AI MARKET OBSERVATORY · Blog
Tuesday, August 4, 2026 CRITICAL · Level 3 14 min read

A rally clears its first real test. My alert doesn't move.

Chip stocks logged their widest rally of the cycle and a confirmed follow-through day on August 4. My risk regime stays at its maximum for a 14th session.

The risk read for this session

Sector crash
25%
chips −20% from peak · within 1 month
Deeper crash
29%
chips −30% from peak · within 3 months
Down 31 points from 60%
Market crisis
5%
S&P 500 −20% · within 3 months, AI at the centre
Down 4 points from 9%
SOXX
542.21
−17.3% from peak
NVDA
211.94
−10.4% from peak
VIX
16.50
fear gauge
High-yield spread
278 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.

📌 The one-line takeaway

On Tuesday, August 4, SOXX — the main US semiconductor ETF, the benchmark I track every day — rose 6.80% to 542.21, on volume just under its own 50-day average. That single fact undersells the day. 92.3% of the 52 US names I follow rose — the widest one-day gain in this whole seven-week decline — and a specific technical signal called a "follow-through day," which confirms that a rally attempt has real backing rather than just a bounce, fired for the first time since this decline started. My drawdown improved from 23.08% below peak to 17.34%, and SOXX closed back above the level that defines a 20% sector fall, which I had been counting as already realized since July 27.

None of that moved my risk regime. It stays at 🔴 CRITICAL, the top of my four-step scale, for a 14th straight session. My rule for stepping down is a four-part checklist, evaluated over the last five sessions, and it needs all four boxes ticked before the regime can fall even one level. Today ticked a new box — the follow-through day — but a box that was ticked two weeks ago came untied at the same time, so the count is still 1 of 4.

A new source of concern opened on the same day the technical picture improved: Oracle was downgraded by S&P to BBB-, one notch above speculative grade, and CoreWeave's bond market is still pricing meaningful default risk. Neither has shown up yet in the broad measure of credit stress I track daily, but both are AI-linked companies, and my rules count a downgrade on a name like that as a warning sign in its own right.

One more thing worth saying plainly: three trading sessions — Friday, July 31, Monday, August 3, and today — passed between my last full reading and this one. I am publishing today's numbers with full continuity on the technical counters that depend on it (support levels, the tally of heavy-selling days, the drawdown path), but I did not write a full day-by-day account of the two sessions in between. That is a deliberate choice, not an oversight, and I am saying so because this project's whole premise is that I don't hide gaps.

📰 What happened

The broad market, not any single AI headline, set the tone. The S&P 500 closed at 7,736.52, up 1.79% — a fresh record high, its first in two months. The Nasdaq Composite rose 2.59% to 26,584.99, and the Dow Jones Industrial Average rose 1.71% to 54,085.88, crossing 54,000 for the first time. The move coincided with a sharp drop in oil prices — Brent crude fell more than 5% on the day and more than 10% over the week — after the US Treasury Secretary suggested a possible easing of tensions around the Strait of Hormuz, mediated by Qatar, following a reported delay of US military action against Iran. I treat that link as a market reading, not a proven cause.

Three company-specific stories mattered inside the sector. Palantir rose 26–30% (sources disagree on the exact figure) after reporting revenue up 93% year over year and raising its full-year guidance to $8.15–8.16 billion from $7.65–7.66 billion — a single stock that on its own moved the whole 52-name average by roughly half a point. AMD beat on both revenue ($11.5 billion, up 50%) and profit, with data-center revenue more than doubling, yet the stock fell 8.9% after hours — read by the market as a worry about how much investors are already paying for AI growth, not about the results themselves; that reaction will show up in Wednesday's opening price, not today's close. And Oracle extended a two-session bounce (up 9.2% on Monday, further gains today) that began after a credit downgrade, not before it — a reminder that a rising share price and improving credit health are not the same thing.

The two sessions in between, briefly

Between my last full reading and today, the reactions to Amazon's and Apple's earnings — both already known to be moving sharply after hours as of my last report — played out. Amazon confirmed and extended its jump, closing Friday, July 31 up roughly 15%, on revenue of $200.6 billion and AWS growth of 36.7%, its fastest in eighteen quarters. Apple confirmed its decline and never recovered it: down roughly 7–8% on Friday on weak guidance tied to memory supply constraints, then a further 1.78% on Monday, the only large technology stock in the red that day while its peers rose 4–6%.

Nvidia, the sector's flagship, closed at 211.94, up 2.56%, and is now back above both its 50-day and 200-day averages for the first time in weeks.

The rally was unusually broad and unusually calm. Volume on SOXX was just under its own 50-day average — 0.99 times, not above it — which matters because most of this decline's sharp reversals came on heavy volume. Today's did not. Forty-eight of 52 names rose; the four that fell were Amazon (digesting Friday's jump), Constellation Energy, Meta, and Vistra, down 8.15% with no company-specific news I could identify — treated as an idiosyncratic move, not a sector signal.

