U.S. Market Recap | Tuesday, 2026-08-18 (ET)

Coverage window: 2026-08-18 09:30–16:00 ET regular session, including post-16:00 ET after-hours earnings moves (readings as of 17:02 ET).

Quote basis: open/high/low/close/volume for single stocks and ETFs are taken from the stockanalysis.com quote API (u timestamp = Aug 18, 2026, 4:00 PM EDT close basis), with key names cross-checked against the CNBC quote API; indices, VIX, Treasury yields, crude oil and the dollar index are taken from the CNBC quote API, read between 16:30–17:01 ET. Earnings figures are taken exclusively from SEC 8-K originals, links in the body. Pre-market prices carry over the 08:07–08:13 ET readings from the 8/18 pre-market list.

Methodology note for this issue (please read first): the reconciliation tables give three columns simultaneously — "vs prior close", "open→close", and "vs pre-market price". The three columns frequently produce different conclusions — for example HD closed only −0.12%, which looks like it "held up", but vs the pre-market price it is −2.02%, meaning anyone who executed at the pre-market price lost money today. Looking only at the closing change systematically overstates the hit quality of the list.

---

0. Today in One Sentence

Today was not risk-off; it was a precise sector execution: AI physical infrastructure was repriced wholesale, while software and defensives rose on the same day. S&P −0.69%, Nasdaq −1.33%, Dow only −0.22%, and S&P 500 constituent advancers vs decliners at 217:261 (A/D 0.83) — that is not the breadth of a broad selloff. What actually collapsed was one complete chain: optical modules and optical components (FN −19.38%, AAOI −15.16%, AXTI −14.23%, CRDO −13.03%, COHR −12.75%, LITE −9.87%, CIEN −8.90%, GLW −7.68%), connectors and EMS (CLS −8.76%, SANM −8.64%, JBL −8.45%, FLEX −7.51%, APH −6.61%), data-center power and thermal management (FIX −7.55%, MOD −7.42%, POWL −6.94%, GEV −6.90%, VRT −6.80%, ETN −5.29%), and memory and semiconductors (STX −9.16%, SNDK −9.01%, TER −8.77%, MRVL −7.82%, WDC −7.43%, MU −7.02%, SKHY −9.20%).

Yet on the very same day, software barely fell: IGV −0.03%, WCLD +1.11%, INTU +4.41%, ADBE +3.58%, CRM +2.71%, NOW +1.52%. Defensives were even stronger: XLE +1.76%, XLV +1.60%, XLP +1.06%, ABBV +3.43%, JNJ +3.33%, LLY +3.60%, GILD +3.25%.

The pre-market list's directional hit rate was extremely high (18/18 on the bearish side, 3/4 on watch closely), but the reason it gave was wrong. The pre-market call was "this is duration de-leveraging caused by long-end rates", yet today Treasury yields fell (30Y 5.31%→5.285%, 10Y 4.724%→4.704%, TLT +0.38%). Software and AI hardware are both long-duration, high-multiple assets — software rose while AI hardware collapsed; rates cannot explain that divergence. Tone into tomorrow: VIX 15.84 (+4.28%) still low, 10Y 4.704%, tomorrow at 14:00 ET brings the July FOMC minutes, and the pre-market brings three retail reports (TGT/LOW/TJX) plus ADI's semiconductor report; the home-improvement chain enters a consecutive verification window after HD.

---

1. Market Overview

Market breadth: S&P 500 constituents 217 advancing / 261 declining / 3 unchanged, A/D ratio 0.83.

Breadth basis note: breadth here is on an S&P 500 constituent basis (not NYSE all-market); quotes were pulled name by name across 502 constituents, of which 481 were obtained, a 95.8% coverage rate; the 21 not obtained were SW/SWK/SWKS/SYF/SYK/SYY/T/TAP/TDG/TDY/TECH/TEL/TER/TFC/TGT/TJX/TKO/TMO/TMUS/TPL/WTW. The missing names cluster in the S–T alphabetical band, so this is not a random gap, but at 4.2% of the total it is not enough to change the characterization of "more decliners than advancers, but far from a broad selloff".

