In 1930 Keynes predicted a fifteen-hour week by 2030. Output per hour has since risen as much as he said it would. The hours barely moved, and most of the money went somewhere else. Scroll to see where.
The blue line is US productivity: how much an hour of work produces, after depreciation, indexed to 100 in 1979. Pay for a typical worker (the median full-time employee, in real terms) also started at 100. Draw where you think it went. Drag from left to right across the chart.
Productivity: net output per hour, total economy (EPI, from BLS). Median pay: real median usual weekly earnings of full-time workers, CPI-deflated (BLS, FRED LES1252881600Q). Annual averages, 1979 = 100.
For a hundred years, productivity bought time. Then one country stopped spending it that way.
In 1870 a worker in the United States, Britain, France, Germany or the Netherlands put in around 3,000 hours a year: six days a week, ten or more hours a day.
By 1938 that had fallen by a third. The eight-hour day, the Saturday half-holiday and paid vacations were won in the same decades that electrification and the assembly line were raising output per hour. Workers took part of the gain as time.
After the war the fall continued, and until about 1980 the United States looked like everyone else: 1,931 hours a year in 1970, heading down.
Then the American line flattened. Germany kept falling to 1,335 hours; France to 1,487; the Netherlands to 1,439. The United States ended 2023 at 1,789, down 7% in half a century. Output per American hour, over the same half century, rose about 180%.
Boppart and Krusell, looking at 25 rich countries, find hours still fall by a little under half a percent a year as productivity rises. The postwar United States, they write, is "a striking exception more than a representative feature of modern economies."
If hours did not fall, pay should have risen with productivity. Whether it did depends on which worker you mean and how you count.
Output per hour in the nonfarm business sector is up 138% since 1979. This is the number people mean when they say productivity has more than doubled.
Average hourly compensation for all workers, counted the way productivity is counted (benefits included, deflated by the prices of what firms sell), rose 56%. Already the gap is wide, and this is the most generous measure of pay.
Count only production and nonsupervisory workers, about 80% of the private workforce, and deflate by what they buy rather than what they make: the Economic Policy Institute's "typical worker" line rises 32%. EPI's matching productivity measure, net of depreciation, rises 94%. Productivity grew about three times as fast as this pay.
Go to the median full-time worker's weekly earnings and the gain over 46 years is 13%. The hourly wage of production workers, before benefits, is up 11%. For the worker in the middle, almost none of the productivity dividend arrived as pay.
So where did the distance between the top line and the bottom one go? Bivens and Mishel split the 1973–2014 gap three ways. Three-tenths is a price effect: what workers buy got dearer faster than what they make. One-tenth went from labor to capital. Six-tenths stayed inside the wage bill but moved up the ladder, to the best-paid employees.
The hypothesis most people start with is the simplest one: companies kept it. The national accounts say that is a growing part of the story, and not the biggest part.
The same technology arrived everywhere. What differed was who sat at the table when the gain was divided.
Each dot is a country. Across is how much output per hour rose between 1970 and 2019; down is how much the working year shrank. The United States: productivity up 116%, hours down 7%.
Every other rich country sits further down the page. Germany cut 30%, Denmark 26%, France 24%, Japan 23%, the Netherlands 21%. Some of that is part-time work rather than a shorter full-time week, but all of it is productivity taken as time.
Each cut was negotiated: the Dutch Wassenaar agreement of 1982 traded wage restraint for working-time reduction; IG Metall struck for seven weeks in 1984 and had the 35-hour week by 1995; France legislated 35 hours in 2000; Danish unions set 37 hours by agreement. Shorter hours were never a gift of productivity. They were a demand.
Which is why this chart matters. Union membership fell from 20% of US workers in 1983 to 10% in 2025, the lowest in the rich world. The institution that historically converted productivity into either pay or leisure went with it. In Denmark and Sweden, where density stayed above 60%, hours kept falling.
The first field studies of generative AI at work show the pattern repeating in real time.
Time savings exist. They are being turned into output and headcount decisions, not into shorter weeks or higher pay, exactly as the computer era's gains were. The four-day-week trials show the alternative is feasible when somebody decides to take it.
"Companies are making more profit." Supported. After-tax profits doubled relative to GDP; labor's share fell eight points; two-thirds of OECD countries saw the same direction, more gently. S&P 500 firms paid more than 90% of a decade's earnings out to shareholders, and the wealthiest 1% of households own half of all shares.
"They pay people just enough to survive and consume." Too strong as stated. Median real pay rose 13 to 30% depending on the measure, and goods got dramatically cheaper. But the median worker got a thin slice of a 138% productivity gain, and most of the missing slice went to the top of the wage ladder before any of it reached shareholders. At the very top of that ladder, CEO pay rose 1,316% while the typical worker's rose 28%.
"That is why people aren't working less." Partly. It does not explain why the Dutch, Germans and Danes, whose median wages also lagged productivity, did work less. The difference is institutional: where unions and legislation set hours, productivity became time; where they weakened, it became profit, top-end pay and consumption of things whose prices rose fastest.
Keynes got the growth right and the hours wrong because he assumed the dividend would be shared. The century of data says it is shared only where someone is organized enough to claim it.