On March 31, 2026, Oracle fired 30,000 people with a 6 a.m. email. Employees opened a message from “Oracle Leadership” telling them today was their last day. Their managers hadn’t been told. HR hadn’t called. IT cut system access within hours.

Oracle posted record profits the same quarter. Net income jumped 95% to $6.13 billion. Oracle fired those people to free up $8 to $10 billion in cash flow for GPU clusters.

Two days later, the Bureau of Labor Statistics released the March 2026 Employment Situation report. You read the headline: total nonfarm payroll employment increased by 178,000, unemployment held at 4.3%. You read the tables and find something different.

The BLS establishment survey tracks employment by industry. Year over year, from March 2025 to March 2026, three sectors contracted:

  • Information: 2,867,000 to 2,791,000. Down 76,000.
  • Financial activities: 9,201,000 to 9,134,000. Down 67,000.
  • Computer systems design and related services: 2,408,900 to 2,368,100. Down 40,800.

Computer systems design grew in 19 of the past 20 years. A 1.7% decline in a single year reverses a trend that most analysts, including the BLS itself in its 2024-2034 projections, assumed would continue.

Two sectors grew enough to offset those losses:

  • Health care and social assistance: Added about 680,000 jobs.
  • Food services and drinking places: Rose from 12,208,400 to 12,361,800.

In aggregate, the economy traded one kind of job for another.

Information sector workers earn $54.61/hour. Finance workers earn $49.02/hour. Health care and social assistance workers earn $33.11/hour. Food service workers earn $20.84/hour.

The jobs disappearing pay two to three times what the jobs appearing pay. A computer systems designer earning $112,000/year can afford a mortgage and childcare. A food service worker earning $41,600/year, assuming full-time hours that many food service workers don’t get, can cover rent and groceries.

The BLS report doesn’t track whether displaced information workers are taking health care or food service jobs. But it does track how long people stay unemployed. In March 2026, 1,821,000 people had been unemployed for 27 weeks or more, up 322,000 from a year earlier. Long-term unemployed now make up 25.4% of all unemployed people, up from 21.3%.

People who lose $54/hour jobs don’t find $54/hour replacements within six months. They sit in the 27-weeks-and-over bucket while employers post $20/hour openings those workers won’t take and can’t afford to take.

The establishment survey counts payroll jobs. The household survey counts people. You get different answers from each.

The household survey shows 162,848,000 people employed in March 2026, down 661,000 from March 2025’s 163,509,000. That’s 661,000 fewer humans working, even though the establishment survey shows 260,000 more payroll positions.

The gap comes from self-employment losses. Unincorporated self-employed workers fell from 9,820,000 to 9,458,000, a drop of 362,000 in one year. Freelance developers and independent consultants who built businesses on specialized knowledge lost clients to AI tools that do the same work for a flat monthly fee.

Another 2,389,000 people left the labor force. The participation rate dropped from 62.5% to 61.9%. In March alone, 144,000 people told the BLS they stopped looking for work because they believe no jobs exist for them.

Follow the $660 billion in hyperscaler capex for 2026 and see who catches it.

Nvidia captures about 90% of AI accelerator spending. One company with 30,000 employees absorbs $180 billion in GPU sales. Oracle fired the same number of people that Nvidia employs, and Nvidia generated more revenue last quarter than Oracle saved.

OpenAI hit $25 billion in annualized revenue with about 4,000 employees. Anthropic reached $19 billion with about 1,500 employees. Businesses pay these companies subscription and API fees, using the savings from the information workers they laid off to cover the cost. OpenAI and Anthropic send most of that revenue to cloud providers for compute. The cloud providers send the capex to Nvidia and TSMC. TSMC fabricates the chips in Tainan.

American companies fire $54/hour American workers. They buy API subscriptions for a fraction of the labor cost. A handful of AI companies with tiny headcounts collect the subscription revenue. Those companies pass most of it to semiconductor firms and cloud infrastructure providers. Shareholders of those firms capture the productivity gains as stock price appreciation.

