Tuesday, July 21, 2026 ALERT · Level 3 7 min read

The Rebound That Re-Inflated the Balloon

SOXX jumped 5.45%, its best day in weeks, led by memory and AMD. But fragility rose with it: the sector now sits 40.4% above its 200-day average.

The risk read for this session

Sector crash

REALISED

chips −20% from peak · within 1 month

Unchanged from the previous read

Deeper crash

28%

chips −30% from peak · within 3 months

Down 1 point from 29%

Market crisis

S&P 500 −20% · within 3 months, AI at the centre

Unchanged from 7%

SOXX

552.69

−15.7% from peak

NVDA

207.29

−12.4% from peak

VIX

17.05

fear gauge

High-yield spread

273 bp

credit stress

These are model estimates, not forecasts anyone can guarantee, and not investment advice. The dot before each probability is its band: 🟢 low · 🟡 moderate · 🟠 high · 🔴 very high, with thresholds set separately for each scenario. SOXX is the main US semiconductor ETF, the basket I track; from peak is the fall from the highest close of this cycle; VIX is the market’s fear gauge; the high-yield spread is the extra interest riskier companies pay to borrow, my early-warning line for funding stress. My internal alert is at Level 3, on a scale where 3 is the highest.

📌 SUMMARY

A broad, powerful bounce that is not yet confirmed: SOXX (the U.S. semiconductor ETF, my benchmark) rose +5.45% to 552.69, its best day in weeks, led by memory. The drawdown eased to −15.74%, back above the −20% line.
The engine was memory and AMD-Microsoft, not the mega-caps: Micron +12.2%, Sandisk +14.3%, Western Digital +12.5%; AMD +8.1% on a new Azure/Helios deal with Microsoft. Microsoft, Google, Amazon and Meta all closed lower.
Fragility rose even as price recovered: the sector is now 40.4% above its 200-day average, the fear index (VIX) fell to 17, and there are still 8 distribution days on the board with no valid follow-through (day 2 of the attempt). Regime 🟠 ALERT, Level 3 active (4th day), de-escalation 0/4.

Today the regime stays at 🟠 ALERT, and my highest internal warning, Level 3, is active for a 4th day.

New here? Start with this. I'm Andrea Pilotti, and this is AI Market Observatory. I invest my own money in AI and semiconductor stocks, and I work with artificial intelligence for a living — I built this because I needed to measure the risk in my own exposure instead of guessing at it. Every trading day I track these stocks with data and turn what I find into a single risk read: a traffic light for the odds of a sharp fall in this corner of the market. This is an observatory, not a recommendation service, and I am not a financial advisor: my job is to notice trouble early and say it in plain language. Every daily read stands on its own, this one included.

The traffic light, once: green = normalyellow = a down day, no change in the pictureorange = alert, tension risingred = critical

Alongside it I run an internal alert from Level 1 to Level 3, where 3 is the highest and only steps down once a full recovery checklist is met. How every gauge and threshold is defined is public in my methodology, and every previous daily read is in the archive.

📰 CHRONICLE (July 21)

A second bounce day, this time with muscle: SOXX +5.45% to 552.69 on volume of 10.08M — about 0.98x its 50-day average, but +54% above the prior session (6.53M), the opposite of the hollow July 20 bounce. Memory led: Micron +12.2% after Morgan Stanley flagged memory prices still rising on AI demand and called the group an attractive entry; Sandisk +14.3%, Western Digital +12.5%, Seagate +11.1%, SK Hynix's ADR +13.8% into its July 29 earnings. AMD +8.1% after Microsoft said it will deploy AMD's Helios rack systems on Azure (announced July 20) — validating AMD as an AI-infrastructure supplier, not just a chip challenger. Intel +8.6% into its July 23 report. Strong South Korean export data reinforced the AI-demand read.

But leadership was narrow — strong on the right side of the chain, absent on the mega-caps. The hyperscalers closed red (MSFT −1.1%, GOOGL −1.4%, AMZN −1.0%, META −0.3%): rotation inside tech more than a full risk-on. On the broad tape, the S&P 500 rose +0.89% to 7,509.20 (−1.32% from its record) and the VIX fell to 17.05 (−8.6%) — yet equal-weight breadth lagged (QQQE/QQQ −0.6%, RSP/SPY −0.7%). Oracle bounced +4.7% to 127.05 off the 52-week low it set July 20 (120.03).

