The reading is not in dispute. On the Wilshire-to-GDP construction the ratio printed 219% for Q1 2026, roughly two standard deviations above its trendline. On the broader total-market-cap version it sits near 230%. The long-run median is about 84%. The dot-com peak was 172%. The ratio first crossed 200% in 2021 and has not been below it since. Three prior instances of this metric getting stretched anywhere near this far were each followed by declines of at least 25%. What is in dispute is whether the series still measures the thing it measured in 1999.
Start with the weakest objection, because it is the one that gets used most. US-listed companies earn something close to 40% of revenue outside the United States, so the numerator is global and the denominator is domestic. That mismatch is real and it widens every year the index leadership gets more software-weighted, since software costs nothing to move across a border. But it is worth perhaps 40 to 60 points of the reading. It takes fair value from 84 to maybe 130. It does not take it to 230. Anyone stopping there has done half the work and reached a conclusion that requires all of it.
The structural problem is different and it is arithmetic. Market cap to GDP divides a stock by a flow. The numerator is the present value of every future cash flow the listed sector will ever produce. The denominator is one year of output. A ratio built that way is only stable, only mean-reverting, if the growth rate and the discount rate stay inside a narrow band across the whole sample. Run the simplest case. Profits growing 4% against an 8% discount rate support a multiple of 25. Move growth to 6% against the same discount rate and the multiple is 50. Nothing became expensive. One input moved two points. The entire mean-reversion argument rests on an assumption that the growth rate embedded in today’s price resembles the growth rate that generated the historical series.
Stated precisely, the AI claim is not about the size of the economy. It is about the split. Labor takes somewhere around 60% of national income and equity holders own what is left after wages are paid. Every prior general-purpose technology, steam through the personal computer, raised output per worker and let the gains divide between the two claims. The argument for AI is that it substitutes for cognitive labor rather than augmenting it, which means the wage bill does not get shared, it gets converted. Move five points of national income from labor to capital and the profit pool expands by roughly a sixth with zero growth in measured output. The numerator rises. The denominator does not move at all.
This is why the indicator is unreadable at this level rather than simply high. It will print its most extreme values precisely when the substitution thesis is working, because a working thesis raises market cap and leaves GDP flat by construction. The cost-deflation channel compounds it. A firm that replaces a $200,000 wage bill with a $20,000 inference bill has shrunk measured output and grown its own earnings in a single transaction, because GDP counts services at cost. Technology that makes things cheaper is numerator-positive and denominator-negative. The ratio has no mechanism for distinguishing that from speculative excess. It was designed for an economy where the split was stable and where output and value moved together.
The version of this that is genuinely outside the sample sits one level deeper. Standard growth theory makes the long-run rate a function of how many people are producing ideas. Researchers are the scarce input, population is the binding constraint, and that constraint has held in every economy in recorded history. AI is the first technology that proposes to produce ideas with capital instead of with people. If compute substitutes for researchers even partially, the rate of discovery becomes a function of investment rather than demography, and investment compounds in a way that population cannot. Electrification made existing workers more productive. It did not remove the human ceiling on the production of new knowledge. Whether AI does is the actual question, and it is not a question that a 1947-to-2020 time series is equipped to answer, because the answer lies entirely outside the period that generated the data.
None of which means the price is right. Being correct about a technology and being correct about an entry price are separate problems, and the 1999 internet thesis was substantively accurate in nearly every particular while costing investors fifteen years. The capital-share argument can also run in reverse: if models commoditize and diffuse cheaply, the surplus lands with buyers rather than shareholders, which is what railroads, airlines and automobiles did to the equity capital that built them. And there is a nearer problem in the numerator itself. Hyperscale AI capex is being depreciated across five and six year schedules against hardware whose competitive life may be closer to three. If those schedules are wrong, current earnings are flattered, and the E in the multiple is soft before the argument about the multiple even begins.
It is also worth noting that this indicator’s failure does not require AI to explain it. It has read overvalued more or less continuously since 2013, across a period in which the market tripled, and Buffett himself has walked back the endorsement, declining to defend any single measure as consistent across time. A metric that has been directionally wrong for over a decade is describing a regime rather than a mispricing. AI is a candidate explanation for that regime. It is not evidence that the regime is durable.
The thesis is falsifiable, and not by the Buffett Indicator. It is falsifiable by the labor share. The BLS nonfarm business labor share series and the BEA figure for compensation of employees as a percentage of national income are where a genuine substitution of capital for cognitive work has to show up, and it has to show up as a persistent decline across an expansion, not a single soft quarter. The series has been drifting lower since 2000 for reasons that predate any of this, so the test is acceleration off that existing trend rather than the level itself. If that acceleration is not visible in the quarterly productivity and costs release within the next several prints, then 230% is not a regime change. It is a price.
