The Economy Is Growing Faster Than We Can Measure
AI GDP is growing at 2,000% per year and the national accounts can't see it — the gap between the real signal and the measured signal is where the alpha lives.
GDP is a lagging indicator of an economy that’s already been replaced.
The Marginal Revolution post on AI in GDP is the most important economic data story nobody is treating seriously enough. Preliminary estimates put nominal AI GDP at roughly $250 billion, growing at over 2,000 percent per year on a quality-adjusted basis in 2024 and 2025. Three compounding forces: data-center capacity, hardware efficiency, and algorithmic progress — the last one being the biggest. Read that again. Not 20 percent. Not 200 percent. Two thousand.
The conventional response is to say the measurement is flawed. That’s probably true. But the direction of the flaw almost certainly understates reality rather than overstating it. National accounts were designed for an era of widgets and warehouse square footage. They’re constitutionally incapable of capturing an input — intelligence — that is simultaneously a commodity, a capital good, and an infrastructure layer. When the Bureau of Economic Analysis struggles to price a GPT-4 equivalent that costs 99 percent less than it did eighteen months ago, you get systematic undercount. The official number is a floor, not a ceiling.
Here’s the deeper issue. Everyone is arguing about whether AI productivity gains are “real” or will show up in GDP. That’s the wrong question. The gains are already real — they’re just accruing in places the tape measure doesn’t reach. A biotech researcher who compresses six months of literature review into a week isn’t generating a measurable output event. A solo founder who ships a product with no engineering team doesn’t create a hiring impulse that registers in payroll data. The surplus goes to the user, not to the national account. Economists call this consumer surplus. In aggregate, right now, it might be the largest silent transfer of productive capacity in recorded history.
The investment implication — and I’m not giving you a trade, I’m giving you a framework — is that the entities best positioned are the ones who own the infrastructure layer before the measurement catches up. IREN just borrowed $3.6 billion to buy Nvidia chips for Microsoft. That’s not a bet on one application or one model. That’s a bet that the compute substrate is as close to a sure thing as anything in markets right now. The algorithmic progress driving that 2,000 percent figure requires physical hardware to run on. The atoms are the bottleneck, and right now capital is racing toward the atoms at a speed that would look irrational if you only read the official GDP tables.
The political economy of this matters too. Washington’s AI policy debate — the White House split Politico reported this week — is organized around questions of safety, export controls, and who gets to sit at which table. Those are real issues. But they’re downstream of a more fundamental problem: policymakers are trying to regulate an economy they can’t measure. You can’t set sensible policy on AI’s labor market effects if your labor market data doesn’t capture AI-augmented output. You can’t calibrate trade restrictions if your trade statistics don’t reflect the cross-border flow of model weights and inference compute. The information deficit isn’t a footnote to the policy problem. It’s the policy problem.
The optimistic read — and I hold it — is that this measurement gap is temporary. The national accounts eventually caught up to the internet. They’ll catch up here too. When they do, the revision upward will be the kind of number that ends careers on the pessimist side of the ledger.
The stakes are not abstract. If AI GDP is genuinely compounding at rates that dwarf any prior technology cycle, the policy interventions designed to slow it down are playing defense against a force that has already lapped them. And the investors, operators, and builders who act on the real signal rather than the measured signal will have accumulated advantages that no amount of catch-up can erase.
The most valuable skill in this economy is not building AI. It’s reading the gap between what the scoreboard says and what’s actually happening on the field.