How the AI crash happens, revisited

The Atlantic piece is old in AI time. It ran on October 30, 2025.¹ Time passed may make it more useful, though, because some parts of the argument look weaker a few months later while others look even harder to dismiss.

The weakest part, to me, is the superintelligence premise. The article treats the race to build a superintelligent machine as one of the reasons the spending might be rational. Ignore it. There is no proven path from current systems to AGI, and the public case for it remains deeply entangled with scaling-law extrapolation, capital raising, and the professional incentives of people who need the story to remain large. Plenty of serious and competent people still do not buy the premise at all. I tend to side with them, although I admittedly lack the expertise to have anything more than “just an opinion, man.” My take: disregard AGI, or “superintelligence” as an economic destination that justifies anything – until proven otherwise.

The stronger argument is more ordinary, and therefore more worrying: the AI boom has become large enough to matter outside the AI industry. The article frames the U.S. as becoming an “Nvidia-state,” pointing to the degree to which AI-related stocks, data-center construction, and chip demand are holding up the broader market story.¹ Even if the exact numbers move month to month, the structure of the concern still holds. A narrow set of companies, suppliers, infrastructure bets, and financing arrangements now carries an uncomfortable amount of the growth narrative.

The circularity is the part that keeps sticking, referred to as “sophisticated financial engineering.” It looks like a machine in which everybody’s valuation is supported by everybody else’s promised spending. Ed Zitron’s Nvidia section in “The Hater’s Guide To The AI Bubble” is useful on this: Nvidia as the market’s weak point, the Magnificent 7’s dependence on GPU buying, and the way the AI trade leans on a small number of companies continuing to spend.² See also Pinocchio’s world.

I guess, the economy can absorb a lot of experimentation when the bill is hidden, subsidized, or treated as venture-backed learning. The mood changes when the token bill arrives. We are getting there now. Uber reportedly capped employee spending at $1,500 per month per agentic coding tool, including Claude Code and Cursor.³ Simon Willison’s comment on that is worth reading: he sees the cap as a rational response to overspending, but also as a real signal about what Uber thinks these tools are worth.⁴ If you assume two actively used tools, that is a $3,000 monthly cap per engineer, before the rest of the software bill. I have a hard time believing it, or that “it” will be good for business as in the bottom line.

The fairest update since October is that the counter-case has improved. The Atlantic itself published a follow-up in May 2026 arguing that revenue and demand were catching up faster than many bubble skeptics expected, especially around coding agents.⁵ That matters. Claude Code, Codex, Cursor, and related tools have made the value proposition much easier to see for software teams than it was during the chatbot phase. If AI becomes a genuine production tool for software, and then for adjacent forms of knowledge work, some of the capex may be less insane than it looked. Maybe. But some of the best coders in the world have come to doubt that, after no lack of trying, George Hotz:

I’m calling it now, the adoption of AI agents into software development will be one of the most costly mistakes in the field’s history. Agents cannot program, and it’s taking longer and longer to realize that they can’t. They are a highly sophisticated statistical model designed to mimic the distribution of programming. The output is broken, but in a way that’s getting harder and harder to detect. Which is exactly what you’d expect from an increasingly accurate statistical model.

Either way, the people who already converted the boom into wealth are not the ones most exposed to the fallout. The executives, founders, investors, and incumbents who rode the valuation wave have already had their payday. The losses, if, or when they come, will travel through markets, pensions, jobs, public budgets, energy systems, and all the places where ordinary people meet macroeconomic “adjustment.”

That is what connects this to my earlier note, “Heading for the cliff”. I do not want to become professionally gloomy about this stuff. There is something personally corrosive about staring at these incentives for too long. But a decent outcome is hard to picture. The industry’s promise is either too small to pay for what is being built, or large enough to tear through labor markets before anyone has adapted. Both versions can be true in different places at once.

I would discount the AGI story. I would discount the idea that scale itself proves destiny. What I would not discount is the financial shape of the thing: circular spending, fragile valuations, visible economic dependency, and a bill that has to land somewhere.

The people who benefited first are rarely the people who pay last.

Footnotes

1. Matteo Wong and Charlie Warzel, “Here’s How the AI Crash Happens,” The Atlantic, October 30, 2025. ↩ ↩²

2. Ed Zitron, “The Hater’s Guide To The AI Bubble,” Where’s Your Ed At, July 21, 2025. ↩

3. TechCrunch, “Uber caps employee AI spending after blowing through budget in four months,” June 2, 2026. ↩

4. Simon Willison, “Uber Caps Usage of AI Tools Like Claude Code to Manage Costs,” June 3, 2026. ↩

5. Rogé Karma, “So, About That AI Bubble,” The Atlantic, May 1, 2026. ↩