On 5 November the chip index had its worst day since March 2020, down 12.2% in a session and 38% below its June peak. A week later Jensen Huang resigned as Nvidia’s CEO. By the definition most people use, that is what a bubble bursting looks like.
Now look at the revenue line over the same months. End-user AI spending kept compounding. The largest lab’s growth sped back up, and its numbers moved from press leaks into a public filing. The money for the next leg of the build-out was raised and put on balance sheets before it was needed.
In The Rental Bill (August) we argued that the technology would be fine and the schedule was the risk. In Ten Years in Six Months (September) we laid out two paths: a straight line, or a sawtooth of sharp drawdowns on the way up. This autumn delivered the sawtooth. This report is the check-in we promised: every reading we said we would watch, with its as-of date.
1. Why this tooth cut deeper
The earlier pullbacks were about one worry at a time: a model price cut, a capex wobble, a financing headline. This one stacked four into three weeks.
- Credit. Meta’s credit default swaps hit a record 127bp on 23 October.
- Politics. The 3 November midterm sweep raised doubts about data-center build-outs and local moratoria.
- Governance. Nvidia’s board opened an internal investigation on 8 November; Huang resigned on the 12th, with CFO Colette Kress as interim CEO.
- Rates. The 10-year Treasury sat near 5.1% by 13 November, and chip-equipment names fell despite beating estimates.
None of those is a revenue problem. All of them change what investors will pay for a dollar of future revenue. Positioning mattered too: by June the chip trade was crowded, so the exit was crowded as well. A deeper cut in that setting says more about who owned the stocks than about who is buying the chips.
2. Why we no longer call it a bubble
A bubble is a price that has left its earnings behind. What happened this autumn ran the other way: prices fell while revenue and earnings estimates rose.
Source. BofA forward P/E, Nov 5, 2026. Nvidia multiple is Proflex arithmetic: the Nov 12 close of $166.19 against the fiscal 2028 earnings Bernstein uses for its $200 target, which it left unchanged when it downgraded.
The chip index trades at 13.7x forward earnings against 17.7x for the S&P 500. The labs grew into their prices too: Anthropic’s private mark was $965B in May, on a run-rate below half of today’s $81B. The things that would have broken the sawtooth argument are a fall in end-user revenue, a cut in capex plans, or a failure to fund the build. None of those happened.
Both are true, and we want to be clear about it: prices can keep falling on a story that is intact. We are describing the business and the funding, not predicting the next move in the stocks.
3. Revenue is running ahead of the bill
In August the question was whether revenue could catch the rent on the capacity being built. Three months later the gap is narrower.
Source. Anthropic, OpenAI, Gemini and xAI. Oct = $131B reported (Anthropic $81B at Oct 31, S-1/A; OpenAI ~$50B at Sep 30, FT/Bloomberg) plus a $22B Proflex estimate for Gemini and xAI, which do not disclose dollar revenue. Jan Anthropic and Jul OpenAI are also estimates.
Anthropic’s run-rate went from about $71B at the end of September to $81B at 31 October — about 14% in one month, from its own amended filing. In August we said growth below 8% per 30 days for a third period would be bearish, and above 20% bullish. October sits clearly above the bear line. OpenAI’s ~$50B at end-September works out to roughly 8–9% per 30 days: inside the band it has held all year, at the low end.
Source. Proflex AI revenue tracker, as of Nov 10. Annualized and de-duplicated: revenue that is another tracked company’s cost is removed. October gross $338B less 33% overlap. About 60% of the gross figure is our estimate, so read the trend more than the level.
Our tracker removes double counting — where one company’s revenue is another’s cost — before adding labs and clouds together. On that basis end-user AI spend is $225.5B annualized, 3.2x since January. Against the $320–400B a year of rent we estimated in August, the rent now runs at 2.4–3.1x the two largest labs’ revenue, down from 2.8–3.5x. The gap is closing from the revenue side.
“The Air Came Out. The Revenue Didn’t.” — the complete 12-page Proflex research report. The rent-coverage table then and now, the Anthropic deal as filed, who raised what in 2026, the neocloud backlog-versus-debt table, the AI spending dependency map, and the signposts with our read on each. Every figure carries its as-of date.
4. The first frontier lab to list
The model makers are where the money starts. Their choices decide whose chips, which clouds and how much memory gets bought. This week the largest of them becomes a public company, and its filing already answers questions we could only estimate in August.
Source. Anthropic S-1 and S-1/A (Nov 2 and Nov 10, 2026). Q3 2025 derived from the reported +1,127% year on year; Q1 2026 from secondary analysis of the filing. Nine-month revenue $33.2B; operating loss $1.26B; cash $131.9B at Sep 30.
