The Criterion That Didn't Fire
The most useful thing a forecast can do is say in advance what would prove it wrong.
I did that. In the AI essay I named a condition: if Microsoft, Alphabet, Meta or Amazon cut capital expenditure guidance by more than ten per cent, the capex argument was in trouble. Not "if sentiment shifts" or "if the narrative breaks" — a number, at four named companies, that anyone could check against a filing.
Between 22 and 30 July, all four reported.
Alphabet raised 2026 capex guidance from $180–190 billion to $195–205 billion. Microsoft disclosed a FY2027 guide of $255–260 billion. Meta lifted the floor of its range from $125 billion to $130 billion. Amazon went from roughly $200 billion to roughly $220 billion, and Andy Jassy attributed the increase to higher memory-chip costs.
None of them cut. All four raised, inside nine days.
And the second-order thing I expected is also missing. If the spending were genuinely unsustainable, the strain should show up first in margins, in the same filings. It hasn't. Alphabet's operating margin expanded from 32.4% to 34.0% year on year while its capital spending nearly doubled. AWS margins rose 650 basis points during the heaviest investment cycle in its history. Meta is the partial exception — earnings per share missed on capex-driven spend and the stock fell around eight per cent — which is one company, not a pattern.
There are two wrong readings of this available and both are tempting, so let me take them in order.
What it doesn't mean
It does not mean the arithmetic was wrong.
The arithmetic was always the load-bearing part. David Cahn's framing is the cleanest version of it: "AI's $600B Question is now roughly AI's $840B Question, assuming that Nvidia reaches something like $210B in run-rate data center revenue by year-end 2025." That was June 2025. The question he was asking is where the revenue comes from to justify the infrastructure, and nothing in the last nine days answers it.
Disclosed AI revenue has grown substantially — and I should correct my own number here, because I have been carrying a stale one. I wrote roughly $50 billion. The current disclosed figures are Anthropic at $47 billion run-rate, Microsoft at $37 billion in AI annual recurring revenue, OpenAI at $25 billion annualised. Call it $109 billion, with the caveat that these are inconsistent measures — run-rates against annualised figures, bundled against standalone — and that only one of those companies files with the SEC in a way that lets you check.
So the revenue has roughly doubled and the gap has not closed. It hasn't closed because the denominator moved too: Cahn published again on 8 July, and his number is now $1.5 trillion.
The gap is real. It is arguably wider than when I wrote about it.
What it does mean
It means I bundled two claims together and only one of them is holding.
The first claim is structural: capital spending of this magnitude requires revenue that does not yet exist, and the two must eventually reconcile. That claim is untouched. Every number above supports it.
The second claim is about timing: that the reconciliation arrives inside 2026–2027. That claim is the one taking damage, and it is taking damage from the specific evidence I said to watch.
These are different claims with different failure modes, and I wrote them as though they were one. The structural claim fails if AI revenue scales to meet the capex — if the technology simply earns its infrastructure the way the argument says it can't. The timing claim fails much more easily. It fails if everyone involved is willing to keep spending for another two years, which is precisely what four earnings releases just said they intend to do.
There is a version of the bear case that survives this comfortably by retreating to "eventually", and I don't want to write that version. "Eventually" is not a forecast. It is a way of never being wrong, which is the same thing as never having said anything.
The base rate I keep having to sit with
The strongest argument against me is not that AI is more valuable than I think. It is a historical pattern, and it is genuinely uncomfortable.
Railroads, electrification, fibre — general-purpose build-outs have run revenue gaps of this shape before, and the gaps closed on something like a five-to-seven-year lag rather than a two-year one. In each case there were people making my argument, with the arithmetic correct, who were right about the destination and wrong about the schedule by enough that being right stopped paying. Fibre is the analogy that should worry me most, because there the overbuild was real and the write-downs did arrive — and they arrived late enough that being early was its own kind of wrong.
Against that: the revenue growth rates here are not a normal base rate either. Anthropic's disclosed progression through the first half of 2026 — $14 billion in February, $19 billion in March, $30 billion in April, $44 billion in early May — is not a company approaching a ceiling. If that compounding continues at even half its current pace through 2027, the gap closes from the numerator without anyone cutting a single dollar of capex, and my structural claim gets tested directly rather than being rescued by a timing argument.
I don't know which of those it is. That is the honest position, and it is a different position from the one my essay implies.
What I'm changing
Three things, and I'd rather write them down than quietly adjust the framing later.
The timing claim comes down in confidence. Not the structure — the schedule. A 2026–2027 window that has passed three quarters with its own disconfirming evidence accumulating is not a window I should still be asserting at the same strength.
The criterion gets sharper. "A ten per cent guidance cut" turned out to be a blunt instrument, because guidance can be restated for reasons that have nothing to do with conviction — Microsoft's headline 2026 number moved from $190 billion to $175 billion this quarter purely because extending server useful life from fifteen to twenty-five years shifts spend from capital to operating expense. That is an accounting change, not a retreat, and a criterion that can be tripped by depreciation schedules is not measuring what I wanted it to measure. What I actually care about is whether these companies are still financing the build-out from operating cash flow, and whether the marginal dollar is going into capacity or into debt service.
And I'm writing down a date. If the four of them are still raising guidance through the first half of 2027, with margins intact, the timing claim is dead and I will say so in those words. Not adjusted, not reframed. Dead.
Why bother
There is an obvious question here, which is why publish this at all when nobody was auditing me.
Because the alternative is what most macro writing does, which is to let a timing claim quietly become a structural one, and a structural one quietly become a mood. The claim was falsifiable when I made it. It is still falsifiable now. The only thing that changes if I don't write this is that the next person to read the essay gets a slightly more flattering version of what I knew and when.
I write this from Cape Town, where the question is academic in the most literal sense. Combined guidance across those four companies for 2026 comes to something like $725 billion. South Africa's entire annual output was $427 billion in 2025, the last completed year. Four firms intend to spend, on data centres alone, something like one and three-quarter times what sixty-three million people produce in a year, and none of it will be spent on this continent. That distance is occasionally useful. It makes it easier to notice that a number can be enormous and still not be evidence of the thing you want it to prove.
The next thing worth watching is not the capex line at all. It is whether the spending is still being funded from operating cash flow, or whether it has started routing through credit — because that is where an argument about valuation quietly turns into an argument about solvency, and the two fail very differently.