689 shares) to declare "Canadian assets are not being sold," and 4 of the values had already drifted. It also missed the one auto-chain name with real volume, STLA (−2.22%, 415,854 shares). The whole passage has been rewritten with the error self-disclosed.
5. [High] Coverage gaps — NSSC (+19.56%, today's only big gainer on real earnings), RGNX (−22.29%), CRWV (455,000 shares), ALAB (−4.34%), AMAT/KLAC/LRCX/TER falling together (the direct test of theme 2), INTC, SMCI and POET were all absent. A dedicated §5.5 has been added to fill them in.
6. [High] XPEV also filed a 6-K the same day — the Dogotix Series A: $600 million from outside investors + $200 million from XPeng Dogotix itself, with redemption rights granted, classified as a connected transaction. I made AAOI the brief's strongest catalyst on the strength of one 424B5 and then ignored a same-day financing filing of the same type — the test was not applied consistently. §5⑦ has been added.
7. [Medium] The Treasury buybacks do not take effect until 2026-09-09, and they cover the 10–20 year and 20–30 year buckets, not the "10/20/30-year tenors" I wrote. The cash is not in the market yet; what is moving now is expectation, not flow. Verified against the Treasury release and corrected; theme 3's strength cut from A to B+.
8. [Medium] §9③'s "momentum decay is clearly visible in COIN/MSTR" is a snapshot artifact — 20 minutes later COIN was already +0.93% and MSTR +1.48%. This is exactly what this brief warned about at the top; the warning was written and then violated anyway. The conclusion has been deleted and the error self-disclosed.

• It also pointed out that §4's scoring table summed to only 51 while stating a total of 86 — the total was made up, not computed. The whole table has been redone; every column now checks cell by cell and matches §3.
• One wording point tightened at its suggestion: 338's exclusion rests on "items already covered by 232 do not apply," not on steel and aluminum being named as exempt categories; 11048's Annex II sweeps in metals/machinery/PCB, and downstream metal products not on the 232 list may still face 50%. "Steel, aluminum and copper are explicitly excluded" is right at the rolled-product level and too absolute at the derivative level.
• Two of its points this brief did not adopt: (a) it suggested cutting theme 2 to C or tagging it "mechanism needs rebuilding" — but it was reading the first draft, and the memory thread has since been rewritten by the fundamentals agent using Samsung's official release and the SSE prospectus, so the mechanism is rebuilt and the A is retained; (b) it questioned treating "AAOI's NTBV of $19.72 as a negative" as mistaking boilerplate disclosure present in every high-multiple offering prospectus for a fundamental — that point is correct, and the text now primarily cites P/NTBV of 6.3x and operating metrics, no longer using the dilution amount as a quality judgment.
• The 3 gaps it listed match the other two agents exactly, all pointing to the same conclusion: this brief's weakest spot is not its judgments but three price series none of them obtained — HRC after 8/10, DRAM/NAND spot in late August, and the actual usage progress of AAOI's May ATM.
• The three agents' shared highest-priority open items: ① whether HRC has turned since 8/10 (SteelBenchmarker #489, out around 8/26) — the sole pillar of the steel thread, and the $1,147 I cite is already two weeks old; ② DRAM/NAND spot quotes for the 3rd–4th weeks of August — the sole pillar of the memory thread, entirely unobtained by this brief; ③ the exact wording of "Q4 $500M" in AAOI's call transcript.
• Three open items the agent raised itself (worth carrying into tomorrow's recap): ① whether the exact wording on Q4's "$500M" in the transcript is "expect" or "guide"; ② whether this $600M ATM replaces or stacks on the 5/14 one (the call mentioned "$538.8M already raised via the ATM") — if it stacks, total dilution must be revised up; ③ the capitalization schedule for the $289.7M of equipment prepayments — the only hard indicator for judging "whether the capacity build earns a return."

---

⑦ XPEV — the first draft missed the second thing that happened the same day, and missed it badly

⚠️ This is a case, found on review, of "a test not applied consistently": today I made AAOI the brief's strongest catalyst because of one 424B5, and then looked straight past a financing filing of the same type on the same day.