Sentiment read: rotation, not risk aversion. Three mutually independent pieces of evidence: ① Dow −0.22% vs Nasdaq −1.33%, a 1.11pp divergence larger than the Dow's own decline; ② credit shows no stress at all — HYG −0.10%, LQD +0.13%, neither high yield nor investment grade moved; ③ the dollar index at +0.01% did not budge, while gold fell −1.97%. In a global flight to safety the dollar and gold should rise together; instead the dollar was flat and gold sold off hard. This rules out both the "systemic risk" and the "global safe-haven" explanations.

---

2. Pre-Market List Reconciliation

Judgment basis: "watch closely" counts as delivered if the stock closes up; "avoid / watch only" counts as delivered if the stock closes down. "vs pre-market price" = today's close ÷ pre-market price − 1, which measures the real outcome of "executing at the pre-market quote".

2.1 Bullish Side (10 names)

2.2 Bearish Side (18 names)

Also: AFRM −1.29% (open→close +2.34%, vs pre-market price +2.79%), SEZL −4.09%, NKE +2.48% (pre-market call "do not short", correct), PYPL −0.07%, CIEN −8.90%, AAOI −15.16%, CRDO −13.03%.

2.3 Hit Rate and Post-Mortem

Three points that must be written down:

① The direction was right, but the attribution was wrong — and wrong in a way today's own data directly falsifies. The core pre-market claim was "today's decline comes entirely from the discount rate rather than fundamentals... this is de-leveraging allocated by duration". Today's facts: the 30Y yield fell from 5.310% to 5.285%, the 10Y from 4.724% to 4.704%, TLT +0.38% — rates were a tailwind, not a headwind; and software, equally a long-duration high-multiple asset, strengthened against the tape (IGV −0.03%, WCLD +1.11%, INTU +4.41%, ADBE +3.58%). IGV and SMH differed by 4.06pp on the day. If this were a discount-rate shock, software could not have performed like that. The correct description: what was repriced is not "duration", it is "the AI physical-infrastructure capex narrative" — which has a boundary (software sits outside it), whereas duration has no boundary.

② Using "thin pre-market volume" to downgrade defensives treated "insufficient evidence" as "negative evidence". The pre-market list explicitly said "JNJ 2,278 shares, ABBV 4,673 shares, VRTX 681 shares... these quotes are not enough to prove money is flowing in", and therefore pushed the whole defensive group down to "watch only". Methodologically that sentence is right, but the inference ran backwards: thin pre-market volume means the direction is not yet priced, not that it will not happen. The result was JNJ +3.33%, ABBV +3.43%, VRTX +2.45%, closing gains of 3.54 / 4.23 / 2.15 times the pre-market readings respectively, with XLV +1.60% on the day. This is the most concrete opportunity cost in this issue.

But this post-mortem has a boundary and must not be generalized into "always add on thin pre-market volume": in the same group, COST (pre-market +0.96% → close +0.82%, 0.85x) and WMT (+0.98% → +0.76%, 0.78x) did not deliver. What delivered was the three pharma names (JNJ/ABBV/VRTX); what did not was the two consumer staples names (COST/WMT). The correct lesson is not "ignore thin volume" but "thin volume means unpriced; the direction has to be established from other evidence" — today pharma had sector-level resonance in LLY +3.60%, GILD +3.25%, PODD +5.96%, and staples did not.

③ HD's "beat" was manufactured by a one-time refund, and the pre-market list treated it as like-for-like comparable. The pre-market note said "$4.92 vs $4.73, a 4.0% beat, is like-for-like comparable", and listed whether the call would quantify the IEEPA refund as the first verification point. The verification point asked the right question; the answer is lethal — see the next section.

---

2.4 HD: The Answer to the First Verification Point, and How It Changes the Conclusion

The pre-market list's exact words: "Whether the 09:00 ET call quantifies the IEEPA refund amount and the 'unplanned fuel and energy cost' amount. No numbers given = the substance of the reaffirmation is in doubt."

Answer: the company gave the numbers, and the number is larger than the entire beat. Management disclosed on the call that this quarter $685M of IEEPA tariff refunds directly offset cost of goods sold, with $730M actually received in the quarter (roughly landing by end-June), and that this "is already the vast majority of IEEPA refunds filed".