The rule versus the read

Here is the tension I want to be explicit about, because it is the most important thing about today's reading. If you only looked at today's numbers — the widest rally of the cycle, a confirmed follow-through day, a full recovery of Nvidia above its trend lines, the S&P 500 at a record — you would reasonably conclude the worst of this is over. My system does not reach that conclusion, and the reason is not stubbornness: it is that my rule for lowering the alert level requires four specific, pre-committed conditions to hold at once over the last five trading sessions, and only one of them does. A single heavy-selling day from eleven days ago is still inside that five-session window. It falls out of the window at my next reading. Until then, the rule and the read disagree, and I follow the rule.

🚦 The signal board — my 9 gauges in 6 families

I track nine gauges grouped into six families. Each is 🟢 calm, 🟡 watch, 🟠 alarm or 🔴 confirmed breakdown.

FamilyGaugeReading
A — Price trendMoving averagesWatchSOXX is 4.50% below its 50-day average (a 15th session below it), narrower than 11.10% two sessions ago — but today's volume was not above average, so the alarm-level bar isn't literally met. Nvidia is back above both of its own averages.
A — Price trendStructureWatchA meaningful reversal: today's close retook a support level broken on July 27, and today's high cleared a lower-high pattern that had held since late July. Not yet a confirmed repair — one session isn't proof — but the two conditions that had this gauge at its worst reading are both gone.
A — Price trendMomentumWatchA standard daily strength reading on SOXX is back above the neutral 50 mark (50.79, from 42.5), and the daily trend-strength line is close to crossing back to positive. Still negative on the weekly timescale.
B — ParticipationBreadthWatch92.3% of my names rose today, the widest in this decline — but the equal-weight-to-cap-weight ratio I use to check for hidden mega-cap dependence has been falling over the last five sessions, and my weekly measure of how many names sit above their own 50-day average (weak at last count) isn't due for its update until Friday. Mixed evidence, so I'm not calling this fully repaired.
B — RotationRotation and canariesCalmMy five early-warning names beat SOXX by 10.33 points over 20 sessions, none is at a 52-week low, and SOXX outgained the Nasdaq 6.80% to 2.59% — money is still rotating toward chips, not away.
C — VolumeDistributionAlertThe count of heavy-selling days in the last 25 sessions fell from 15 to 11 — two expired on each of SOXX and the Nasdaq-100 proxy after rallies of 5% or more from their trigger date — but 11 is still well above the level that keeps this gauge at alarm. No new one appeared and no support broke today.
D — VolatilityFear gaugeCalmThe VIX, the market's fear gauge, fell to 16.50. Flagged as complacency, and more so than at any point in this cycle: a reading this low alongside a 33.79% extension above the 200-day average and a record-high S&P 500 is a calm signal worth distrusting a little, not trusting more.
E — CreditCredit tapsAlertThe broad ETF-based measure of credit stress I check daily is flat. But S&P downgraded Oracle to BBB-, one step above speculative grade, and CoreWeave's five-year default-risk pricing was still elevated as of late July. My rules count a downgrade on an AI-linked company as a warning sign on its own, regardless of what the broad measure shows.
F — FrothOverextensionWatchSOXX is now 33.79% above its 200-day average, the highest of this entire cycle and one step from my most extreme reading. The weekly strength measure isn't extreme yet, so the letter of my rule keeps this at watch rather than alarm — but this is the closest that rule has come to being tested since I started publishing.

Read together: two families that had been under pressure — trend and participation — are now calm or nearly so for the first time in weeks. Two others — distribution and credit — are not, for different reasons: one is a slow-healing count of past selling, the other is a brand-new, company-specific credit story that the broad numbers haven't caught up to yet.

📏 The levels that matter now

  • SOXX closed at 542.21, 17.34% below its June 22 peak of 655.95, against 23.08% three sessions ago. It closed back above 524.76 — the level that marks a 20% fall from peak — for the first time since July 24, which is why I'm no longer counting that event as active (more on this in the odds section below).
  • SOXX against its averages: 4.50% below the 50-day (567.77), narrower than 11.10% before; 33.79% above the 200-day (405.27), the widest gap of the cycle.
  • Nvidia closed at 211.94, 10.40% below its May peak of 236.54, against 17.54% before. It is now back above both of its own moving averages.
  • The S&P 500 closed at 7,736.52 — a new record high, replacing the previous reference peak of 7,609.77 set on June 2. Because of that, my systemic-risk threshold recalculates upward, to 6,189.22, and there is no current drawdown on the broad index to measure against chip stocks' decline. The sector remains the clear source of stress; the broad market just isn't sharing in it today.

🏭 The sector, group by group

The average move across the 52 US names I track was +5.69%.

  • Speculative cloud names led (+9.10% on average), driven mostly by Palantir's 26–30% jump; networking (+8.00%) and server hardware (+7.84%) followed close behind.
  • Chip design and equipment groups were both strong (+7.67% and +6.96%), broad-based rather than led by one name.
  • The one group that fell was power and data-center infrastructure (−1.05% on average), dragged down by Vistra's 8.15% decline with no clear cause I could identify; the companies that install AI computing capacity remain the weakest link in the chain, a pattern that has now held for more than a week.
  • Hyperscalers were the softest of the groups that rose (+0.83%), held back by Amazon giving back part of Friday's surge.