XPeng filed a 6-K on 2026-08-24 (EX-99.1 original¹⁷, with a parallel HKEX announcement under 9868), titled Discloseable and Connected Transactions in Relation to the Subscription, the Grant of the Redemption Rights and the Adoption of the Dogotix 2026 Share Incentive Plan:

• Robotics subsidiary Dogotix completed its Series A: outside investors subscribed for 296,025,600 Series A preferred shares for a consideration of $600 million; XPeng Dogotix itself subscribed for 98,675,200 shares for a consideration of $200 million; executives subscribed for a further portion. Roughly $900 million in aggregate.
• The announcement title explicitly names "the Grant of the Redemption Rights" — i.e., investors may demand a repurchase if agreed conditions are not met, which is a contingent obligation, not pure equity financing.
• The transaction is classified as a connected transaction.

The media's actual reading also treats it as two things side by side, not just the earnings: Investing.com's headline is XPeng stock falls on Q2 results miss; robotics arm raises $900 mln.

So XPEV's −5.41% this morning (2.465 million shares) cannot be attributed entirely to delivery stagnation.

⚠️ At the same time, the first draft's presentation of XPEV's earnings figures was "cherry-picked from both ends," and the other half has to be supplied:

Every number is correct, but deliveries were framed YoY, gross margin QoQ and losses YoY — all three happened to pick the ugliest basis, making "delivery stagnation confirmed" look harder than it is. A more balanced statement is: XPEV hit its delivery guidance but growth has stalled, per-vehicle economics are improving, losses are narrowing directionally, and today's genuinely new information is that $900 million robotics financing with redemption rights attached.

Final call: Avoid (unchanged), but the rationale shifts from "delivery stagnation" to "mediocre earnings delivery + a same-day connected-transaction financing carrying a contingent repurchase obligation."

---

5.5 ⚠️ Coverage-Gap Corrections — names with real volume this morning that the first draft missed

Last Friday's recap failure was precisely a coverage gap (missing MRNA/MRK/TSLA, the largest names of the day). I rescanned about 250 tickers by pre-market volume; the following are names the first draft never mentioned that genuinely have volume. They are added here rather than pretending the brief covered them from the start.

Two methodological problems that must be written down:

① The four semicap names falling together is a test of my own theme, and I did not run that test. AMAT/KLAC/LRCX/TER all fell 2.6–3.3% this morning. If the market really were reading "Samsung's record returns = a capex peak," semicap should be the first casualty — and they are falling, which directionally supports that reading. But the evidence is dirty: the entire AI hardware chain is falling today (SMH −1.36%), and semicap falls on any AI de-risking day, so it cannot separate "capex peak" from "broad AI de-risking." Conclusion: semicap's decline is compatible with "capex peak" but does not constitute independent evidence. The genuinely clean test is in §8.

② Silver is not cooperating, which weakens the debasement leg. Gold +0.90%, BTC +1.08%, and silver −0.27%. Silver is usually the highest-beta leg of a debasement trade. Its failure to rise suggests the money buying gold and the money buying coins today may not be the same "anti-debasement" pool, but two flows with separate reasons (gold = long-end expectations, crypto = short covering). §2 theme 3's strength is cut from A to B+ on this basis.

---

6. Bearish / Avoid List

---

7. Within-Theme Ranking

Theme 1: AI optical-comms chain (bearish; ranked by strength of the negative)

Ranking logic: the closer to "pricing the optical module itself," the heavier the damage; the closer to contract manufacturing/systems integration, the lighter. FN belongs last for a reason — it is a contract manufacturer and bears no module price risk. If FN clearly outperforms AAOI today, the market is making precise distinctions between links rather than selling indiscriminately, which is evidence the theme is "orderly" rather than "collapsing."

Theme 2: Debasement trade (bullish)

Theme 3: US steel and aluminum (bullish — but the ranking basis has been switched from "tariff exposure" to "HRC spread + cash-flow quality")

---

8. Opening Verification Signals

Pre-market (before 09:30)

• Direction and volume of the leader: AAOI at 621,000 shares pre-market, −12.59%. Watch whether it is gap-and-go or gap-fill — the nature of an ATM means rallies get sold, so "opens low, bounces, but the bounce fails below $115" is the shape that confirms the theme; if it recovers straight above $120, the market is treating the $600 million ATM as "ammunition" rather than "cashing out," and this brief's main line is wrong.
• ⚠️ The pre-market volume threshold must be applied in both directions: this brief tagged NUE (4,026 shares), STLD (8,563 shares), CMC (892 shares) and FN (7,204 shares) as thin. That threshold has to apply equally to bullish and bearish names — it cannot be invoked only when volume works against my conclusion. Today's thin names include both gainers (steel) and decliners (FN, CIEN), and both sides have been down-weighted.