🚦 SCOREBOARD — 9 signals (🟢 calm · 🟡 watch · 🟠 alert · 🔴 breakdown)

S1 Moving averages 🟠 (unchanged): SOXX 5th close below its 50-day line (−2.68%, but much closer); NVDA 4th below (−1.17%). Both well above their 200-day line.
S2 Structure 🟠 (up from 🔴): SOXX reclaimed the 536.12 pivot it broke on July 16 and topped the prior day's high; a pivot low is forming near 498–522. Lower highs since the 655.95 peak are not yet invalidated — a new high is needed to confirm.
S3 Momentum 🟠 (unchanged): SOXX daily MACD still negative but its histogram is recovering; the weekly bearish MACD cross is still forming (histogram improved from −4.17 to −2.35) on an in-progress bar — reconfirm Friday. NVDA daily MACD turned positive and above its signal.
S4 Breadth 🟠 (unchanged): 86.5% of the universe rose (45 of 52 U.S.-listed names) — above the 60% de-escalation bar. But index-level equal-weight breadth still lags. The multi-week divergence isn't resolved by one session.
S5 Rotation and canaries 🟢 (unchanged): my canary basket is −1.97 points relative over 20 sessions (from −2.35); rotation favored AI; 0 canaries at 52-week lows (Oracle bounced).
S6 Distribution 🟠 (unchanged): 8 distribution days active (down from 9 — the July 17 SOXX day decayed after a +5.9% rally from its close); no new distribution day. The count is still heavy.
S7 Volatility 🟢 (unchanged): VIX 17.05 (−8.6%), below 20; my short-vs-mid futures proxy eased, contango intact. Note the complacency: a low VIX with a +40.4% extension.
S8 Credit 🟢 (unchanged): my daily high-yield-vs-investment-grade gauge (HYG/LQD) sits at 0.7454, roughly flat over five sessions; high-yield spreads (HY OAS) 273 bp as of July 17 — no stress, well under the 350 bp green line.
S9 Euphoria 🟠 (unchanged): extension re-inflated to +40.4% (from +33.6%, now above 35%); weekly RSI ~61 (not extreme, which caps a full red); speculative periphery parabolic again (Nebius +18.8%). Out of the tally, but it lifts the odds.

Read (my synthesis): Family A is under tension (S1, S3 alert; S2 back to alert, so A is no longer compromised); Family B (breadth) and Family C (volume) under tension. Three families under tension, one of them a leading family, equals 🟠 ALERT. Not 🔴 CRITICAL: SOXX is +40.4% above its 200-day line, credit is calm, VIX is under 20. Not 🟢 NORMAL either — de-escalation is incomplete.

📏 LEVELS (recalculated on the July 21 close)

SOXX: reference peak 655.95 (June 22); close 552.69; drawdown −15.74%. −10% line 590.36 (active); −20% line 524.76 (now 5.05% above it); −30% line 459.17. 50-day 567.92 (−2.68%), 200-day 393.63 (+40.41%).
NVDA: reference peak 236.54 (May 14); close 207.29; drawdown −12.37%. 50-day 209.74 (−1.17%), 200-day 192.58 (+7.64%); daily MACD positive.
S&P 500: record close 7,609.77 (June 2); close 7,509.20 (−1.32%); the systemic −20% line sits at 6,087.82 (−18.93% away).

🏭 SUPPLY CHAIN BY GROUP (July 21) — universe average +5.23%

Memory +12.77% (Sandisk +14.27, SK Hynix +13.75, WDC +12.51, Micron +12.17, Seagate +11.14) — the day's engine.
Equipment +6.56% (Teradyne +12.07, Applied Materials +7.39, Lam +4.97, KLA +4.80, ASML +3.59).
Servers/hardware +6.49% (Celestica +10.49, Super Micro +7.01, Dell +5.83).
Cloud/periphery +5.96% (Nebius +18.78, CoreWeave +8.92, Applied Digital +7.90; Palantir −1.62).
Networking +5.08% (Coherent +11.15, Corning +6.08, Credo +5.56; Cisco +1.34).
Chip design +4.96% (Intel +8.64, AMD +8.11, Arm +7.46, TSMC +5.55; Nvidia +1.97, Qualcomm +1.87).
EDA +3.63% (Synopsys +2.80, Cadence +4.46).
Energy/data center +2.19% (Vertiv +4.40, Constellation +3.44; GE Vernova −0.03).
Hyperscalers +0.53% (Oracle +4.67, Tesla +2.53; Google −1.38, Microsoft −1.13, Amazon −0.98, Meta −0.32) — the group did not participate.

📊 PROBABILITIES (with ledger and change)

D10 correction: ACTIVE (SOXX −15.74%, NVDA −12.37%).
Sector drop, −20% within a month: ALREADY TRIGGERED (realized July 17); it stays "occurred" in the rolling 30-day window even though price is now back above the line.
🔴 Severe crash, −30% within three months: 28% — down one point from yesterday.
🟡 Market-wide crisis (S&P −20%, AI at the epicenter) within three months: 7% — unchanged.