Using the primary figures from the SEC 8-K income statement¹ (net sales $47,861M, cost of sales $31,746M, gross profit $16,115M, pre-tax $6,315M, income tax $1,549M → effective tax rate 24.53%, diluted shares 996M) the picture can be reconstructed precisely:

After-tax per-share contribution of the refund = $685M × (1 − 24.53%) ÷ 996M shares = $0.52/share. This is a derived calculation in this issue, based on the company-disclosed $685M and the effective tax rate from the 8-K income statement; it is not a figure the company disclosed directly.

Conclusion rewritten: this quarter HD was not "a solid clear of a low bar" but "an operational YoY deterioration masked into a beat by a one-time tariff refund". Excluding the refund, gross margin was −117bp YoY and adjusted EPS was 7.0% below consensus. That explains why the pre-market +1.95% was completely erased after the call, the close was −0.12%, and vs the pre-market price it was −2.02%. Management also said the refund "will be fully offset by unplanned costs for the full year", and that Q4 gross margin is expected to be roughly flat YoY — meaning this benefit will not repeat, while the cost pressure it offset will persist.

This is a direct input into tomorrow's LOW/TGT earnings: for any retailer that posts "a beat with an unexplained source of gross margin", the first thing to do is check for a similar one-time tariff refund.

---

3. Theme Verification

3.1 The Most Important Finding Today: What Was Repriced Is the "AI Capex Narrative", Not "Duration"

Three sets of same-day, opposite-direction facts rule out the "rate-driven" explanation:

So what was the real driver? Two verifiable, same-day pieces of information:

1. FN's gross margin guidance replaced the pricing benchmark for the entire chain (already locked down pre-market, SEC 8-K²). The new information today is its range of contagion: not just optical modules, but optical fiber (GLW −7.68%), connectors (APH −6.61%), EMS contract manufacturing (JBL −8.45%, CLS −8.76%, SANM −8.64%, FLEX −7.51%), and data-center power and thermal management (FIX −7.55%, MOD −7.42%, GEV −6.90%, VRT −6.80%, POWL −6.94%) all fell in lockstep. This is a wholesale re-rating of "unit profitability across the AI hardware supply chain", not six optical module names being dragged down together.

2. The Wall Street Journal report on large tech companies' off-balance-sheet AI commitments (published over 8/17–8/18 and widely circulated on 8/18): nine tech companies with roughly $3 trillion in combined off-balance-sheet AI-related commitments (about $1.2 trillion in leases on facilities not yet in service, $1.9 trillion in chip/energy/equipment purchase commitments), against roughly $600 billion of trailing capex. Note the mechanism: the report points at the financing structure and sustainability of this AI build-out (commitments far larger than what is on the books), not at "capex getting smaller".

⚠️ One falsification that must be spelled out: some financial sites attributed today's optical module decline to "Anthropic revenue missing expectations". On checking, that attribution does not hold. Anthropic reaching a $65 billion annualized revenue run rate was reported by Bloomberg on 8/17³, roughly 7x the year-ago level and a clear acceleration from $47 billion in May, and mainstream coverage uniformly framed it as "strong", not as a miss; "below the rumored $80 billion" is a frame that site added itself. This issue does not adopt that attribution.

3.2 Equipment vs Memory: The Falsifiable Pre-Market Forecast, and the Answer Is "It Holds"

The pre-market list wrote: "If this judgment holds, semiconductor equipment (AMAT/LRCX/KLAC/TER) should rebound faster than memory (MU/SNDK/WDC)... this is a falsifiable forecast, note it down before looking at the outcome."

The forecast holds: equipment outperformed memory by 2.50pp on the day and by 1.54pp intraday, with AMAT and LRCX even turning green intraday. But the reason it holds is not the one given pre-market (discount rate allocated by duration) — because rates fell that day. A better fit to the data is "the give-back is proportional to the prior gain":

The verification point for SKHY (SK 海力士 / SK Hynix ADR): it did not hold. The pre-market note explicitly said "what to watch tonight is whether SKHY's close holds, not that pre-market pop" — the answer is −9.20% (open→close −3.81%, full-day volume 22.37 million shares). Seoul's 8/18 close of +1.03% was indeed survivorship bias, and the pre-market warning on this point was entirely correct.

