The CEOs Walked It Back. The Layoffs Didn't.
The same CEOs who sold boards on replacing people with AI are now walking it back. The receipts from Microsoft, Nvidia, Uber, and the BLS say what I've been saying all along: AI is a tool that makes talented people better, not a one-for-one swap for them.
Over the last six months I’ve run into the same conversation more times than I can count, maybe six, maybe more. It usually opens with an executive, sometimes a founder, sometimes someone on a board, telling me they’re carrying too many people. Too much headcount, too much cost, and now that AI can handle a big chunk of the work, it’s time to cut. There’s almost always a number they’ve already circled. And there’s almost always a name they point to, someone who said it first.
For a while, that name was Dario Amodei. Anthropic’s CEO spent a good chunk of last year warning about a “white-collar bloodbath,” the idea that within a few years AI might erase half of entry-level white-collar jobs. Sam Altman floated similar warnings from the OpenAI side. When two of the most credible voices in the industry talk like that, boards pay attention, and a fair number of them went past attention into actual headcount decisions.
So when one of those conversations lands in my lap, I give back roughly the same answer every time, give or take. Sure, there are always efficiencies to be gained with AI, and you should always make sure you’ve got the right people in the right spots. But you hang onto the people who have the knowledge, the skill, and the technical ability your company needs to stay competitive and stay an industry leader. You didn’t get where you are because you were so awesomely hyper-efficient with automation. Every good startup starts with a group of really talented people, and when the company actually works, those people are still the center of it. So keep that in mind before you start making irreversible calls about your future.
I’ve been making some version of that argument for about a year. This week it got a whole lot easier to make.
The Walk-Back
A few days ago, speaking in Australia, Sam Altman said he’d been “pretty wrong.” He put it about as plainly as you can: “I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened.” That’s the same Altman whose earlier warnings kept getting repeated back to me in budgeting meetings.
Amodei had softened his own forecast a couple weeks before that. The new framing isn’t “half your staff is gone.” Now it’s that AI automates something like 90% of a given job, and the leftover 10% stretches out to fill the whole role and makes people 10 times more productive. Read that one slowly, because it’s a full reversal in spirit. What got pitched as a bloodbath turned, pretty quietly, into a productivity story. Same person, opposite conclusion, about a year and a half apart.
I’ll be fair about it. Changing your mind when the evidence shifts is a strength, not a weakness, and if the data is what pushed Altman and Amodei to revise, then good, that’s what they should do. I’m not interested in scolding anybody for updating.
But the new view lands in a world where a lot of leaders already acted on the old one. And there’s a detail I keep coming back to: the same week Altman called himself “pretty wrong,” a survey made the rounds claiming 99% of executives still expect AI-driven layoffs over the next two years. The boardroom conviction is now running out ahead of the very people who planted it. The CEOs are easing off the gas while the market keeps it floored.
The Receipts Nobody Planned For
If you’re a leader who cut hard, this next stretch is the uncomfortable part.
The story isn’t getting revised because AI turned out to be a dud. It’s getting revised because the replacement math was never as clean as it got marketed, and the evidence showing up this month makes that pretty concrete.
Microsoft cancelled most of its licenses for Claude’s coding tool and pointed its engineers at GitHub Copilot CLI instead. Nvidia made a similar call. Neither of these companies is exactly dragging its feet on AI, they’re about as sophisticated as it gets, and they still pulled back on agentic coding tools for a reason their own people said out loud. Here’s Nvidia’s VP of applied deep learning, Bryan Catanzaro, to Axios: “For my team, the cost of compute is far beyond the costs of the employees.”
Sit with that one for a second. The whole pitch, said and unsaid, was that AI would come in cheaper than the humans. At Nvidia, on real work, the compute is costing more than the people it was supposed to replace.
Uber’s COO said the company burned through its entire 2026 AI budget by the end of April, and even after spending that fast, they still can’t draw a line from the spend to better products. “That link is not there yet,” he said. And MIT Technology Review, working off Bureau of Labor Statistics data, found “scant evidence that AI has yet had any large-scale impact on the U.S. labor market.” For the jobs that are supposed to be most exposed to AI, the unemployment rate is actually lower than it is for the jobs people consider safer.
I’ve written versions of all this before, scattered across a few posts. That compute costs aren’t headed for zero, in Tokens Aren’t Going to Zero. That frontier AI still trips over real engineering, and that every agent session starts over from scratch, in The Tokens-for-Engineers Trade Has Three Holes. That the AI bill is turning into a top line item before anyone’s built the governance to handle it, in When Your AI Bill Becomes Your Biggest Line Item. I’m not running a victory lap here. The point is just that the proof isn’t coming from me anymore. It’s coming from Microsoft, Nvidia, Uber, and the federal labor numbers.
What the Narrative Actually Cost
The false story was never that AI is powerful. It is powerful. The false story was that powerful automatically meant a clean one-for-one swap, where you cut a person, buy a token budget, and pocket the spread. That’s the version that got sold, and it’s the version a lot of its loudest sellers are now backing away from, quietly.
The leaders who bought it are the ones left holding the consequences.
If you trimmed your team last year on the theory that AI would soak up the work, here’s roughly where you tend to land now. The tools picked up some of the work, not all of it. The cost model you ran for AI wasn’t as close to reality as you’d hoped. And the people who carried the knowledge of how your systems actually fit together, the ones who knew what quietly breaks when you touch the other thing, they’re gone. You can buy more tokens next week. You cannot buy back 15 years of somebody knowing why the company does it this way instead of that way.
The part that worries me most is the early-career squeeze, and it’s the one the labor data hints at. Even with no broad displacement showing up in the headline numbers, the 22-to-25-year-olds trying to break into software and analyst roles are getting hit hard. A company decides AI can handle “junior work,” so it stops hiring juniors. Fine, maybe that pencils out this quarter. But your seniors came from somewhere. Stop hiring and training people at the bottom and in five years you’ve got nobody in the middle, and in ten you’ve got nobody at the top. That’s not a cost saving. That’s eating your own pipeline and writing “efficiency” on the receipt.
Right People, Right Spots, Better Tools
None of this is an argument against AI. I’d be about the last person to make that one. AI belongs inside nearly every workflow a serious company runs, and the leaders refusing to touch it are setting up a different and equally expensive mistake, which is a post for another day.
The real argument is about how you hold the thing. Is it a tool that makes your people better, or is it something you expect to swap in for them one-for-one? Those are two genuinely different strategies, they lead to two genuinely different places, and that gap is most of the whole game.
Put AI in the hands of talented people and you get a team that ships more, moves faster, and keeps compounding the institutional knowledge that already makes you hard to beat. Try to swap those talented people out for AI and you tend to end up with a thinner team running tools nobody fully understands, a compute bill bigger than you forecast, and an expensive rebuild a year or two down the line, in a tighter labor market, paying a premium to buy back the exact capability you just let walk.
The CEOs who pushed that second approach are the ones now saying they were “pretty wrong.”
You don’t have to wait around for them to say it twice. Keep the right people in the right spots. Hand them the tools. And remember that you didn’t get where you are because you were so awesomely hyper-efficient with automation. You got here on the backs of a group of really talented people. That part hasn’t changed, and it’s going to outlast this month’s prediction, and probably the next walk-back after it too.