The 21,000-Person Quiet Layoff and the AI Non-Revolution

The Severance Slide

Twenty-one thousand people.

That is roughly the population of Geneva, or Galveston, or the entire graduating class of a large university system. According to Oracle's fiscal 2026 annual report, released in late June, the company shed that many employees in a single fiscal year — a 13% workforce drop, taking headcount from 162,000 down to 141,000.

The filing notes the cuts were driven by "management and product changes, performance issues, strategic shifts and acquisitions." AI gets one mention: the workforce decline was "partly driven by the adoption of AI across its operations." Oracle spent $1.84 billion on severance and exit costs this fiscal year, roughly five times the $374 million it spent the year prior.

You read that correctly. Twenty-one thousand jobs gone, and AI shows up in the filing like a polite footnote.

Layoffs.fyi tracks another 119,800 cuts across 196 tech companies so far this year. The aggregate story is no longer about any one employer.

The market is less polite than the 10-K. CNBC reported on June 26 that Oracle's stock had just finished its worst week since August 2001 — a 19% slide in five trading days, the kind of drop normally reserved for the dot-com bust itself. Capital expenditures surged 162% to $56 billion. Free cash flow went negative to the tune of nearly $24 billion. Debt sits at $130 billion. The company is planning to raise another $40 billion in debt and equity next fiscal year to keep building data centers for the AI workloads that are, in the same breath, replacing the humans on its payroll.

The setup is almost baroque. Oracle is borrowing against the AI future to build the AI present, while shrinking the workforce the AI present no longer needs. Larry Ellison, the co-founder, has slipped down the world's-richest list behind Larry Page, Sergey Brin, Jeff Bezos and Michael Dell. Seventy-one percent of analysts still recommend buying the stock. They are betting on AI scaling. The org chart is no longer arguing.

The Khanmigo Non-Event

Now flip the page.

In 2023, Sal Khan stood on a TED stage and declared that AI was about to deliver "probably the biggest positive transformation that education has ever seen." Khan Academy's chatbot, Khanmigo, was pitched as "an amazing personal tutor" for every student on the planet. Sam Altman chimed in shortly after, promising AI tutors capable of "personalized instruction in any subject, in any language, and at whatever pace they need."

Three years later, Khan called Khanmigo a "non-event."

This piece in The Atlantic, published the same week as the Oracle filing, walks through what actually happened. Access exploded — from 40,000 students in 2023 to nearly a million this year. Actual uptake did not. A Stanford review of all the available research on AI in K–12 schools concluded the educational benefits were generally limited. Laurence Holt, author of The Science of Tutoring, calls it the "5 percent problem": about 5% of students use ed-tech tools as intended, which means the tools tend to widen inequality by supercharging the already-motivated rather than democratising access.

Khan himself has hedged the digital-transformation talk. "I think our biggest lever is really investing in the human systems," he said in an April interview. The AI tutor turned out to be very good at emitting math problems and very bad at making a fifteen-year-old care about doing them.

The Pattern, Not the Hype

The two stories sit next to each other for a reason.

AI is excellent at work that is mechanical, structured, codifiable, and most importantly already documented in a giant internal org chart. Enterprise software had that profile in spades. So does insurance underwriting, mortgage processing, basic legal discovery, and a long tail of back-office operations. The 21,000 people Oracle shed were, in many cases, doing work that could be written down — which is the only kind of work a language model is reliably good at.

AI is poor at work that is relational, embodied, situational, and resistant to documentation. Teaching is the canonical case. So is most of nursing, most of skilled trades, most of management above a certain level, and most of the jobs where the deliverable is getting another human to care about something. Khan Academy's "non-event" is, mechanically, the same observation the AI-in-medicine and AI-in-construction worlds keep rediscovering: a chatbot can prescribe, but it cannot persuade a patient to take the prescription; it can coach, but it cannot make a student sit down and try.

The headline version of AI disruption has been, for three years now, "every industry is about to be transformed." That framing has aged poorly. What is actually happening is sharper and more useful: AI is rewriting the industries whose work was already halfway written down, and leaving largely intact the ones whose work was never written down in the first place.

A 21,000-person layoff and a chatbot tutoring non-event, in the same week, is the cleanest evidence of that line you are going to get this year. It is not a story about AI replacing humans. It is a story about AI being very, very good at replacing a specific subset of humans whose jobs were already half-articulated in a wiki.

The disruption is real. It is just uneven. And the industries that think they are next on the list might want to check which side of that line they actually sit on.

The org chart knows. The market knows. Khan Academy, eventually, figured it out.