Stocks for Thoughts · The Wage Bill, part two

Nobody taxes a machine.

The market has priced twenty‑five trillion dollars on AI taking a cut of the world's wages. The Treasury already spent that money.

The market values its claim at twenty‑five trillion dollars. The Treasury spent its claim years ago, on pensions and hospitals and interest. Both can be paid, in principle — a government can tax the same income somewhere else. The point is that it currently doesn't, and that the arithmetic of not doing so is larger than anyone has bothered to write down. So that is what follows.

Two things about that table are worth sitting with. The first is scale, and it needs the two numbers put on the same footing. The Trustees express Social Security's seventy‑five‑year shortfall as 1.5% of GDP — and that figure is already an annual one, the average adjustment they calculate is needed in each year from now to 2100. A twenty per cent displacement costs about 2.2% of GDP a year. Larger than the gap Washington convenes a commission about every few years and never closes.

The second is direction. Every previous wave of automation hit the bottom and the middle of the income distribution, which was fiscally cheap, because those people weren't funding much. This one is walking uphill towards the people who are — and the tax system's grip loosens as it climbs. Wages are 84% of the income tax paid in the thirty‑to‑fifty‑thousand band, 59% at a million, and 36% above ten million. Past a certain altitude income stops looking like a wage and starts looking like a capital gain, and the code notices.

All of the above depends on displacement actually happening, which is slow, contested, and easy to argue about for a very long time.

This next part doesn't.

Acemoglu, Manera and Restrepo went and measured what the American tax code charges for the two ways of getting a piece of work done. Hire a person: an effective rate of 25.5%. Buy the equipment or software that does the same job: about 5%.

Not a loophole. Not aggressive planning. That is the statute working exactly as written.

So thirteen months ago, in the middle of the largest corporate capital build in history, Congress made the discount on automation permanent. I don't think that was aimed at AI. Full expensing has been a mainstream Republican tax position for twenty years and it was sold as a manufacturing measure. But intent is not the same as effect, and the effect is a twenty‑point wedge that widens precisely as the substitution gets easier.

The reassurance is always 1800. It worked out, look at us now.

The received version is Robert Allen's: output per head up 46% between 1780 and 1840, working‑class real wages up 12%. Sixty years of the machines working and the workers not getting paid.

Nicholas Crafts went back at the numbers in 2020 and found something else. Labour's share of British national income was 61.0% in 1770 and 60.2% in 1860 — essentially unmoved. Capital's share nearly doubled, but the counterpart wasn't wages. It was land, which collapsed from 21.8% to 8.5%. The great redistribution of the Industrial Revolution ran from landlords to factory owners. The workers weren't robbed.

So why were wages flat? Because output was. Total factor productivity ran at 0.32 to 0.38 per cent a year before 1830. Crafts's own verdict: "more like a story of paradoxically slow productivity growth than of pro-rich growth."

Three things fall out of that, and all three are about now.

The aggregate saw nothing while the specific was carnage. By 1851, at the height of it, textiles, metals and machine‑making together were 13.4% of employment. Agriculture and most of services were, in Crafts's words, largely unaffected. The Industrial Revolution touched about an eighth of the workforce.

The handloom weavers are the whole skill argument in one trade. Two hundred and forty thousand of them in 1820. Forty‑three thousand by 1850. Weekly real pay from 276 pence at the 1805 peak to 75 pence by 1830 — down 73% while they were still employed. The evidence does not show them moving into better‑paid replacement work: a skilled craft was destroyed and the jobs that took its place went to cheaper unskilled operatives. Only the second phase — engineers, mechanics, clerks — was the kind that raises wages, and it arrived a generation late.

And the clock.

If AI runs on that clock, everything about it is real and the payoff lands two decades after the capital was committed. Which is a problem for a market, and a much bigger problem for a Treasury that has to fund the gap in between.

Resist the obvious next step, though. The lags do shorten — Comin and Hobijn measured two centuries of it, and a technology invented ten years later is adopted 4.3 years faster — but the rate at which they shorten depends entirely on which moment you call the invention, and four defensible ways of running the trend forward put AI's payoff anywhere between 2022 and 2034. A sixteen-year spread is not a date. It is the reason this argument has to be built on the tax code, which you can read today, rather than on a clock nobody can set.

Britain got through a forty‑year wage pause. It is worth asking how, because the answer is not the one people assume, and it is not transferable.

There is a refinement worth making, because it cuts against Britain too. Excise fell on beer, malt, sugar, tea and soap — so the Exchequer depended absolutely on working people spending. It held up because the population was growing 1.1% a year. Consumption per head stagnated; there were simply more heads. Britain's tax base was rescued by demography.

Ours is going the other way.

Britain had the worse balance sheet and got through by handing the bill to people with no voice. We have the better balance sheet and nobody to hand it to. Which means the cost has to land somewhere else, and there are three doors. It is not Margin's job to tell you which one to walk through.

One paragraph on Europe, because the comparison runs the opposite way to the one people expect. Continental states raise 24 to 26 per cent of GDP from taxes on labour, against 16.3% in the United States, and they are ageing considerably faster — old‑age dependency around 41 to 43 today in France, Germany and Italy against 31.7 in America, and 74 to 76 in Italy and Spain by 2050. The American advantage is on the revenue side and the demographics. It does not extend to the balance sheet. On the IMF's general‑government gross measure the United States is at roughly 126% of GDP for 2026 — below Japan at 204%, Italy at 138% and Greece at 137%, but above France at 118% and Spain at 98%. American borrowing is now on the same scale as Europe's most indebted states, which was not true a decade ago.

The bears in this particular argument are the people saying it doesn't matter, and they have a real case.

Nobody official thinks this is happening. The 2026 Social Security Trustees Report — the document that determines when the fund runs dry — contains zero mentions of artificial intelligence, automation, or technological displacement. The Medicare report: zero. The Chief Actuary's testimony to the Senate in March: zero. The Congressional Budget Office does model AI, as a productivity tailwind worth a tenth of a point a year, and never carries it through to trust fund solvency at all. In CBO's optimistic case, AI reduces federal debt.

So every official projection of the American fiscal future currently assumes AI either doesn't happen or helps. That is either a warning or a reassurance, and reasonable people read it both ways.

The measured labour effect is still tiny. The clearest finding anyone has is Brynjolfsson, Chandar and Chen's, drawn from ADP payroll records for millions of workers a month: a 6% employment decline for twenty‑two to twenty‑five year olds in the most exposed occupations, against a 6 to 9% rise for older workers in the same jobs. It is a difference‑in‑differences estimate on one payroll provider's panel, not a clean causal identification, and it says nothing about pay. Aggregate studies find close to nothing. And adoption is still thin: the Census Bureau's fortnightly survey put firms using AI in any business function at 17 to 20 per cent through the first half of this year, with 20 to 23 per cent expecting to within six months.

And states tolerate tax asymmetry for a very long time. The gap between the tax on labour and the tax on capital has been widening since about 2000 and no government has fixed it. Betting that this is the one that finally forces the issue is betting against a twenty‑five year losing streak. That is the strongest thing anyone can say against this note, and I don't have a clean answer to it.