Labor's share of US GDP was 64% in 1974. It was 56% by 2024. Citrini Research's 2028 scenario projects 46% by the end of the decade. That is eighteen percentage points of national income that used to flow to wages and now flows somewhere else. The somewhere else is the owners of compute.
This is the single most important macroeconomic fact of the next five years, and almost no public debate is anchored to it.
What labor share actually measures
Labor share is wages plus benefits divided by GDP. When it falls, capital's share rises by definition — not because workers got lazy, not because immigration shifted the curve, but because the productive value generated in the economy is increasingly owned by the people who own the production assets rather than the people who operate them.
The 1974 to 2024 decline tracks four overlapping forces: globalization, union decline, the shift from manufacturing to services, and automation. MIT research finds automation alone explains 50% to 70% of the rise in US wage inequality from 1980 to 2016. Automation is the dominant variable, not the residual.
The 2024 to 2028 projection is different in kind. It compresses what took fifty years into four. The mechanism is generative AI applied to the largest remaining pool of cognitive labor — white-collar work — combined with capital deployment on a scale previous automation waves never had. Hyperscalers spent up to $660 billion on AI capex in a single year. That capex does not employ a meaningful number of workers per dollar deployed. It substitutes for them.
Eighteen percentage points of GDP is roughly $5 trillion at current US output. That is the order of magnitude of the redistribution under discussion. It is larger than the entire federal discretionary budget. It is larger than total US healthcare spending. It is the prize.
The compute owners' share, mechanically
Where does the eighteen points go? Not evenly. The income that used to flow to white-collar wages flows up the compute stack in a specific order: chip designers (Nvidia), hyperscalers (Microsoft, Google, Amazon, Meta), frontier model labs (OpenAI, Anthropic), and the companies that own the deployment surface for AI agents.
Each layer is concentrated. The Yale Law and Policy Review's antimonopoly analysis describes severe concentration at multiple layers simultaneously. There is no competitive layer in the stack that would compete the rents away. Nvidia has a 90%-plus share of training accelerators. Three hyperscalers run the cloud. A handful of labs train frontier models. Each layer reinforces the next.
The ownership of those firms is also concentrated. Index funds nominally hold the equity, but the economic exposure flows through a small set of asset managers, sovereign wealth funds, and ultra-high-net-worth households. The eighteen percentage points of national income do not arrive in millions of household checking accounts. They arrive on the balance sheets of a few thousand entities.
Why this trajectory does not self-correct
Labor share has self-corrected before. Postwar union density, the GI Bill, the minimum wage, and tight labor markets pushed labor share back up in the 1950s and 1960s. The mechanism was political: organized labor was strong enough to bargain wages up, and full employment kept the bargaining power durable.
None of those conditions hold now. Union density is at a generational low. The AI displacement cycle is creating slack faster than retraining can absorb it. The political constituency for raising labor share is the constituency being structurally weakened by the trajectory itself, which is the dynamic Acemoglu's work has been documenting for two decades.
The displacement spiral that Citrini describes — layoffs reduce demand, demand pressure increases AI investment, AI investment increases displacement — has no internal brake. The brake has to come from outside the loop. That means policy: a compute tax, mandatory profit-sharing, public ownership of compute infrastructure, antitrust action that breaks the concentration of the stack, or some combination.
None of those policies are on a credible legislative path in any major economy at the speed the labor-share decline is occurring. Antitrust takes five to ten years. The labor-share trajectory takes four. The arithmetic is not subtle.
If the next eighteen percentage points of national income are about to redistribute from labor to compute over four years, the political question is not whether to redistribute. It is who writes the rules for the redistribution that is already happening?