From March 2025 to March 2026, the S&P 500 component driven by AI infrastructure went up. Over the same period, 661,000 fewer Americans had jobs. Both facts appear in the same BLS release if you read both surveys.

The Structural Problem

A consumer economy runs on consumer spending. Consumer spending runs on wages. The median American household owns too little stock for capital gains to replace lost wages.

Replace millions of $54/hour jobs with $20/hour jobs and you watch consumers cut spending. People who used to buy cars and furniture now buy groceries and pay rent. Car dealers lose revenue. Home buyers disappear because mortgage underwriters look at income, and median income in affected metros drops.

Local and state governments lose tax revenue at the same time. A worker earning $112,000 pays far more in income tax and sales tax than a worker earning $41,600. Budgets shrink at the exact moment those governments need to fund retraining programs and expanded unemployment insurance.

The BLS data shows early signs. Average weekly earnings in total private fell from $1,279.05 to $1,278.40 in March. Employers trimmed hours. The diffusion index in February sat at 49.2, meaning more industries shrank employment than grew it. March recovered to 56.8, but February showed how thin the margin is.

The ai-2027.com scenario, published April 3, 2025, predicted that by late 2026, “AI has started to take jobs, but has also created new ones.” The authors gave displacement one paragraph in a 10,000-word geopolitical narrative. They predicted the government would respond with “job training programs and unemployment insurance.”

Both parties in Washington treat AI displacement as a secondary concern behind the US-China competition. Republicans see AI as a national security asset. Democrats see the AI companies as donors. Neither has proposed legislation that addresses the transfer of value from labor to capital that the BLS data records month after month.

Retraining programs have a poor track record. The federal government promised retraining to workers displaced from steel and coal towns in the 1980s and 1990s. Researchers who studied those programs found that participants who completed retraining took large pay cuts. Many never returned to stable employment. Youngstown and Flint, forty years later, still report lower labor force participation and household income than national averages.

AI displacement moves faster and hits wider. Steel town losses played out across decades, in specific geographies, among workers with specific manual skills. AI displacement is playing out across months, in white-collar offices in most major metros, among workers whose main asset was general cognitive ability applied to information tasks. If a $20/month API subscription handles those tasks, the skill that made those workers valuable loses its market price.

Those 537,000 open tech postings on job boards in March? Research from Greenhouse, ResumeUp.AI, and HR professional surveys puts the ghost job rate at 22-30%. Companies post listings with no funded position behind them.

A 2025 LiveCareer survey found that 45% of HR professionals “regularly” post ghost jobs and 48% do so “occasionally.” Only 2% said they never do it. Among companies that post ghost jobs, 43% do it to project growth to investors. 62% do it to make existing employees feel replaceable.

Discount the 537,000 tech postings by 27% and you get about 392,000 real openings. TrueUp tracked 90,524 tech layoffs so far in 2026, at a rate of 963 per day as of early April. Factor in the skills mismatch between what the real postings require (AI/ML engineering, cloud architecture) and what the displaced workers know (database administration, project management, QA testing), and the real opening count shrinks further for the people who need it most.

What You’re Watching

The March 2026 BLS report records the early phase of a class restructuring. American workers climbed into the middle class on two employment ladders over the past eighty years. Manufacturing carried them from the 1940s through the 1980s. Information work carried them from the 1990s through now. Manufacturing towns lost their base over thirty years and never recovered.

Information work is entering the same trajectory, but the companies driving the displacement employ almost no one. Nvidia has 30,000 employees. OpenAI has 4,000. Anthropic has 1,500. Between them, they collect tens of billions in revenue from selling tools that replace hundreds of thousands of workers.

I don’t know what a viable policy response looks like. I don’t think the people in Washington writing AI policy know either, because they’re focused on China. But the aggregate payroll headline (178,000 jobs added!) hides what the BLS tables show: 661,000 fewer people working, $54/hour jobs replaced by $20/hour jobs, and 1,821,000 people stuck in unemployment for six months or longer.

You can read the BLS tables yourself. They’re free, they’re public, and they contain more information about the next decade than any AI scenario document.

Data sources: BLS Employment Situation, March 2026 | layoffs.fyi | TrueUp Layoff Tracker