The Rescue Is the Crash

Why superintelligence may arrive in time to deliver AI’s productivity gains, and why that may not save the people betting on it

Yes, superintelligence could arrive in time to deliver the productivity gains AI has promised before the bubble pops. But if it does, it probably won’t save the people who are betting on it.

In my recent post, “The Party Trick Economy,” I argued that AI’s public face is still mostly demos. Models write a working PlayStation controller in code and pass impressive benchmarks, but the gains in day-to-day business productivity haven’t shown up yet. That gap is the bubble. Trillions of dollars are being spent on the bet that the gains will arrive, and the bill comes due before they do.

There’s a hopeful counter-story making the rounds, and I’ll admit I find it tempting. It goes like this. The labs don’t need ordinary businesses to figure out AI. They need to build an AI that can improve AI. If recursive self-improvement takes off in 2027 or 2028, around when I expect artificial superintelligence anyway, the capex suddenly looks prescient. The productivity flood arrives, the skeptics eat crow, and the bubble turns out to have been a foundation.

I think that scenario is genuinely possible. I also think it rests on a confusion. It treats the bubble bursting and the gains arriving as opposing forces, racing each other. I suspect they’re the same force.

The Gains Are Real, Inside the Building

The clearest productivity gains from AI are showing up in exactly one place: the companies that build AI.

Anthropic’s IPO filing, made public in late October, is the best evidence so far. As Axios reported, the company disclosed that Claude wrote more than 90% of the contributions to its own codebase in the prior month. According to the filing’s coverage, its newest model cut customers’ cost per task by roughly 40% compared with the generation before it. Its revenue run rate went from about $4 billion in July 2025 to more than $47 billion by mid-2026.

The capability curves point the same direction. METR, the independent evaluator, measures how long a task (in human working time) an AI agent can finish on its own. That number used to double about every seven months. Recent analyses suggest it’s now growing closer to tenfold a year. The frontier models are approaching the point where METR’s own task suite can no longer measure them reliably.

This is what the early stage of recursive self-improvement would look like. It doesn’t start with AI transforming the economy. It starts with AI transforming the AI lab. The most productive workplace in the world right now is the one making the tool, and the tool’s main job there is making the next version of itself.

The Gains Are Missing, Everywhere Else

Outside the labs, the evidence for big productivity gains ranges from thin to nonexistent.

  • The executives say no. A February 2026 NBER working paper surveyed nearly 6,000 executives in the US, UK, Germany and Australia. Eighty-nine percent reported no measurable effect of AI on labor productivity over the previous three years.
  • The industry data says no. In July, Barclays told clients that industries which embraced AI showed little sign of faster productivity growth. It called the evidence “surprisingly fragile.” Only 14% of US workers said they used AI daily on the job in the second quarter.
  • The bottom line says mostly no. An MIT study found that only about 5% of firms had achieved measurable profit-and-loss gains from generative AI.

The best explanation comes from a Federal Reserve discussion paper. It found that gains at the level of individual tasks aren’t adding up across the whole economy. Making one step of a process 10% faster does nothing if the bottleneck sits three steps downstream. A paralegal who drafts a contract in half the time still waits on the same partner, the same client and the same court calendar.

In fairness, there is a case on the other side. A March 2026 NBER paper drawing on about 750 executive surveys did find positive gains in every measured sector, largest in finance and high-skill services. Economists also point to the “productivity J-curve.” Transformative technologies tend to depress measured output while companies reorganize, and the gains come later. Electricity took decades to show up in factory productivity. But “wait a couple of decades” is not a timeline a $5 trillion capex cycle can survive.

The Race Between Two Clocks

The hopeful scenario comes down to which of two clocks runs out first.

The first clock is capability. It runs inside the labs, and it’s speeding up. The second clock is patience: how long investors, lenders and boards will keep funding a buildout whose returns haven’t shown up yet. As The Economist put it, until business adoption speeds up a lot, the revenue needed to justify roughly $5 trillion in AI capex stays out of reach.

Here’s why I take the hopeful version seriously. The labs don’t need the rest of the economy to adopt AI in order to reach ASI. They need compute, talent and runway. A public listing buys a lot of runway. If the self-improvement loop closes before the money runs out, the diffusion problem changes character. Today, adoption stalls on integration: connecting systems, retraining staff, rebuilding workflows. A system smart enough to do that integration itself removes the bottleneck the Fed paper describes.

Even then, some bottlenecks don’t care how smart you are. Data centers need electricians, permits and grid connections, and analysts already flag a shortage of skilled trades as a constraint. Insurers have to underwrite AI for mission-critical work, and I suspect companies will trust it well before insurers do. Courts, unions and regulators move at human speed. ASI could compress the timeline. It can’t erase it.

So the race is real, and it’s close enough that I won’t call it. But suppose the hopeful side wins. Then the more interesting question is what “winning” actually means.

Whose Bubble?

“The AI bubble” isn’t one bubble. It’s a stack of bets, and ASI pays off very few of them.

