AI Leaders Called for a Slowdown. What the Capex Filings Actually Show
🧠 AI Leaders Called for a Slowdown. What the Capex Filings Actually Show
Welcome back, everyone! 📊 I’m your Quant Analyst, filtering out market noise using data, statistical modeling, and systematic insights. 👩💻✨
Last weekend produced the strangest AI headline of the year: the people building frontier models publicly asked the industry to build them more slowly. Semiconductors sold off. To cut straight to the chase, after reading the filings rather than the reaction: what was actually proposed was a limit on capability advancement, not on spending — and the last reported quarter shows hyperscaler capex accelerating, not slowing.
▲ The debate is about how fast models get smarter. The capital is committed to how much compute exists. Those are not the same variable
📌 First, the actual numbers
On September 12, 2026, Anthropic CEO Dario Amodei published an essay, “We Must Pace the Frontier,” arguing the industry should slow “the pace at which we improve the capabilities of AI models.” OpenAI’s Sam Altman (“I agree with Dario that we need to pace the frontier”) and xAI’s Elon Musk (“Dario is right”) publicly backed it. Meanwhile, from the companies’ own 10-Q filings, combined quarterly capital expenditure at Microsoft, Alphabet, Amazon and Meta reached $165.1 billion in the June 2026 quarter — up 27.2% on the prior quarter and 87% year over year. First-half 2026 total: $294.9 billion.
Sources: TechCrunch and Axios on the essay and endorsements; capex computed by me from SEC EDGAR XBRL filings.
1. Read what was actually proposed
The three steps in the essay are: embedded third-party evaluators with real access to frontier labs; coordinated safety standards among leading companies in democratic countries, with government mediation on antitrust; and an attempt at international coordination including with China on the most dangerous capabilities.
Notice what is not in that list. There is no proposal to reduce data-centre investment, no proposal to stop deploying existing models, and no proposal to cut compute purchases. The target is the rate of capability improvement — which is primarily a statement about frontier training runs and how they are evaluated before release.
That distinction is the entire investment question, because training and inference sit on very different parts of the capex line.
2. Is the money slowing?
I pulled quarterly capital expenditure for the four largest buyers straight from their filings. Amazon reports true 90-day figures; the others require differencing year-to-date cash-flow periods, which is how you get the fiscal-quarter gaps most summaries have.
Combined Hyperscaler Capex, Eight Quarters
*$ billions per calendar quarter, my calculation from SEC XBRL cash-flow data. Microsoft’s fiscal year ends in June, so its quarters are aligned to the calendar here.
From $58.9 billion to $165.1 billion in eight quarters — up 180%. The most recent quarter is the largest sequential jump in the series. Whatever is being debated in essays, the committed capital was still compounding as of June 30.
The caveat that matters more than the chart
These filings cannot possibly reflect the news. The latest data ends June 30, 2026; the essay was published September 12. Nothing here is evidence that the companies will or will not respond. What the chart gives you is the starting point — and therefore the size of what would have to change. A spending line growing 27% a quarter does not go flat quietly; you would see it in guidance first, and the September-quarter reports are the first place it could appear.
3. Why semiconductors fell harder than the hyperscalers
The market’s reaction looked lopsided, and it was — but it was also correct, for a reason you can quantify. Compare Nvidia’s quarterly revenue with the same four companies’ capex:
A Remarkably Stable Coupling
*Nvidia’s fiscal quarters end in late January, April, July and October, so each point is matched to the nearest calendar quarter — an approximation, not an exact overlay. Two quarters are missing where Nvidia reported no standalone fiscal-Q4 figure.
Over eight quarters in which both series grew roughly 175–180%, the ratio stayed inside a band of 53% to 63%. Nvidia’s revenue and these four companies’ capital spending have moved almost in lockstep.
That is the transmission mechanism, and it explains the asymmetry precisely. For a hyperscaler, cutting capex is a cost saving — free cash flow goes up on the day of the announcement. For Nvidia, the same decision is a revenue cut. Identical news, opposite sign, which is why one layer sold off harder than the other.