Intraday (first 30 minutes / full day)

Options and Sentiment

• VIX 15.92 (+5.22%, 08:10 ET CNBC) — a bounce off Friday's low of 15.13, but the absolute level is still low. ⚠️ Pre-market VIX liquidity is thin; treat this reading as indicative and do not rely on it alone.
• There are two IV events this week: NVDA (Wednesday after the close) and MRVL (Thursday after the close). Options IV on AI hardware names will keep rising from Monday through Wednesday, which means the cost of shorting this chain via options is getting more expensive, and post-earnings IV crush will hurt option holders on both sides.

Risk Notes (specific to today)

1. Gap-and-reverse: AAOI −12.6%, XPEV −4.4% and PDD +4.2% are all earnings/event gaps, and pre-market gaps filling is the norm.
2. Lone-gainer risk: PDD is the only China ADR that is up, with no sector following.
3. ⚠️ One correlation that has already broken down, and it belongs here: last Friday's recap found that miners (HUT/CIFR/RIOT/WULF/CLSK) are no longer crypto beta but AI-infrastructure beta. Confirmed again today: BTC +1.08%, while IREN −2.65%, CIFR −1.78%, HUT −1.08%, MARA −0.62%, CLSK −0.67%, RIOT −0.25%, with only WULF +1.47%. Trading "BTC up → buy miners" will lose money today.
4. Calendar overhang this week: Wednesday's core PCE (est. 3.3% YoY, unchanged) + NVDA earnings, Friday's Warsh debut. The July FOMC was 9-3 to hold, with the three dissents favoring a hike, and the probability of a hike at the 9/16 meeting was about 31.6% as of 8/20 (⚠️ that reading predates Friday's blowout PMI and will be pushed higher, i.e., this brief is citing the value that is unfavorable to its own conclusion**).** Every directional call today should be handled on a "hold ≤ 2 days" basis.
5. Data release: 08:30 ET Chicago Fed National Activity Index (July, prior −0.02) — low impact, but it lands before the open.

---

8.5 Final Reading Before the Open (08:32 ET)

⚠️ Pre-market volume and price are cumulative. The body of this brief reads at 08:00–08:24 ET; the table below is the last refresh before publication, used to confirm which conclusions are stable in time and which are drifting.

⚠️ The most important row in this table is not the biggest gainer — it is the two that did not move. NUE's pre-market volume was 4,026 shares from 08:08 to 08:32, and STLD's was 8,563 shares throughout — not a single share traded in 25 minutes. That means the +4.05% / +3.86% they display is not "sustained buying," it is the residue left by one small order a long time ago, and the quote sits still only because no new trade has come along to update it. Control group: over the same window CLF's volume rose from 94,465 to 100,535, AA's from 30,878 to 32,476, and PDD's from 920,000 to 1.004 million — those are live quotes. Conclusion: the only names in the steel theme with real trading support before the open are CLF and CENX; the pre-market gains of NUE/STLD/CMC/RS/KALU are not used as fact anywhere in this brief.

⚠️ Last print before publication (08:45 ET, 45 minutes to the open): AAOI −12.28% (706,000 shares), COHR −5.70%, SNDK −5.48% (569,000 shares), MU −3.44% (1.376 million shares), PDD +3.58% (1.087 million shares, a clear retreat from +4.66% at 08:32), XPEV −4.43% (2.851 million shares), IBIT +2.13% (7.233 million shares), SPY −0.21%, QQQ −0.65%.

Two drifts from 08:32 that have to be pointed out:

1. PDD fell back from +4.66% to +3.58% while pre-market volume kept rising (920,000 → 1.004 million → 1.087 million shares). That is real trading pushing the gap back down, not a stale quote. The verification point §9① sets for PDD (whether it breaks the opening price in the first 30 minutes) is therefore all the more worth watching.
2. NUE's pre-market volume finally moved from 4,026 to 8,883 shares, and the gain narrowed at the same time from +4.22% to +3.85%. Those few thousand new shares traded lower — consistent with this brief's judgment that steel's pre-market quotes do not constitute fact, and directionally tilted toward a pullback.