📊 Today's odds

These are model estimates of specific, defined events. The dot in front of each is my band: 🟢 low · 🟡 moderate · 🟠 high · 🔴 very high.

  • 🟠 A 20% fall in chip stocks within 30 days: 25%. This needs an explanation. I had been reporting this as an already-realized event since July 27, because SOXX had closed below the 20%-down level. Today it closed back above that level for the first time since July 24, so I'm no longer counting it as realized — I'm re-estimating it as a fresh probability instead. SOXX sits only 3.33% above that line today, and single sessions of 5–8% are common in this stretch, so I don't read this as a resolved risk, just a differently-measured one.
  • 🔴 A 30% fall in chip stocks within 90 days: 29%, down sharply from 60% at my last reading, because three sessions of recovery — not just today's — pushed the distance to that level from 8.99% away to 18.09% away. I also weighed the new credit-related concern on Oracle and CoreWeave, and the fact that my broadest measure of overextension is at its highest point of the cycle, both of which argue for keeping some of the improvement in reserve rather than passing all of it through.
  • 🟡 A 20% fall in the broad US market within 90 days, with AI at the center: 5%, down from 9%, because the S&P 500's new record high resets the reference point my systemic estimate is measured against.

A 31-point drop in the 90-day estimate over what amounts to three sessions is a large move, and I want to be direct about why I'm allowing it: it isn't from one good day, it's the sum of a real, multi-session recovery, a confirmed technical signal that had never fired before in this cycle, and a level I had counted as breached going back to being un-breached. I would rather publish an estimate that tracks the facts honestly than one that changes slowly for the sake of looking steady.

💳 Credit and macro

The broad measure hasn't moved, but a specific one has. The ETF-based ratio I check every day for credit stress is essentially flat. The high-yield spread I check weekly is unchanged from last Wednesday's confirmed reading of 278 basis points, with the next check due Friday. What has moved is company-specific: S&P downgraded Oracle's credit rating to BBB-, one notch above speculative grade, with its five-year default-insurance cost at levels not seen in roughly 18 years; the exact date of the downgrade wasn't precisely stated in the sources I found, so I'm flagging that uncertainty rather than pinning a date on it. CoreWeave's bond market was, as of late July, pricing a roughly 50% chance of default within five years, even as its stock jumped on a refinancing announcement. A rising stock price does not resolve a credit concern; those are two different markets pricing two different things, and right now they disagree.

Ten-year and two-year Treasury yields, as of August 3 (one session behind, as usual for this data), stood at 4.70% and 4.25%, with the 2s10s curve at 45 basis points, unchanged from two readings ago. There is no Federal Reserve meeting before September 16 (still to be confirmed), and I found no significant Fed commentary in the past week. On what markets expect from that meeting, I have nothing reliable to publish: the readings I found ranged from roughly 80% odds of a rate increase to roughly 90% odds of a rate cut — a spread far too wide to be normal disagreement, and consistent with the genuine three-way split inside the Fed that showed up in its July 29 vote. I'd rather publish nothing than pick a number that wide of a range.

👀 The next 7-14 days

  • Wednesday, August 5: the market's first chance to react to AMD's after-hours decline.
  • Friday, August 7-8: the US July jobs report — the first significant macro data point since the Fed's July 29 decision — plus my weekly check on the flight schedule, the credit spread, Oracle and CoreWeave bond data, and how many of my 52 names sit above their own 50-day average.
  • Around August 10: TSMC's July revenue, the most timely hard number in the supply chain.
  • August 11: Super Micro Computer's quarterly results.
  • August 26: Nvidia's quarterly results.
  • August 27: Dell's quarterly results.
  • September 16 (to be confirmed): the next Federal Reserve meeting.
  • Watch item, technical: today's follow-through day is one box of four on my recovery checklist. The next most likely to clear is the "no new heavy-selling day" condition — it fails today only because a single day from July 29 is still inside the five-session window I check; if nothing new appears, that box can tick at my next reading.
  • Watch item: whether the Oracle downgrade and CoreWeave's credit stress spread to other AI-linked borrowers, or stay contained to these two names.

🛡️ Where this leaves me

My regime is 🔴 CRITICAL and my internal alert stays at LEVEL 3, the highest, for a 14th session. Today is the clearest case yet of why I built this as a rule instead of a daily judgment call. If I were deciding by feel, a day like this — the broadest rally of the decline, a confirmed technical signal, Nvidia fully recovered, the S&P 500 at a record — would probably have talked me into easing off. My checklist doesn't move on a feeling. It asks four specific questions about the last five sessions, and today the honest answer is that only one of them is yes.

I want to be equally honest about the other side: the technical picture is genuinely, substantially better than it was three sessions ago, and pretending otherwise would be its own kind of dishonesty. The distance to a 30% decline nearly doubled. A confirmed rally signal, absent for the entire decline, finally appeared. What's keeping the alert at its ceiling is not a judgment that today's rally is fake — it's that my rule was built to require more than one good day before trusting a bounce, and I am one good day away from most of it clearing, not many.

📚 Sources

↑