The $3.2tn AI sell-off was a leverage event before it was a verdict on AI

US equities lost approximately $3.2tn on Thursday as the Nasdaq 100 fell 6.4% and the S&P 500 dropped 4.4%. A 1.4% Nasdaq rebound on Friday still left the index down 4.9% for the week. Chipmakers, data-centre companies and other concentrated AI exposures led the decline after Democratic midterm victories raised the probability of slower permitting, fewer subsidies and tighter environmental and utility rules.

The political news mattered because investors had treated infrastructure construction as close to an exogenous fact. The industry expects roughly $1tn of AI investment next year. New Democratic governors in states responsible for large shares of planned capacity had campaigned on pauses or moratoriums, while local ballot measures repeatedly rejected data-centre expansion. Even when a pause is temporary—as Texas governor-elect Gina Hinojosa clarified hers would be—the possibility of delay changes the present value of projects whose debt, equipment orders and customer commitments assume a particular energisation date.

But politics alone does not explain the speed. The FT reports that hedge funds and retail investors were already heavily concentrated in the same AI trade after a year in which it dominated equity returns. JPMorgan estimates that hedge funds reduced borrowing over a 20-day period at the fastest pace since the Covid crash. Falling prices produced losses and margin pressure; leveraged funds sold to reduce exposure; those sales pushed prices lower and forced further deleveraging. The election was the catalyst for a position unwind that had been prepared by crowded ownership, higher borrowing costs and deteriorating AI credit.

Friday’s rebound was consequently weak relative to the fall. The S&P recovered less than 10% of Thursday’s loss, compared with a historical average of roughly 46% after declines of that size. Some investors argued that the move exceeded the change in fundamentals: much of the 2027 construction pipeline is already permitted, and stronger October employment data did not support an immediate recession narrative. Others saw the rout as the first time political limits, infrastructure debt and uncertain returns were being priced together.

The useful distinction is between a technology judgment and a financing judgment. The sell-off does not show that AI demand disappeared or that models stopped improving. It shows that valuations had bundled those beliefs with a stronger and less defensible assumption: that capital, permits, power and public consent would remain available on schedule. When that assumption moved, leverage determined the size of the first-day loss. The next test is slower. If projects are cancelled and credit remains expensive, the market has identified a durable constraint. If permits continue and forced sellers finish unwinding, much of the rout will have been a violent repricing of ownership rather than of AI’s underlying economics.