Twenty-five billion dollars in, twenty-one billion out.
That is the entire non-China AI industry in the first quarter of 2026: $25 billion in sales against roughly $21 billion in depreciation on the hardware that generated it. A four-billion-dollar sliver separates the sector from running in place.
For all the trillion-dollar mystique, the aggregate arithmetic is the arithmetic of a business working hard to stay exactly where it is. Revenue has just caught up with the cost of the gear wearing out.
And the market's verdict on the companies leaning hardest into the boom was punishment, not applause. Alphabet and Tesla both slid after their quarterly reports, dragging the major indexes into weekly declines — Alphabet's fall triggered "in part by an increase in its already massive AI spending plans."
Read that mechanism carefully. Investors dumped the stock because the company announced it would spend more. Capex is being treated as a confession, not a strategy.
The tidy story here is the obvious one: AI is a bubble, the returns aren't there, and the smart money is right to flinch every time a hyperscaler opens its wallet. But that story runs into a strange wall.
The returns look thinnest at the exact moment the technology looks most capable.
The same quarter that produced break-even margins also produced an AI model that disproved the Jacobian Conjecture — an 87-year-old open problem human mathematicians had failed to crack. Terence Tao, the greatest living human mathematician, turned to AI to understand the solution. AI solved an open question in quantum cryptography. Erdős problems have become, in Noah Smith's phrase, "almost child's play" for the best models.
So we have machines that are demonstrably smarter than us at the hardest cognitive tasks we know how to define, sitting inside an industry that can barely clear its own depreciation. Something in that pairing has to give.
Machines demonstrably smarter than us at the hardest cognitive tasks we can define, sitting inside an industry that can barely clear its own depreciation.
Either the economics are genuinely broken — intelligence itself hits diminishing returns, as François Chollet has argued, a ball that's already "pretty damn spherical" — or the output is real and simply isn't showing up where we're looking.
That second possibility is not a hopeful hand-wave. It is a documented historical failure mode, and the St. Louis Fed just published the receipt.
In September 1996, the official data said U.S. nonfarm labor productivity had grown an average of just 0.89% a year between 1989 and 1995. Weak. Disappointing. The kind of number that tells you the computer revolution was mostly hype.
It was wrong.
By February 2000, that same 1989–1995 period had been revised up to 1.40%. As of June 2026, after further revisions, it stands at 1.51% — nearly double the real-time figure.
What changed was not the economy. It was the accounting. The 1999 comprehensive revision of the National Income and Product Accounts began treating software expenditures as capital investment rather than a business expense that vanishes on purchase. Reclassify the software, and a chunk of the "missing" 1990s productivity suddenly appeared in the record — right where practitioners had insisted it must be.
Here is the part that matters for policy. Alan Greenspan argued in 1996 that the productivity gains from the information-technology boom were real but not yet visible in the official data. He used that judgment — his belief in output the statistics couldn't yet see — to justify delaying preemptive interest-rate hikes.
He was leaning against the numbers on his own screen. And the revisions ultimately proved him right to do so.
Now run the 2026 tape against the 1996 template, feature by feature.
Measured productivity growth today is, in Smith's words, "decently robust, but nothing really amazing" — the same underwhelming reading that greeted the PC and the internet. Capex is enormous and front-loaded, which is exactly what makes near-term margins look terrible, because depreciation lands immediately while the output gains dribble in over years.
And the single biggest suspect in the 1990s mismeasurement — software treated as an expense rather than an asset — has a direct 2026 analog in every dollar of AI capability that shows up as cost today and as productive capital only later, if the statistical conventions ever catch up.
The break-even margin, in other words, may be a measurement artifact wearing the costume of an economic verdict.
That $21 billion in depreciation is a hard, booked number. The output those chips are generating — code written faster, research accelerated, tasks compressed from hours to minutes — is precisely the category of gain the national accounts are worst at capturing, because so much of it accrues as time saved rather than revenue billed.
We even have a tell that the statisticians know they're flying blind.
FRED just added 137 new data series tracking generative-AI adoption in the United States, collected through the Real-Time Population Survey. They measure the share of working-age adults using AI for work and outside of it, comparative uptake against computers and the internet at the same stage — and, tellingly, "the reported time savings due to generative AI use."
You do not build 137 new series to track a phenomenon you believe your existing accounts already capture. Time savings is exactly the output that never shows up cleanly in output-per-hour statistics. The Fed's own data division is quietly conceding that the standard instruments have a blind spot, and trying to build a workaround in real time.
That the St. Louis Fed published its 1996-vintage productivity comparison in the same window — explicitly framing it as relevant to "how any lags or limitations might impact monetary policy" — is not a coincidence. It is the institution flagging its own history to itself.
Which brings us to the stake.
If the AI industry's break-even margins and flat measured productivity are read as proof that the boom is a bubble to be leaned against, monetary policy could repeat the exact tightening mistake the 1990s revisions ultimately forgave — hiking into a productivity surge it cannot yet see.
The precise danger is not a dramatic surprise hike. It is passive overtightening: a policy stance that stays too restrictive because the dashboard shows weak productivity and thin corporate returns, when the true underlying trend — invisible for now, like 1989–1995's real 1.51% — would justify patience. (The FOMC's communication posture under Warsh is a separate question, owned by the sibling piece; the point here is only that a 1996-style measurement blind spot is exactly the lever such a stance would pull the wrong way.)
Greenspan's contribution in 1996 was not forecasting genius. It was epistemic humility — trusting that real gains could exist below the resolution of his instruments, and refusing to tighten against a number he suspected was wrong.
The instruments in 2026 are, on the evidence of FRED's own scramble, no better calibrated for a general-purpose technology than they were thirty years ago. Software fooled them then. AI, whose entire value proposition is cognitive labor that shows up as time rather than tonnage, is built to fool them worse.
The four-billion-dollar sliver between AI's revenue and its depreciation is real. Whether it's the whole story depends entirely on a question the official statistics are structurally unequipped to answer yet — and the last time we asked it, the answer took the accountants until 2000 to admit.