---

4. After-Hours Earnings Moves

Basis: the after-hours prices below are read at 17:02 ET, benchmarked to each name's 16:00 ET close (i.e. the pre-release baseline; all three reported after the close). The data source does not provide after-hours volume, so the after-hours percentage changes are materially less certain than regular-session ones and serve only as a directional read on tomorrow's catalysts.

4.1 KEYS — An Earnings Report in Direct Conflict with Today's Main Line

SEC 8-K Ex-99.1 original⁴, whose headline reads "Second consecutive record quarter with orders over $2 billion; full-year outlook improved".

Why this matters: KEYS's commercial communications +56% and a second consecutive quarter of orders above $2 billion say that AI data-center test demand is accelerating — and orders are a leading indicator. That directly conflicts with the wholesale selling of AI physical infrastructure today.

But the more notable thing is the market's reaction function: KEYS had already been dragged down −5.58% in the regular session before the release, and this big beat after the close only pulled it back +1.54%. Measured from the prior close, KEYS is still −4.13%. A report with non-GAAP EPS beating consensus by roughly 27% (market expectation about $2.42), record orders and raised guidance still nets out to a decline — the classic shape of "good news ignored", indicating that the current pricing pressure on the AI hardware chain outweighs single-stock fundamentals.

Falsifiable proposition for tomorrow: if KEYS can gap up, hold, and pull ANET/VRT along with it, today was sentiment overshoot; if KEYS cannot rally even on this report, the repricing of the AI infrastructure chain is not finished.

4.2 TOL — Order Growth Comes Entirely from Community Expansion, the Same Structure as HD

SEC 8-K Ex-99.1 original⁵, released 8/18, FY26 Q3 (ended 7/31).

EPS of $2.97 beat the market's roughly $2.93 by a slim 1.4%, and the adjusted gross margin of 25.6% came in 35bp above its own guidance. All full-year guidance was reaffirmed (about $10.5B revenue, 26.1% adjusted gross margin), and the FY26 buyback was raised from $650 million to $700 million.

The structure that must be pointed out: total orders +5.0% looks like demand recovering, but net signings per community fell from 5.6 units to 5.4 units (−3.6%) — the order growth comes entirely from an expansion in community count (company guidance is FY26 community count +8%~10%), not from improving demand per store. This is structurally identical to HD's "revenue +5.7% but comps only +1.7%, with roughly 70% of the growth from acquisitions and new stores": revenue growth across the entire housing chain comes from laying down more floor space, not from demand per unit.

---

5. Flows and Sentiment

5.1 Sector Rotation (SPDR Sector ETFs, sorted by today's performance)

An easily overlooked but critical detail: XLU (−0.36%) and XLRE (−0.45%) were the only two defensive sectors to fall, and they are precisely the most typical "bond proxy" assets. If today had been a rate-driven flight to safety, with yields falling XLU/XLRE should have led the gains; instead they fell, while XLV/XLP, which are more weakly tied to rates, rallied hard. This is a fourth independent piece of evidence, again pointing to "this is not a rate story". XLU's decline also carries an industry meaning: the recent rally narrative for utilities was itself tied to AI data-center power demand, and today it was sold alongside GEV/VRT/POWL — the same chain.

5.2 Style and Breadth

• Software vs semiconductors: IGV −0.03% vs SMH −4.09%, a 4.06pp gap; WCLD +1.11%. This is the cleanest style signal of the day.
• Growth and small caps: ARKK −3.18%, IWM −1.30%, Russell 2000 −1.30% — high beta under pressure, but far less so than the AI hardware chain.
• Thematic ETFs: XSD (semicap-weighted) −5.75%, PAVE (infrastructure) −2.42%, BOTZ (robotics/AI) −3.36%, XME (metals and mining) −3.78%, ITB/XHB (homebuilders) −1.59%/−2.07%.
• Breadth: S&P 500 217 advancing / 261 declining (A/D 0.83). 16 of the 20 biggest decliners belong to the AI physical infrastructure chain; among the top 20 gainers, 5 are pharma (LLY/ABBV/JNJ/GILD/PODD) and 4 are software (INTU/ADBE/TYL/GDDY).

