The productivity dividend from artificial intelligence is starting to show up in the data. It is just not showing up in wages. In the second quarter of 2026, US workers' share of gross domestic product fell to 52.8 percent, the lowest reading since records began in 1947, even as corporate profit margins reached a record 14.9 percent. "Productivity growth protects margins, not income," EY-Parthenon economist Gregory Daco told Fortune, before adding a line that ought to unsettle anyone drawing a salary. Asked where the floor was for labor's share, he said: "I don't think there's a floor."
The scale of the capital side is hard to picture. AI server imports hit a $450 billion annualised pace in September, up from roughly $50 billion a year through 2023, one of the fastest single-category investment surges in modern US history. Data-center investment is projected to reach $31 trillion by 2050. Meanwhile hours worked rose just 0.3 percent in the quarter, and the economy grew at a modest 1.7 percent. Money is pouring into the machines that do the work; far less of it is reaching the people who used to do that work.
Into this landscape, Anthropic's economics team released an interactive model projecting AI's effect on growth, jobs and wages through 2030. Its scenarios are notably more restrained than the double-digit-growth talk of some AI chief executives. The "substantial" case, grounded in a survey of 10,000 Americans, sees 5.4 percent growth in 2030 with only a 0.7 percent rise in AI-driven unemployment. Even the "extreme" case reaches a bonkers 15.4 percent growth rate only alongside an 8.9 percent jump in joblessness. One of the authors' more honest conclusions is that, across the full range of reasonable assumptions, "the range of outcomes is wide."
Not everyone buys even the middle of that range. LSE economist Ben Moll and Google DeepMind's Alex Imas argued that forecasts of double-digit growth rest on shaky foundations: that machines do most of the economy's work by 2035, that buyers keep spending, that AI never destroys value along the way. "Each of these assumptions may fail in the real world," they wrote, noting pointedly that even Anthropic's own model only reaches double digits in its extreme scenario.
Underneath the macro numbers sits a human timing problem. Gartner predicts that by 2029, one in three workers laid off because AI took their job will have to be rehired, sometimes at higher cost, as firms discover they shed institutional knowledge they still needed. The warning is that treating AI purely as a way to cut headcount can be a false economy. The optimistic reading, floated by Anthropic's economists, is that demand for manual labor could rise and offset losses in desk work. The pessimistic reading is already sitting in the second-quarter figures. Which one wins will depend less on how clever the models get, and more on who ends up owning the gains.