Anthropic reported $16.96B of revenue in Q3, up 1,127% year on year, with an operating loss of $1.26B and $131.9B of cash at 30 September. The IPO range is $70–78 a share, raising about $61B at the midpoint, alongside a $26B placement at the IPO price from Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC.
In August we said the disclosure that would change our view was the revenue-recognition note in an eventual public filing. We now have it: marketplace sales are booked gross, about 44% of nine-month revenue. That makes the headline larger than the cash Anthropic keeps, which our tracker already adjusts for. It does not make the demand smaller.
What would options pay on your stock?Enter a ticker and see estimated covered-call income and hedged downside in 5 steps.Try the calculator →5. The build-out is raised and on the books
The worry in August was who pays the rent if revenue is late. This year the spenders answered by raising capital before they needed it: stock, bonds and, this week, the largest lab’s IPO.
Source. Company Q3 releases and calls (Oct 21–29, 2026). Combined ~1.8x the ~$410B of 2025. Microsoft is calendar-year and was trimmed to ~$170B on Oct 28, citing third-party capacity delays. Alphabet, Amazon and Meta all point to higher 2027 spending.
Source. Cash and liquid investments at Sep 30, 2026, from company Q3 filings; SpaceX approximate, post-IPO. Before the Anthropic IPO and later bond deals: Anthropic adds ~$85B (new shares at the $74 midpoint plus the $26B placement) if the deal prices as filed.
Capital raised and sitting on a balance sheet is the opposite of the late-cycle pattern we worried about, where spending runs on promises. It does not remove the risk; it moves it to shareholders and bondholders while revenue catches up. And both are true here as well: free cash flow turned negative at Alphabet (−$9.1B in Q3) and Meta (its first negative quarter since 2012), and some money is still circular — suppliers buying customer stock with cash that returns as compute orders.
6. Nvidia without Jensen: what changed, and what didn’t
The founder’s exit removes the industry’s most recognizable executive. It does not remove the installed base, the software ecosystem or this quarter’s order book. The market is repricing the multiple and the governance, not the business.
Source. Stockscan price history; Fox Business. Nov 12 close is 31.7% below the Oct 6 record. Bernstein halved its target to $200 and left its earnings estimates unchanged; Mizuho reiterated Outperform at $250. Fiscal Q3 results are due Nov 17.
Nine firms stepped back this week, and the targets came down with them. But look at what Bernstein did: it halved its target and left its earnings estimates unchanged. That is the pattern of the whole autumn in one stock — the price came down, and the earnings outlook did not. The key question for 17 November is whether the interim team delivers the guided $108.0B ±2% at a 74% gross margin and reaffirms the data-center outlook.
7. Paid in every scenario: memory and optics
If Nvidia keeps its share, it needs high-bandwidth memory. If TPUs and custom chips gain, they need it too. The same goes for the optical links that connect them. That is why memory and optics posted the most dramatic numbers of this cycle.
Source. SK Hynix Q3 results (Oct 28, 2026); Seoul Economic Daily; VLSI. Bars from zero. Q3 revenue KRW 104.6T, up 328% year on year; operating profit up 632%.
Micron’s fiscal Q4 revenue was $54.2B, against roughly $11.3B a year earlier, and it says the vast majority of its calendar 2027 HBM supply is already under agreement. Lumentum’s revenue rose 137%, and Marvell agreed to buy it in a $115B all-stock deal. Booked volume at higher prices gives memory a visibility it has rarely had. The classic risk is still the cycle: when all three memory makers accelerate capex together, pricing power has usually peaked within a few quarters.
Where this leaves the story
In September we closed with: the technology is no longer the variable; the path is. Two months later we would add one line. The path is now being paid for in advance. Revenue is further ahead of the rent than it was in August, the largest lab is about to report every quarter in public, and the spenders raised the money before they needed it.
The air that came out was price. The revenue stayed.
Six readings that tell the straight line from the sawtooth — lab growth per 30 days, revenue-recognition disclosure, 2027 capex, enterprise measured ROI, GPU rental prices and credit for the builders — with where each stands now. Plus three scenarios (rebuild, funding squeeze, capex cut) with what holds up best and what is most exposed in each, and the green lights and red flags we are watching.
Figures are from the Proflex Finance research report “The Air Came Out. The Revenue Didn’t.” (November 2026), which dates and sources every number; data as of 13 November 2026. OpenAI figures come from press reports of internal documents, and private-company figures are unaudited. Figures marked as estimates are Proflex arithmetic, not disclosed numbers. Prices are as reported on the dates shown and will have changed — verify anything you intend to act on against current sources.
What would options pay on your stock?Enter a ticker and see estimated covered-call income and hedged downside in 5 steps.Try the calculator →