The AI Trade Just Broke. The New Test Is Revenue, Power, and Cash Flow.

Last week, Goldman Sachs’ prime brokerage book saw the largest technology-stock selloff on record. Rich Privorotsky, head of European FICC and equities trading at Goldman Sachs, gave the number in an interview from the bank’s London trading floor. He did not call it a correction or a pullback. He said the crowded “long AI, short everything else” trade was being painfully deconstructed.

This was more than profit-taking. For two years, the AI trade worked like a faith: as long as hyperscalers kept raising capex and compute demand kept expanding, the AI supply chain could absorb any uncertainty. That machine is now loosening. The old consensus is breaking up. The new one is taking shape: AI is no longer a story contest. It is an exam on revenue conversion, power delivery, financing capacity, and free cash flow.

$1.1 trillion is no longer a free pass

The market expects AI capex to reach $1.1 trillion next year. Two years ago, that number would have lifted almost any AI asset. Not now.

Privorotsky’s words: “It is not enough to simply say you are spending a lot of money. You also need to prove that backlog is converting into substantive revenue.” Hyperscalers must keep pushing capex higher and show the return. Cloud growth, monetization, and the speed at which free cash flow reaches breakeven will feed straight into a credit market that is already sensitive.

Capex has changed meaning. It used to be a positive signal: investment, commitment, future. Now it is a question. When will the money come back? Who finances it? If revenue conversion disappoints, who carries the credit risk? $1.1 trillion is no longer the answer. It is the opening question.

Power and financing: the physical limits of AI mania

AI is not just software. It ends up on chips, in data centers, and at the power grid. Privorotsky named a specific bottleneck: next year’s power gap is roughly 40 to 50 gigawatts. His question was direct. If you cannot secure the energy that actually powers data centers, can you still use the chips you ordered in advance?

That is the physical limit of AI capex. The market can trade expectations. Data centers cannot run on expectations. Chips can be ordered early. Without power delivery, compute does not form, and revenue does not arrive. Investors have moved into de-risking mode: short AI and AI capex trades to lower growth expectations, and put money into bonds. Privorotsky said that as growth expectations came down, bonds rallied. That move was “completely not in anyone’s script.”

Financing is tightening too. Privorotsky pointed to the more leveraged, most advanced parts of the AI trade, especially emerging cloud providers. They face a financing test: can they raise money to deliver on AI spending? If capital costs rise and credit spreads widen, marginal borrowers get squeezed out first. The AI label used to blur credit differences. Rising capital costs expose them again. Stable cash flow, cheap financing, real customer orders, and leverage-supported narrative will now be priced separately.

From savings glut to spending glut

Privorotsky offered a larger macro frame. For most of his career, the system ran a savings glut: too much money chasing too few uses. That has changed in what he called a seismic shift.

Over the past year, the market expects more than $1 trillion in AI capex, while most governments run fiscal deficits. We are in a spending glut. Capital is occupied. Government deficits compete with private capex for funding. Capital costs rise. Marginal borrowers get pushed out.

The pressure shows up in CCC credit spreads, which have widened sharply, and in sovereign debt, including French government bonds with a large fiscal deficit. Energy prices remain high. Risk appetite is weaker. Privorotsky noted that the S&P 500 has looked stable while the median stock is down more than 10%. Financial conditions have tightened sharply and forced the Fed to do more. The AI trade is no longer a tech-sector story. It sits inside macro rates, credit spreads, energy prices, and financial conditions.

Afraid of under-delivery, afraid of over-delivery

Market sentiment is split in a strange way. Privorotsky described two investor conversations this week. One group said, “I’m afraid we won’t be able to deliver compute, and that scares me.” Another said, “I’m afraid we will over-deliver compute, and that scares me.” Two opposite outcomes now make up the market’s anxiety.

That sentence explains the current AI trade. The market is not simply bearish on AI. It has lost a single view on the pace and return of AI capex. Fear of under-delivery means supply bottlenecks, power gaps, and financing constraints could delay revenue. Fear of over-delivery means capex is overheating, capacity will overshoot, and returns will fall. Both fears exist at once. The market has moved from directional faith to pricing uncertainty.

That is why “long AI, short everything else” is breaking. The old trade assumed AI was the only winner and other assets would be drained. The new trade has to accept three things: AI will diverge internally, AI and macro will interact, and AI capex will feed back into credit and rates. Privorotsky said he used to believe “as goes oil, so goes everything else.” The logic has shifted to “as goes AI, so goes everything else.” The line is a slight exaggeration. It still captures a fact: AI now drives macro risk appetite and financial conditions.

The new consensus: four shifts

The old consensus breaking does not leave a blank. Four shifts are taking shape.

First, the trade moves from burning-cash narrative to revenue verification. Capex will stay high, but the market will focus on cloud growth, order conversion, monetization, and free cash flow breakeven. Hyperscalers have to prove that $1.1 trillion is capital that earns a return, not a bottomless pit.

Second, the trade moves from single long-AI positioning to credit differentiation. Rising capital costs squeeze marginal borrowers. More leveraged AI trades, especially emerging cloud providers, face financing tests. Credit markets become the referee. CCC spreads, sovereign debt pressure, and financial conditions all feed into AI asset valuations.

Third, the trade moves from infinite demand to supply constraints. Power gaps, data center delivery, and chip digestion become hard limits on AI capex. Record compute spot prices show strong demand. If power arrives late, revenue arrives late.

Fourth, the trade moves from concentrated bets to a barbell. Privorotsky said conviction is not high, but heavily de-rated electrification assets, civil aerospace assets, and parts of transport offer a margin of safety and deserve cautious optimism. Outside AI, the market is looking for assets that have been oversold, de-rated, lightly positioned, and offer convexity.

Not bearish on AI, but repricing it

Privorotsky is not fully pessimistic. He said spot compute prices hit new highs every day and demand remains “insatiable and endless.” AI’s real-economy demand is still strong. What changed is how the capital market prices AI.

The market used to pay an unlimited premium for future AI. It now asks current AI to prove it can make money. Capex used to be proof of faith. It is now an exam on monetization. The AI trade used to be a directional bet. It is now a set of equations across revenue, power, financing, credit, and cash flow.

For investors, the AI trade will not disappear. It will get harder. The crowded trade breaking brings volatility and reallocation opportunities. For hyperscalers, earnings season now turns on monetization quality, not capex size. For highly leveraged AI companies, the financing window and credit spreads decide survival. For the macro market, AI capex is tied to rates, energy, credit, and financial conditions.

Closing

The consensus trade broke because the market stopped paying an unlimited premium for one narrative. The new consensus reprices AI around arithmetic and monetization. It sorts credit instead of treating AI as one trade, and it favors a barbell over a one-way bet.

Privorotsky’s interview leaves one useful line: the market is now afraid of both “not being able to deliver” and “over-delivering.” Those opposite fears show that the AI trade has entered a more mature and more painful phase. AI remains a large capex cycle. The capital market no longer pays for a story alone. It wants revenue, cash flow, power, financing, and returns.

AI is not done as a narrative. It is being asked to hand in its homework.