AI Costs Now Top Payroll, and That’s Saving Jobs

Artificial intelligence (AI) was pitched to boardrooms as the cheaper hand: software that works around the clock for less than a salary. In 2026 the bills tell a different story. At several large employers, running the AI now costs more than paying the people it was built to replace, and that arithmetic is quietly reshaping the layoff wave everyone was bracing for.

The reprieve for human workers is real, and largely accidental. Companies are not keeping staff out of mercy; they are keeping them because the meters attached to every AI agent kept spinning long after this year’s budget ran dry.

The Compute Bill Came In Higher Than the Payroll

The clearest signal came from inside the industry that builds the technology. Bryan Catanzaro, vice president of applied deep learning at chipmaker NVIDIA, described the economics of his own engineering team in terms that undercut the entire cheap-robot pitch.

For my team, the cost of compute is far beyond the costs of the employees.

When the company whose chips power the boom concedes that its software habit costs more than its engineers, the cost case for mass automation starts to wobble. Microsoft offers the corporate version of the same lesson. It canceled most of its direct Claude Code licenses roughly six months after handing access to thousands of employees, moving engineers onto a cheaper in-house coding assistant once usage scaled past what it wanted to pay.

Uber went further into the red. Praveen Neppalli Naga, the ride-hailing firm’s chief technology officer, said it had burned through its entire 2026 budget for AI coding tools in just four months, even while it pushed staff to use them harder through internal adoption leaderboards. Three of the most aggressive AI adopters in corporate America, in other words, hit the spending ceiling before the year was half over.

Goldman’s Math: 25,000 Out, 9,000 Back

None of this means the cuts stopped. A note from Goldman Sachs, authored by economist Elsie Peng, pulled apart the two forces AI exerts on jobs and found both running at once. The takeaways are sharper than the panic headlines suggested.

• 25,000 jobs a month wiped out by AI substitution, where software replaces a worker outright, over the past year.
• 9,000 jobs a month added back through augmentation, where AI makes existing staff more productive.
• 16,000 net monthly displacement, with the pain falling hardest on Gen Z and entry-level white-collar workers.
• 22,000 to 28,000 net monthly losses projected by the fourth quarter of 2026 as agents grow more capable.

That net figure is smaller than the doom forecasts, but it is not nothing, and the cuts are very real where they land. Meta’s move to cut 8,000 jobs and reassign 7,000 staff into new AI teams shows how a single company can deliver a month’s worth of national displacement on one Wednesday. You can read the full breakdown in Goldman Sachs’ analysis of AI and the US labor market, which leans on actual payroll data rather than projections.

Altman Steps Back From the Apocalypse

Even the executive who helped popularize the fear is dialing it down. Speaking at a Commonwealth Bank of Australia conference in Sydney on May 26, OpenAI chief executive Sam Altman said the rapid job collapse he once warned about has not arrived on schedule.

“I’m delighted to be wrong about this,” Altman told the audience, adding that “I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened.” He has previously said entire job categories would be “totally, totally gone.”

What shifted his thinking was small and human. After he experimented with letting AI answer his messages under a label noting the replies were automated, the response convinced him that people still value genuine contact at work. The augmentation side of Goldman’s math points the same way: AI is good at speeding people up, less reliable when left to run alone. One study where AI beat human forecasters at startups found that blending the two actually made predictions worse, a reminder that handing the wheel entirely to the machine often backfires.

So the reprieve has two engines. The technology cannot yet do the whole job, and where it can, the bill arrives faster than the savings.

Where the Money Goes: Tokens, Agents and the 85% Problem

The cost spiral is structural, not a billing glitch. Inference, the computing needed to actually run a model after it is trained, now eats roughly 85% of the average enterprise AI budget. And those budgets have ballooned, even as the price of an individual unit of AI output keeps falling.

The paradox sits in those last two rows. Token prices, the per-unit cost of AI output, are projected to fall nearly 90% by 2030, yet total bills keep climbing because the new agentic systems consume far more tokens per task than a simple chatbot. Goldman expects agentic AI to drive a 24-fold jump in token consumption by 2030, reaching about 120 quadrillion tokens a month. NVIDIA chief Jensen Huang has floated a future of 100 AI agents working alongside each employee; at current consumption rates, that is a future of 100 running meters per desk.

The spending itself is not slowing. Gartner’s 2026 worldwide AI spending forecast has agentic software outlays rising 141% this year alone. What is slowing is the willingness to let those bills replace people who cost less to keep.

Who the Cost Wall Protects, and Who It Doesn’t

The ceiling does not shield every worker equally. It buys time for some roles while leaving others exactly where they were.

• Customer-facing and relationship roles, where Altman’s point about human contact carries real commercial value, look the safest.
• Senior staff whose judgment is hard to encode into an agent remain largely untouched by the automation push.
• Entry-level and Gen Z white-collar workers stay the most exposed, the very group Goldman flags as bearing the brunt.
• Routine document and coding tasks sit in the middle, automated where the math works and reversed where the monthly bill spikes.

That mix explains why the labor data looks contradictory. A firm can lay off a graduate analyst whose work an agent now handles cheaply, then quietly cancel an expensive automation pilot in the next department over because the compute ran hot. The cost wall is not a policy. It is a spreadsheet, and it protects whoever happens to be cheaper than the tokens that week.

What Resets the Clock

None of this is locked in. Much of today’s AI pricing is propped up by venture capital (VC) money and by cloud giants subsidizing usage to win market share. As that capital discipline tightens, analysts expect the pricing on the application programming interfaces (APIs) that businesses build on to normalize, potentially pushing inference costs up 30% to 50% within 12 to 24 months.

Higher prices would, oddly, protect more jobs by making automation even less affordable. The opposite risk is efficiency: if the next generation of models makes each task cheap enough to run at scale, the economics flip back toward the machine, and Goldman’s fourth-quarter projection starts to look conservative.

For now, workers are caught between a budget and a breakthrough. If the next models make every task cheap enough to run flat out, Goldman’s 22,000 figure becomes the floor rather than the ceiling, and this reprieve ends without an announcement. If the meters keep running ahead of the savings, the spreadsheet that paused this round of cuts will pause the next one too.