Start with who’s holding what. The frontier labs are at the top. Below them are the hyperscalers renting out compute and the chipmakers selling it. Then come the data center developers and the lenders financing them, then thousands of startups wrapping a model in an interface. At the bottom are pension funds and index investors who own a slice of all of it. When people ask whether ASI arrives “before the bubble bursts,” they usually picture all of these layers getting rescued together. History says that isn’t how it works.

The railways. Britain’s Railway Mania of the 1840s roughly tripled the country’s rail mileage. The new lines cost more than forecast and earned less. The mathematician Andrew Odlyzko, who has studied the episode closely, says it was a disaster for most investors. About a third of authorized lines were never built. And yet the railways that did get built carried Britain’s economy for the next century. The technology won. The people who financed it mostly lost.

The fiber. The pattern repeated in the dot-com era. More than 60 telecom companies failed between 2001 and 2003, leaving hundreds of billions of dollars of fiber in the ground, unlit. That “dark fiber” later became the backbone of the modern internet, and now of the data centers themselves. The bandwidth boom was real. It just arrived after the companies that built for it had gone bankrupt.

ASI sharpens this pattern instead of breaking it, because there are two ways it can play out. In one, a single lab gets there first and the field becomes winner-take-most. Everyone who bet on the other labs, and on the startups built on top of them, gets wiped out. In the other, several labs get there at about the same time and intelligence becomes a commodity. A 40% drop in cost per task with every generation is a preview of what happens to margins then. Either way, most of the $5 trillion stack doesn’t get its money back. And that leaves aside the possibility, which I’ve written about before, that a true ASI may not stay ownable at all.

The Rescue Is the Crash

Here’s the part the hopeful scenario skips. The productivity gains that would justify the AI buildout are, by definition, a labor shock. A shock that big hits the same economy that’s supposed to pay for the buildout.

In “The Party Trick Economy,” I argued that an AI trustworthy enough for mission-critical work would snap the economy’s backbone, because every company would rush to replace workers at once. That is the promise of the boom, stated plainly. When economists say “productivity gains,” they mean the same output from fewer people. At ASI scale, that means a lot fewer people, very quickly.

Now follow the money. Wages are most consumers’ income. Consumer spending drives the ad markets, retail sales and subscriptions that pay for most of the software and cloud services AI is supposed to supercharge. A company that replaces half its back office with agents sees its margins jump. An economy where every company does that at once sees demand fall. Measured productivity could soar in the same quarters that GDP, consumer credit and corporate revenue fall apart.

Then there’s politics. A labor shock that size won’t go untaxed. I’ve argued before that ASI-generated wealth will force questions about taxation, dividends and representation that no one is ready for. Windfall taxes, mandatory dividends and partial nationalization all stop being fringe ideas in a world where a handful of companies own the means of producing intelligence. Every one of those policies comes straight out of the returns investors were counting on.

So to the market, the “rescue” scenario looks like a crash with a different cause. The story wouldn’t be that AI didn’t work. It would be that AI worked too well, too fast, for the economy around it to absorb. Chart the S&P 500 and the two scenarios might look the same. The difference is what the headlines blame.

What to Watch

If you want to know which way this breaks, these are the signals I’d follow over the next 18 months.

SignalWhat it would tell us
Labs’ own disclosures about AI-written code and AI-run researchWhether the self-improvement loop is actually closing
METR time horizons outrunning METR’s ability to measure themWhether capability is still compounding faster than diffusion
Share of workers using AI daily (14% in Q2 2026)Whether adoption outside the labs is finally speeding up
Entry-level white-collar hiring and unemploymentWhether the gains are arriving as a labor shock
Insurers writing policies for AI in mission-critical rolesWhether institutions trust the technology, not just the demos
Price per task falling with each model generationWhether intelligence is commoditizing and margins are collapsing
Lab revenue growing faster than their customers’ revenueWhether value is concentrating rather than spreading

The worrying combination is the first two signals racing ahead while entry-level hiring collapses, before adoption and insurers catch up. That would mean the gains arrived, landed on workers first, and the economy couldn’t absorb them.

The Real Question

So, is it possible that ASI arrives in time to deliver the productivity gains before the bubble bursts? Yes. I’d even say it’s more likely than most bubble-watchers think, because the gains are already compounding in the one place that matters for ASI: the labs.

But the question assumes the bubble and the gains are rivals, and I don’t think they are. The Railway Mania gave Britain its railways and took its investors’ savings. The telecom crash gave us the internet’s backbone and took Global Crossing with it. ASI could give us the productivity the boom promised and, in the same stroke, destroy the economy that was supposed to cash it in.

The question isn’t whether the gains arrive before the burst. It’s who’s standing where when they do, and whether we’ve built anything by then to catch the people who aren’t holding the winning ticket.

Sources

Author: Shelton Bumgarner

I am the Editor & Publisher of The Trumplandia Report

Leave a Reply