5.3 Characterization

Neither risk-on nor risk-off is accurate; today was "an asset switch within themes". Four independent pieces of evidence support this characterization: ① no credit stress (HYG −0.10%, LQD +0.13%); ② the dollar did not move (+0.01%) while gold sold off hard (−1.97%), inconsistent with safe-haven demand; ③ long-end yields fell and TLT rose, inconsistent with a rate shock; ④ bond-proxy defensives (XLU/XLRE) fell while cash-flow defensives (XLV/XLP) rose. VIX at 15.84 was +4.28% but still low — the market is not treating this as a systemic event, it is simply moving money out of one theme and into two others.

---

6. Outlook for Tomorrow (2026-08-19, Wednesday)

6.1 Theme Persistence

6.2 Tomorrow's Calendar (ET)

8/20: BABA earnings (the falsification point for today's +2.76%). 8/26 after the close: NVDA earnings — the entire AI chain remains in a fundamentals vacuum until then.

6.3 Names to Watch (Ticker + Verification Point)

Write the verification point first, then talk about the conclusion.

6.4 What to Avoid

1. Do not bottom-fish any name in the AI physical infrastructure chain. Today is Day 1 of diffusion, and most names closed near the day's lows (FN −6.06% intraday, AAOI −10.12% intraday, CRDO −7.85% intraday); the decline has not yet shown a high-volume stabilization pattern.
2. Do not treat "falling rates" as a reason to buy growth. Both the 30Y and 10Y fell today and the AI hardware chain sold off regardless — that transmission channel is currently disconnected.
3. Do not put on directional macro positions ahead of the FOMC minutes (14:00 ET).
4. Always check retail earnings "beats" for one-time items first. The lesson from HD: a $685M tariff refund manufactured the entire beat, and that number exists only in the call, not in the 8-K.
5. Still no participation in KLAR / FN / BIDU. The three fell −22.81% / −19.38% / −12.73% respectively today and have all entered a price-discovery phase with no fundamental anchor.

6.5 Input Notes for Tomorrow's Pre-Market List

1. Suggested rewrite of the theme structure: replace "duration de-leveraging" with "repricing of the AI physical infrastructure capex narrative", and state its boundary explicitly — software sits outside it (IGV −0.03% vs SMH −4.09%). That boundary is the first filter for tomorrow's screening.
2. Tracking list additions that are mandatory (completely uncovered in today's pre-market): GLW, APH, JBL, FLEX, CLS, SANM, AXTI, VRT, GEV, FIX, MOD, POWL, ETN, MRVL. 16 of today's 20 biggest decliners belong to this chain, and the pre-market list covered only 6 of them, all optical modules.
3. Stop using pre-market volume to downgrade defensives. Today proved it: thin pre-market volume on defensives = not yet priced, not that it will not happen. The suggested change is to confirm with the first 30 minutes of post-open volume rather than veto on pre-market volume.
4. Add one step to the retail earnings verification process: 8-K figures → (mandatory) one-time items from the call → compare against consensus only after stripping them out.
5. Keep and reinforce the "verification point first" style. Four of today's 5 verification points produced clear answers (HD's IEEPA, FN's industry-wide question, BABA's KWEB follow-through, SKHY's close), and two of them — HD and FN — directly rewrote the conclusions; a verification point carries more information than the conclusion itself.

---

Data Retrieval and Execution Failures (Internal)

