What Would an AI Crash Look Like?

The key is that a bubble can form even if lots of investors are rational. So long as there are enough investors trading on momentum — or blind optimism — sophisticated traders will find it profitable to ride the bubble rather than to bet against it. And while bubbles pop, only some set off wider economic and financial crashes like the one the world endured in 2008.

That account comes from A Crash Course on Crises: Macroeconomic Concepts for Run-ups, Collapses, and Recoveries (Princeton University Press, 2023) by Markus Brunnermeier of Princeton and Ricardo Reis of the London School of Economics. At an admirable 111 pages, it’s a great resource — and one I’ve gone back to as talk of an AI bubble has returned in recent weeks. But Crash Course is about “macro-financial crises,” not bubbles per se. That makes it helpful not just for thinking through whether there’s an AI bubble, but for imagining what an AI crash would look like.

Crash Course sketches three models of the run-up to a macro-financial crisis, and each one suggests things to watch for with AI.

The first is the classic speculative bubble already described, and in the model it’s not enough for everyone to be overly bullish about AI. The real bubble dynamic comes from the combination of naive investors extrapolating — line goes up — combined with investors who know valuations are out of whack but hang on for fear of selling too early. This is what a lot of the AI bubble worry has been about so far: As my colleague Edward Harrison wrote of the AI-heavy S&P 500 in August, “We’re in the unusual situation where fund managers almost uniformly say US stocks are overvalued, yet everyone is piling in.”

That sounds bubblicious, and it’s the dynamic that “gets the party going,” says Reis. But it’s not enough on its own to create the sort of crash that he and Brunnermeier write about.

The second model in the book involves misallocation — investors mistakenly backing the wrong firms in the hot sector because some policy or distortion biases what gets funded. The third concerns the role of shadow banks that are vulnerable to runs and amplify dips in asset prices.

Applied to AI, Crash Course’s brief tour of crisis economics suggests three gauges of the risk:

1. Are investors buying or lending to AI firms without any conviction except the belief that they can unload later to the ‘greater fool’?
2. Are there barriers — like, say, VCs investing in anything labeled AI — that are preventing money from reaching the most promising firms?
3. How much of the AI boom is funded by debt and and are those lenders vulnerable to runs?

‘Yes’ to the first one would signal a bubble. ‘Yes’ to all three would warn of a crash.

Predictions

The AI boom will require lots of nuclear power: “Soaring demand for electricity will drive a $350 billion nuclear spending boom in the US,” says Bloomberg Intelligence. — Will Wade, Bloomberg News  

The material of the 21st century will be… metal-organic frameworks, for which the Nobel Prize in Chemistry was awarded this week. The recipients’ work pioneered “creating molecular constructions with large spaces through which gases and other chemicals can flow… [with potential applications like] capturing carbon dioxide and harvesting water from desert air.” — Charles Daly, Bloomberg News

The risk of another World War is growing: “Memories of the last world war have faded, and the current generation of leaders and experts — from China and Russia to the US and elsewhere — is showing signs of waning humility and growing hubris, similar to European leaders in the summer of 1914.” — Andreas Kluth, Bloomberg Opinion

NATO may build a “drone wall” along Europe’s eastern border. It would take years, and no one is quite sure how it would work. But the idea is gaining momentum and would involve a system that would likely take cues from Ukraine. — Gerry Doyle and Jake Rudnitsky, Bloomberg News

The dollar isn’t going anywhere: It’s still on one side of 89.2% of all FX trades, according to the BIS. — Daniel Moss, Bloomberg Opinion (Although, counterpoint: Gold’s rally is helping China challenge the dollar.)

What Are the Chances…

Every quarter we look back at the markets cited in the newsletter: Here’s Q1 and Q2. The second quarter was rough in terms of accuracy, as lots of plausible-but-unlikely events kept happening — mostly involving Donald Trump.

This quarter markets did better. Seven markets that we’ve cited closed, five to ‘Yes’ and two to ‘No.’ At the time we cited them, the average forecast for the Yes outcomes was 67%, and for the No outcomes 36%. Six out of the seven leaned in the direction of the actual outcome. Traders expected the Fed to cut in September and Francois Bayrou’s ouster as France’s prime minister. Way back in November of last year they expected Trump to implement “large” tariffs.

As for misses, Polymarket put just a 35% chance on Jimmy Kimmel returning to air by mid-October and that happened.

Here are a few other markets we’ve cited and where they are now:

Amateur Forecasters vs. Traders

Twice this year we’ve cited forecasts about a potential ceasefire between Russia and Ukraine — in May and in August — and both times we compared the real-money trading platform Polymarket to the no-money crowd forecasting platform run by the think tank Rand. (Disclosure: I was once a paid forecaster and freelance writer for the platform before it was absorbed by Rand.)

In May, Polymarket put the chances of a ceasefire this year at 45%, while Rand’s consensus forecast was just 26%. In August, Polymarket was at 35% while Rand was at 11%.

Today, Polymarket is much closer to Rand: Polymarket puts the chances at 11% while Rand is down to 4%. We won’t have an official outcome until the end of the year, of course, but right now it looks like the amateur forecasters were ahead of the traders.

Keep an Eye On

The big news in prediction markets this week was that Intercontinental Exchange Inc., owner of the New York Stock Exchange, announced that it was investing $2 billion into Polymarket. (You can read more about what that means for prediction markets and for finance from Bloomberg Opinion’s Matt Levine, John Authers and Aaron Brown.)

Perhaps just as important, though, was a report from the Forecasting Research Institute, a think tank.

FRI reports that large language models now outperform the median human forecaster, a reversal from a year ago. The very best human forecasters still were more accurate than AI, but FRI’s researchers project that will reverse by 2026 as AI continues to improve.

Other forecasting organizations have seen similar trends. Last month the forecasting website Metaculus reported that the startup ManticAI had secured 8th place in its summer competition, the highest result by an AI system to date.

Prediction markets can be a great way of harnessing the dispersed wisdom of many human minds. But they may be headed to the mainstream just as AI is proving its ability to exceed the wisdom of crowds. In the meantime, as more investment flows to prediction markets, expect traders there to rely more and more on AI to make their predictions.