Chip Rout First, AI Slowdown Is Not a Cycle Break
US Stocks · Options · News · Views
The tape barely broke. Semis repriced the multiple.
Monday did not hurt at the index level. SPX closed 7,620, down 0.48%. The Nasdaq finished 26,186, down 0.6%. The Dow slipped about 0.3%. The PHLX Semiconductor Index dropped about 5.9%, with all 30 members lower, from roughly 1% to double digits. SOXX fell 5.63% to 497.40 from 527.07 after a 495.15 low. Software took the other side. IGV closed 106.64, up about 5%, with cybersecurity and software volume expanding and names like ADBE finding buyers again. The seesaw is clean: sell compute, buy software that has not been replaced yet.
Rates and oil tightened the discount rate another turn. The 10-year yield tagged 5% intraday, the first print there since 2023, then settled near 4.96%. Crude faded from the session high and still held above $100: WTI near $101, Brent near $106. Futures put roughly a 90% chance on a 25 bp hike Wednesday. Banks had their own cut. At the Barclays conference, BAC said third-quarter sales-and-trading revenue looks roughly flat year over year, with investment-banking fees down at least 10%, guided to $1.6 billion to $1.8 billion versus $2.0 billion a year earlier.
A weak open is not a breakdown. QQQ opened 703.33, printed 702.74, and closed 709.18. The 702 shelf held. The damage is in chips, not the tape. The index is still sliding inside a range. Semis already ran from the upper trendline to the lower rail. Before buying the dip, separate a discounted leader from a high-multiple name that still has accelerate written into the price.
Pacing is not a halt. None of this is law yet.
The weekend twisted the pricing logic. Anthropic's Dario Amodei published We Must Pace the Frontier. OpenAI and xAI publicly backed a slower cadence. The first market read is a haircut to the AI capex story. Do not hear pace as shut it down. Slowing the frontier is not a freeze on training, and it is not zero technical progress. Recursive self-improvement is still in the speech, and a six-to-twelve-month capability jump is still the scare line. The trade still has to split slower from stopped.
The three-step plan is still a lab-and-government sketch. Step one: third-party evaluators such as METR sit inside labs with employee-level access. Anthropic says it can do that unilaterally. Step two: labs in democracies coordinate safety standards and the pace of uncontrolled progress. That needs a limited antitrust exemption from Washington. Step three is red-line talks that include Beijing: bioweapons, test mechanisms, limits on recursive improvement. A full Chinese pause is called a fantasy. Export controls stay in place to stretch the lead three to five years before any deal. Chinese foreign-ministry channels have already rejected a pause. That road does not exist.
Keep two investment sentences. Neither Washington nor Beijing will kill this business. AI continues. The variable is speed. No statute landed today. The tape killed an expectation, not rewritten purchase orders. Without rules, the market cannot re-cut next year's contract slope. It can only clip the acceleration premium inside the multiple.
The Hill is moving. Capex may want the off-ramp.
Congress did not wake up this weekend. The Klobuchar-Cruz-Thune AI framework is warm again, and reporting treats it as the viable path before 2027, possibly dropping soon. Policy is split: support the infrastructure, while the street fights large data centers, buses included. Applications are plentiful. Permits are not. The build was already lagging. The three-step lab plan and the Senate draft point the same way. This was not a weekend invention.
The ethics language is real. So is the ledger. Data-center and model spend has been running ahead of enterprise adoption. Sam Altman has already said the company misjudged how fast businesses would actually use the tools. Plenty of firms are still on the old stack. The software-replacement scene did not arrive on the lab timetable. If the capex is not strictly required, slowing development under a safety banner is a convenient step down. The labs know the debt, the targets, and the bubble risk better than the tape does. The watch item is not whether training continues. It is when growth rates get formally rewritten.
Semis take a multiple cut. The hyperscalers get relief.
For semiconductors, this is a far-dated negative, not a same-day contract void. Fast growth and ever-expanding demand sit under a lot of five-year CAGRs and multiples. If regulation slows the pace, the long-run EPS path comes down and the valuation ceiling comes with it. Signed deals still ship near term. The question is whether the next contract gets added at the old slope. Today cannot rewrite next year's table. There is no enforceable bill, only a sketch.
That is not a bubble pop. The cycle is not over. AI is not off. The fear case is disorderly expansion: invent demand, invent orders, roll credit forward. A lighter throttle at least stops one more round of storytelling. Names still bought at high multiples on a full-year high-growth print need a wider margin of safety. A down day does not make them cheap.
The four hyperscalers sit on the other side of the ledger. They are the bulk of AI capex. The second quarter just started to show some return on the spend. A slower development race cuts growth pressure, eases funding strain, and cleans up the balance sheet. Someone holding the reins and saying slow down is a mild positive, not a hit. Holding those four still makes more sense than chasing chips at the highs. If a framework lands by year-end or next year, 2027 capex can change shape. 2028 is the outer date. Cash-flow repair is a time question for them and a multiple question for high-multiple chips.
NVDA closed 210.96. Even haircut to 20 to 30 times, that still maps as not expensive against the current year through the January 2027 fiscal year. What is expensive is any name that still has accelerate forever inside the price. Semis are still buyable. The high is not a chase.
486 is intact. 107 is not. Wait for Wednesday.
SOXX failed the falling trendline twice last week and today ran from the upper rail to the lower one. It closed 497. Support at 486 is still underneath. Single-name location decides the trade. One ETF does not. If the dip looks cheap, check the name in hand first. Do not treat the semiconductor basket as a single switch.
IGV broke 101, went back to a range, and popped today on software and cyber. It closed 106.64, short of 107. A dime still counts as a miss. ORCL wears a software wrapper and is tied too tightly to OpenAI. When software rallies and Oracle does not, the mapping is symmetric: a slower OpenAI pace is not relief for that ticker. Software gets a breather because the AI-replaces-software story just got pushed out. Names tied to compute orders do not get that breather.
CTA marked 7,642 as a trigger over the weekend. SPX closed 7,620, so that print is already behind. The next acceleration sell is 7,399, then 6,902. The next-month flow sketch: plus $4.2 billion if the tape is sideways, plus $21.5 billion if it rises, minus $72.9 billion if it falls. If this tape weakens, the mechanical supply gets louder than today's session.
If QQQ loses 702 again, the pullback adds a gear. Wednesday still has the rate decision and the quarterly SEP. Inflation, unemployment, and the dots land in one packet. September's chop is not done. The index had range. The close did not change regime. If the two-way tape is unreadable, stay out of the day trade. Let the regulatory text and the dots rewrite the growth path before deciding whether chips are a discount or still a wait.