• stockanalysis batch quotes hit a whole-batch rate limit on the third batch: all 27 names including MRVL/GLW/APH/JBL returned FAIL at once (contending for the same endpoint as the concurrent S&P 500 breadth scan). That entire batch was re-pulled via the CNBC quote API, so the open/close/volume for those names in the body are on a CNBC basis, interleaved with names on a stockanalysis basis; the two routes were cross-checked on FN/COHR/LITE/GLW and the closing prices and percentage changes matched exactly.
• Coverage gap in the S&P 500 breadth scan: only 418 of the 502 constituents were obtained on the first pass, and the 84 failures clustered in the S–T alphabetical band (a classic one-sided bias, not a random gap). After retrying with reduced concurrency the count was filled in to 481 (95.8%), still missing 21 (SW/SWK/SWKS/SYF/SYK/SYY/T/TAP/TDG/TDY/TECH/TEL/TER/TFC/TGT/TJX/TKO/TMO/TMUS/TPL/WTW). The body discloses the coverage rate and the missing list. Had the first-pass result of 418 names been used directly, A/D would have been 0.833 rather than 0.831 — the characterization is unchanged in this instance, but this bias pattern must be checked explicitly every time.
• The stockanalysis screener API (/api/screener/s/f, /api/screener/a/f) still returns 404 (consistent with the 8/18 pre-market issue), so the all-market movers scan is unavailable and breadth instead uses a self-built basis of name-by-name quotes across S&P 500 constituents.
• The after-hours volume field (epv) returned null for all three names, so the after-hours percentage changes for KEYS/TOL/ZTO cannot be assessed for reliability via volume weighting; the body carries a basis warning on this.
• Direct SEC access requires an explicit User-Agent and must use --compressed: the first WebFetch of the TOL 8-K returned 403; after adding a UA, curl still returned 0 bytes; after adding --compressed it worked (23KB). Keysight's 8-K was located via data.sec.gov/submissions/CIK*.json to find the accession, after which Ex-99.1 was retrieved successfully.
• Three pieces of false information identified and discarded:
  1. "TOL fell 4.9% after hours, EPS $3.73 beating expectations of $3.60" — the search summary had stitched together FY2025 Q3 figures ($3.73 is precisely the year-ago EPS). The actual FY2026 Q3 EPS was $2.97, with after-hours +0.10%. A textbook case of "a search summary conjuring an earnings report out of thin air".
  2. "FN closed at $477.37, down 20.25%" (24/7 Wall St.) — stockanalysis and CNBC independently agreed on $482.59 / −19.38% (prior close $598.58); the media figure was not used.
  3. "Optical modules fell because Anthropic revenue missed expectations" — Anthropic's $65 billion annualized revenue was reported by Bloomberg on 8/17, framed as roughly 7x YoY and an acceleration from $47 billion in May, and mainstream coverage uniformly characterized it as strong; "below the rumored $80 billion" is a frame that site added itself. This attribution has been explicitly falsified in the body and was not used. The same piece also inverted the meaning of the WSJ off-balance-sheet commitments report (writing it as "capex is smaller than expected", when the report actually says the commitments are far larger than book capex); the body rewrites it per the original meaning.
• WSJ $3 trillion off-balance-sheet commitments report: this issue obtained only secondary retellings (Seeking Alpha, TipRanks, cryptobriefing and others, all consistent), and did not obtain the WSJ original (paywall), so the body describes it on a "as reported" basis, and the amount breakdown ($1.2 trillion leases / $1.9 trillion purchase commitments) is a secondary retelling.
• HD's IEEPA $685M comes from retellings of the call (consistent across multiple financial outlets), not from the 8-K original; the 8-K contains only a qualitative statement that "the guidance includes IEEPA tariff refunds" with no amount. The gross margin/EPS reconstruction based on that amount is a derived calculation in this issue (the inputs — 24.53% tax rate, 996M shares, $31,746M cost of sales — are all primary from the 8-K income statement), and the body labels it as derived. If the media retelling of $685M is wrong, the conclusions in that section need to be recomputed.
• Items not obtained: ① NYSE all-market advance/decline counts (the endpoint has no such symbol, so the S&P 500 basis was used instead); ② KEYS's absolute Q3 order figure of $2.091B, seen only in secondary sources — the 8-K original says only "orders over $2 billion", and the body uses the original's wording; ③ consensus estimates for KEYS/ADI are secondary retellings ($2.42), and where used to compute the size of the beat they are labeled "approximately".
• No risk-auditor QC was run: this issue explicitly labels every derived value, secondary source and unclosed assumption inline in the body (the derived-calculation note in Section 2.4, the Anthropic falsification box in Section 3.1, the missing after-hours volume warning in Section 4, the basis note in Section 5.1, the breadth coverage rate in Section 1). This is a fallback, and it is not equivalent to independent QC.

---

⚠️ Risk disclaimer: this recap is only a post-close review of information and observations, and does not constitute investment advice. Data may differ in timeliness or basis; please refer to company disclosures and SEC filings as authoritative, and do not use this directly as a basis for trading.