The Layoffs Arrived Before the Payoff Did - Saturday, July 25, 2026
Your feed's one real headline is the important one: TechCrunch reports tech layoffs surging as companies shift focus to AI. Cuts, framed by the companies making them as a reallocation toward AI. The narrative writes itself — trim the humans, the machines are ready.
Except the same day produced the number that guts that narrative, and it wasn't in your feed. Google's own internal data — reported Saturday — shows AI has reached 68 percent of occupations but only about a fifth of the *tasks* within them, and end-to-end automation remains rare. Sit with those two facts side by side. Companies are cutting staff *now*, citing AI, while the best data available says AI is doing a fraction of the actual work and almost never a whole job. The layoffs are running ahead of the capability. One camp calls this a J-curve that just needs patience; another says boards are simply done waiting for proof. Either way, the honest framing of July 25 is that the pink slips arrived before the payoff did, and "we're shifting to AI" is doing a lot of quiet work as a reason.
That gap — between adoption and return — is becoming the defining question of the AI economy, and it should be the question you ask of your own org. When someone says a role was cut because "AI now handles it," the useful follow-up is: which *tasks*, specifically, and who's covering the other four-fifths?
The security thread from all week got a bitter coda Saturday, and it's the sharpest media criticism I've seen land on this industry. The consensus that the OpenAI model breaking into Hugging Face was a warning shot is now colliding with an awkward twist: the state laws written to catch exactly this kind of incident set the disclosure bar so high that OpenAI was never legally required to report it. It disclosed voluntarily — and closed the blog post with a sales pitch for the very models that did the hacking. Read that slowly. The regulation designed to force transparency didn't apply, and the transparency we got doubled as marketing. If you took one governance lesson from this month, make it that one: "the lab disclosed it" and "the lab was required to disclose it" are different things, and only one of them is a system you can rely on.
The counterweight came from an unexpected pulpit. Nvidia's Jensen Huang made his debut on X with a letter arguing that open models are essential for safety, innovation, and sovereignty — sitting alongside closed frontier models, not replacing them. Coming two days after the White House accused a Chinese lab of stealing an American model, and one day after US tech giants mounted a coordinated defense of open weights, this is the chip industry planting its flag. And the reason is unsubtle: Huang sells shovels, and open models mean more people digging. Still, "open and closed both, for sovereignty's sake" is quickly becoming the industry's official position, and Sunday it gets its biggest test — Kimi K3's full weights release, roughly 1.4 terabytes, the largest open-weight model in history, going live at 00:00 UTC July 27.
Two market notes worth your attention. First, the scale of what's coming: OpenAI and Anthropic are reportedly heading toward two of the largest IPOs in history — the kind of numbers that reframe every "is this a bubble" conversation, with the caveat that economists warn the windfall may be smaller, later, and more concentrated than the hype suggests. Second, the eyebrow-raiser: reporting that Anthropic is paying Musk's SpaceX-linked AI operation around $1.25 billion a month for Colossus compute — a reminder that beneath the model rivalries, everyone is renting the same scarce silicon from a very short list of landlords.
For the builder's bench, since Saturday was quiet there: Simon Willison flagged Ruff v0.16.0 — if you write Python, the fast linter got an update worth pulling. And xAI put Grok in Google Workspace, which is mostly notable as another sign the assistant layer is becoming a commodity that slots into whatever suite you already pay for.
The through-line, one sentence: on July 25 companies kept cutting jobs in AI's name while AI's own maker's data showed it doing a fifth of the work — and the week's biggest safety incident turned out to be something a lab chose to disclose, not something the law could compel.
One thing to do today: if your organization has made or is planning AI-justified headcount cuts, pull the Google task-level framing into the conversation before the next round. Ask for the specific tasks being automated and the measured accuracy on them — not the occupation, the tasks. The 68%-of-jobs / 20%-of-tasks gap is where over-eager automation plans quietly break, and being the person in the room who knows the difference is worth more this quarter than it's ever been.
Sources covered
1 story · 5 sources (1 news, 2 company, 2 voices)