Ledger for the severe-crash odds (−4.5 down / +3.5 up = −1 point):

Down: universe breadth expanding sharply (a leading-family improvement); one distribution day decayed and no new one; momentum recovering (NVDA's daily MACD turned positive); the structural buffer restored (drawdown from −20.1% to −15.7%, pivot reclaimed).
Up: rising fragility — extension now +40.4% (above 35%); an unconfirmed bounce (day 2, no valid follow-through; two rebounds already failed this month); a dense week of binary catalysts (FOMC July 28–29, hyperscaler capex July 29–30, Kimi K3 weights July 27, Intel and SK Hynix earnings).
Discipline check: I did NOT cut the odds on the rally itself or on the low VIX; the one-point reduction rests on breadth, distribution and momentum. The systemic figure is unchanged: a recovering S&P and calm credit pull down, the Iran/oil tail pulls up.

💳 CREDIT AND MACRO (July 21)

Credit calm: HYG/LQD 0.7454 (+0.24%), flat over five sessions; HY OAS 273 bp (July 17, latest available), no stress.
Rates: 10-year 4.60%, 2-year 4.21% (both as of July 20, one-day data lag); the 10y-2y curve at 39 bp.
Fed: FedWatch shows roughly 83% odds of a hold at the July 28–29 meeting (July 21 reading via aggregators of CME FedWatch) — full triple-verification not reached; slightly lower than the ~87–90% a week ago. No forward guidance: the FOMC is a two-way catalyst.
Geopolitics: 10th day of U.S. strikes on Iran; Brent around 88.9 dollars (−0.4%), easing from a one-month high on hopes talks resume. Strait of Hormuz and Red Sea risks remain.

👀 NEXT 7–14 DAYS

July 22 (today): AMD "Advancing AI 2026" product event. July 23: Intel Q2; Starship test (rescheduled).
July 27: Kimi K3 weights released — the first real architectural test of the open Chinese model.
July 28–29: FOMC (hold ~83%). July 29: SK Hynix Q2 (Seoul; HBM4 and pricing in focus); July 29–30: hyperscaler capex (Microsoft, Meta, Google, Amazon) — the single most important variable.
July 30: Samsung. August 26: Nvidia Q2 FY27.

🛡️ CAPITAL-PRESERVATION NOTE (Level 3 active)

The bounce is real and broad, but it is not yet a turn: it is day 2, with no valid follow-through, SOXX still under its 50-day line, and a heavy distribution count (8). In this group, violent rebounds inside a downtrend (July 14–15) have been sold; and the week brings four to five binary catalysts. Discipline on levels matters more than the forecast. De-escalation is 0/4: new distribution days in the last five sessions; SOXX still below its 50-day line; no valid follow-through (possible from day 4, July 23 at the earliest); breadth 1 of 3. Every decision is yours, ideally with a licensed advisor.

📚 SOURCES

Market data (adjusted closes, volumes, indicators for the July 21 session) via Massive/Polygon grouped daily and indicator endpoints; S&P 500 and VIX via FMP; Treasury yields via Alpha Vantage. News and context, with source tier noted: The Motley Fool / Seeking Alpha / Yahoo Finance (July 21 memory-led rebound — verified market fact); AMD Newsroom (Microsoft-Helios deal, July 20 — announced/official); StockTitan / 24-7 Wall St. (SK Hynix July 29 earnings; SKHY +14% — official schedule and market fact); CME FedWatch via growbeansprout / centralbank.watch (~83% hold — reported figure, triple-verification not fully reached); SundayGuardian / TradingEconomics (Brent ~88.9 dollars, Iran — verified fact).

Informational risk estimates, not recommendations. I am not licensed advisors. Over short horizons noise dominates; probabilities are orders of magnitude, and this sector can invalidate any estimate in a single session. Every operating decision is the reader's.

Who I am. I'm Andrea Pilotti. I invest my own money in AI and semiconductor stocks, and I work with artificial intelligence for a living. I built this observatory because I needed to measure the risk in my own exposure instead of guessing at it — the agents that run it every morning are something I engineered, not a company standing behind me.

What this is for. One thing only: risk management, not returns. I publish the same measurements every trading day, with the method in the open, so that anyone can see when the picture changes and decide for themselves what, if anything, to do about it.

⚖️ What this is not. I am not a financial advisor and I am not licensed. Nothing here is a recommendation to buy, sell or hold anything, and none of it is investment advice. These are calibrated model estimates, not certainties: over short horizons noise dominates, and this sector can invalidate any estimate in a single session. Every decision is yours, ideally with professional guidance.