Broad market (08:32 ET): SPY −0.16%, QQQ −0.61%, DIA −0.02%, IWM −0.09%, SMH −1.36%; GLD +0.90%, IBIT +2.11%, USO −1.37%, XLE −0.28%, XLU +0.40%. The structure is exactly as in §0⑤: the Nasdaq and semis lead lower, the Dow and Russell are flat, hard assets strengthen, oil keeps falling. That structure did not change in 30 minutes, and it is the most stable judgment in this brief.

---

9. Final Conclusions

① The 5 Stocks Most Worth Watching Today

② Today's 3 Strongest Themes

③ Directions to Avoid Today, and Why

1. AI power and neocloud data centers (CRWV / BE / NBIS / IREN / OKLO / SMR / PWR) — no catalysts of their own today, purely day 4 of high-beta de-risking since 8/18. A decline with no catalyst is neither predictable nor tradable.
2. Miners (HUT / CIFR / RIOT / MARA / CLSK) and neocloud data centers (CRWV / NBIS / IREN / APLD / CORZ) — their correlation with BTC has broken down; BTC is up today and they are down. Taking a directional view on an instrument whose logic has broken down is betting on a factor you do not understand.
3. Energy and oil services (XLE / USO / XOP / OIH) — today is Iran sanctions announcement day, and oil is falling. It already rose 5% last week, so the market has clearly priced the version it expects; but the press conference has not been held and its content is unknown, so this brief does not treat "already priced" as an established fact. Neither long nor short today — this is a take no direction on an event day position.
4. The already-known hard-asset elasticity leg (COIN / MSTR / NEM) — up +8.20% / +6.10% / +3.09% respectively on Friday, so today is "do not chase," not "already weakening." ⚠️ The first draft wrote "momentum decay is clearly visible" based on the +0.24%/+0.81% readings at 08:08, and that conclusion was overturned by itself within 20 minutes: by 08:26 COIN was +0.93%, MSTR +1.48%, IBIT +2.11%. This is exactly the "pre-market volume and price are cumulative; a single snapshot is not a fact" warning at the top of this brief — the warning was written and then violated anyway. The conclusion has been deleted.
5. Any AI hardware stock within 3 days of earnings (MRVL Thursday, NVDA Wednesday) — do not catch the knife and do not chase it short.
6. ⚠️ New: any US steel or aluminum position bought on the rationale of "Canada tariffs are bullish." The rationale is wrong — steel, aluminum and copper are explicitly excluded from 338, and 232's 50% has been in force for 14 months. The sector can be bought, but the rationale has to be swapped to the HRC spread; because anyone with the wrong rationale will not find a sell signal when steel prices peak — they were never watching that signal.
7. ⚠️ New: CENX's "7.6x TTM P/E" and KALU's pre-market gain. The former includes a $287.9M one-off gain on a plant sale (≈15x ex-item), a low-PE trap; the latter's +2.56% rests on 84 shares and had already flipped to −0.50% by the time of writing.

④ Final Judgment in One Sentence

Today is neither "trade war day" nor "sanctions day" — the pre-market prices rejected both media frames (Canadian stocks did not fall; oil is falling). Today is "the day AI hardware got de-risked": optical comms, memory and AI power are being sold on all three legs while money moves toward hard assets, US steel and aluminum, and defensives; at the index level almost nothing happened (SPY −0.18%), and the entire problem is inside the index. AAOI's $600 million ATM is the hardest single point at the filing level today, but it is neither the cause of the ebb nor a "first" — it is 2026's 4th ATM, and this group of stocks has already fallen together with no news 5 times in two months at larger magnitudes. The lesson worth remembering today is this: the thing that looks most like news is often just the base rate. And all of it is happening on Day 0 before Wednesday's NVDA earnings + core PCE and Friday's Warsh debut: every position taken today is essentially a bet on Wednesday, so the holding period should be handled as ≤2 days.

---

⚠️ Risk note: 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 post-earnings IV crush and guidance reversals occur; automatically generated content may contain stale information or factual errors, so please defer to company disclosures and SEC filings, and do not use this directly as a basis for trading.