4. Where I think the market may be misreading it
Here is my actual view, stated as a claim I could be wrong about. The essay is about training; a large and growing share of this capex serves inference. Pacing the frontier means fewer or slower capability jumps — it does not mean fewer people using the models that already exist. If anything, a slower capability cycle lengthens the useful life of deployed models, which makes inference demand more predictable, not less.
So a literal reading of what was proposed touches the part of the spend tied to frontier training runs, and leaves the part tied to serving users largely alone. The market appears to have priced it as a cut to the whole line.
Three ways I could be wrong about that
One: the training/inference split is not disclosed. Companies report capex as one line, so my claim that inference is the larger and more durable share is an inference of my own, not something I can source to a filing. Two: the 59% coupling ratio is a ratio of two aggregates, not a customer-concentration figure — Nvidia also sells to neoclouds, sovereigns and enterprises, and these four also buy silicon that is not Nvidia’s. The stability is striking but it does not license a statement about who buys what. Three: expectations, not spending, set prices. Capex could keep rising while the multiple on it compresses, and holders of semiconductor stocks would still lose money. “The spending continued” is not the same claim as “the stocks were cheap.”
5. What I would actually watch
- Capex guidance in the October reports, not the capex itself. Spending is contracted years ahead and turns slowly; guidance turns immediately. The September-quarter calls are the first moment a response could surface.
- Whether anyone quantifies the commitment. Three CEOs agreeing on a principle costs nothing. A published evaluation threshold, a named third-party evaluator with access, or a stated delay between capability milestones would be the first evidence that this changes behaviour rather than tone.
- The coupling ratio itself. If Nvidia’s revenue starts falling below ~50% of combined hyperscaler capex, it means the spend is rotating toward custom silicon, networking, power and buildings — a far more important structural story than any one quarter’s growth rate.
- Whether the non-signatories move. The agreement covers frontier labs in democratic countries. If capability progress continues elsewhere at the same pace, the competitive logic that produced this spending does not change, whatever anyone has signed.
For the layer-by-layer version of who depends on whom, my earlier piece on what Nvidia’s growth actually says about every other chip stock works through the same transmission problem from the semiconductor side.
Quick FAQ
Q. Is this the top for AI stocks?
I can’t tell you that and I’d be sceptical of anyone who says they can from one weekend of headlines. What I can say is what would have to be true for the bear case: hyperscaler capex guidance would need to come down in October, and the coupling ratio says semiconductor revenue would follow it closely. That is a checkable prediction with a date attached, which is more useful than a view.
Q. Doesn’t a safety agreement among competitors just get broken?
Possibly, and the essay concedes as much by asking for government mediation on antitrust and for international coordination. The investable point is narrower: an agreement that is never quantified cannot be verified, and cannot be modelled either. Until a number is attached, it is a sentiment event, not a cash-flow event.
Q. Should I be reading CEO essays for investment signals at all?
For the direction of policy, yes. For the direction of spending, the cash-flow statement is a far better source and it arrives on a schedule. This post exists because those two things said different things in the same week.
💡 Quant Strategy & Takeaways
What was proposed is a limit on capability advancement, not on compute spending. Through June 2026 the four biggest buyers were still accelerating — $165.1bn in a single quarter, +27% sequentially. Nvidia’s revenue has run at 53–63% of that capex for two years, which is exactly why a spending scare is a cost saving for one layer and a revenue event for the other. Watch October guidance, not September headlines.
When market volatility spikes, remove emotion and focus strictly on the numbers! 🤖
Do you read a safety agreement as bullish or bearish for the companies selling the compute? Let me know in the comments! 📈✨
Disclaimer: This article analyses public filings and reported public statements for educational purposes and is not financial or investment advice. Capex figures and the Nvidia coupling ratio are my own calculations from SEC XBRL data; quarter alignment between companies with different fiscal calendars is approximate. Statements attributed to named individuals are as reported by the cited outlets. Always do your own research or consult a licensed financial advisor before investing.
Comments
Post a Comment