History suggests an AI financing crash is on the cards. But at what scale – and what comes after?
The Bank for International Settlements – the central banks’ central bank – has put an AI bust at the top of its list of threats to the global financial system, naming the unwinding of “circular” AI deals alongside inflation and sovereign debt. There is roughly $1.8 trillion in opaque, often off-balance-sheet financing woven through the AI build-out. Chipmakers have stakes in the labs buying their chips, infrastructure is leased back on contracts with hidden exit clauses, government money is involved, the same collateral is quietly pledged more than once.
History is instructive. Every technological revolution – railroads, electricity, fiber, the dot-coms – has paired game-changing technology with a bubbly financing cycle. The tech survives even as its financiers go bust. We now know with a high degree of confidence that AI will create an inflection point; a moment when order-of-magnitude change takes place. That makes point predictions useless. Instead, consider the robustness of your strategy against multiple possible futures.
There are two questions we cannot yet answer. The first is whether AI’s economic payoff arrives roughly on the promised schedule. The capital being committed assumes a steep productivity curve, yet a recent NBER study found a large majority of firms report little measurable impact on output so far. The second is the question systemic-risk analysts are circling. When the financing unwinds, does the damage stay largely contained inside the technology sector, creating a sharp but bounded repricing, as in 2000? Or does leverage, private credit and rehypothecated collateral create a wider crash, as in 2008?
Cross the two questions and four distinct futures appear. In The Productive Shakeout, the froth burns off and over-leveraged players fail, but the technology and its strongest operators endure, with the wider economy collecting the dividend. The Long Hangover is quieter and slower: no Lehman moment but the gains don’t arrive on schedule, valuations deflate over years, data centers are left half-empty and a tech recession lands. Governments might decide AI should be managed like critical infrastructure, running data centers like public utilities.
The third scenario is Crisis, Then Cornucopia. The financial scaffolding cracks, credit freezes and a real recession follows – yielding, in the long run, to societally beneficial results. The Reckoning is its nightmare twin: no productivity gains and full systemic contagion, like an AI-flavored 2008 with nothing to cushion the fall. Emergency responses from policymakers would be required – or societal disaster would unfold.
In all scenarios, the big question is whether we emerge with an AI deployment boom. This is Carlota Perez’s turning point in textbook form: a crash marking the transition between the installation frenzy and the golden age of deployment. The point is not to pick the scenario you believe in, but to find moves that pay off across several. Three things can be done now.
Separate the two bets The technology and financing questions are distinct. You can be convinced AI is durable and still refuse to wager your balance sheet on circular financing holding together.
Watch the edges, not the headlines Inflection points announce themselves at the fringes first. AI’s leading indicators are already visible. These numbers will move long before problems become obvious.
Stress-test your own place in the loop Are your unit economics quietly subsidized by below-cost compute that will one day reprice? Have you signed take-or-pay commitments that turn into anchors the moment demand softens? Preserve optionality: in any scenario, it’s the firms that survive that get to buy the upside.
Seeing around this corner does not mean predicting which future will arrive. It means being solvent, and ready, for all four.
Rita Gunther McGrath is professor of management at Columbia Business School and a Duke CE educator
