Eighty-Two, Then Fifty-Three

In July, Bank of America's global fund manager survey recorded 82 percent of respondents naming long global semiconductors as the most crowded trade. It was the highest reading the survey has produced for any position.

In August the same question returned 53 percent.

Between those two surveys, nothing material changed in the industry the trade was about. Data centre construction did not stop, chip demand did not reverse, and no major capital expenditure programme was cancelled. What changed was who was holding the position, and the mechanism of that change is worth more attention than the collapse that made the headlines.

The most concentrated expression of the trade

Situational Awareness, a fund run by Leopold Aschenbrenner and named after the essay that did more than any other document to popularise the AI scaling thesis, had grown to roughly 45 billion dollars in assets. Its strategy, as described in contemporaneous reporting, was to own the suppliers of the build-out, chips, data centres and power, while shorting software companies expected to be disrupted by the same technology. Filings as of 31 March showed positions including Nebius, Bloom Energy, SanDisk, CoreWeave, SharonAI and IREN. Leverage was reported at up to 400 percent.

During July the fund lost roughly two thirds of its value, assets fell to about 10 billion dollars, and it was obliged to sell its leveraged public equity positions to Citadel at a discount. It has since reportedly returned with around 400 million dollars.

What is striking is how little of this requires the thesis to have been wrong. On the available evidence it was not. The fund was not carried out by a change in the fundamentals it had underwritten. It was carried out by the volume of capital expressing the same view with the same instruments at the same time, amplified by borrowing, in a month when the survey said the position was more crowded than any position the survey had previously measured.

Crowding is a risk factor and it is not priced like one

Standard risk frameworks treat concentration within a portfolio carefully and concentration across the market barely at all. A fund can measure its own exposure precisely while having no visibility into how many others hold the same exposure, which is exactly the variable that determines the price at which it can exit.

The BofA survey is one of the few regular instruments that measures it, and it is a survey of opinion rather than of positioning, which is a real limitation. But the two-month swing suggests something the level alone does not: that a substantial share of the July reading was not conviction. A position genuinely held on analysis does not evaporate in four weeks absent new information. Momentum does.

That distinction matters for anyone assessing systemic risk in this cycle. The question is not whether AI infrastructure spending is justified by eventual returns, which is unresolved and will stay unresolved for years. The question is how much of the capital currently expressing that view would remain if the price stopped rising, and the honest answer from the July-to-August move is: materially less than the headline number implied.

The contagion channel that did not fire, and why

A 45 billion dollar fund at reported leverage of up to 400 percent unwinding inside a month is the sort of event that historically transmits. This one largely did not, in the sense that the broad market absorbed it without a liquidity event.

Two features explain most of that. The positions were in liquid, heavily traded names rather than in anything structured or private, so there was a market to sell into. And a single well-capitalised counterparty, Citadel, took the book at a discount, which converted what could have been a forced sale across many venues into a negotiated transfer. The discount is the price of that orderliness and it was paid by the fund's investors.

Neither of those conditions is guaranteed next time. The lesson generally drawn from a collapse that was absorbed is that the system is resilient. The more careful reading is that this particular collapse had two specific shock absorbers, and that a similarly crowded position expressed in less liquid instruments, or arriving when no buyer of that size wanted the book, would not have had them.

What it does not license

It is not evidence of an AI bubble, and the reporting does not support that claim. A crowded trade and a mistaken trade are distinct, and the events of July speak to the first. The build-out those companies serve is proceeding.

It is also not evidence about the individuals involved, beyond the obvious. A great deal of the coverage has focused on the manager's age and on his having come to the fund from AI research rather than from a trading desk. That is true and reported, and it is not much of an explanation: crowded, leveraged positions in correct theses have been unwound violently by people with thirty years of experience for as long as there have been markets.

The durable observation is duller and more portable. Being right about a direction says nothing about survivability of the path, the path is determined by who else is positioned alongside you, and that variable is measured seriously by almost nobody. Eighty-two percent, then fifty-three, in a month when the facts did not move.

Sources

• Bank of America Global Fund Manager Survey, reported by 24/7 Wall St. and Benzinga. Source for long global semiconductors at a record 82 percent in July 2026 and 53 percent in August.
• CNBC, 30 July 2026 and 31 July 2026. Source for the forced unwind of leveraged public positions, the sale to Citadel at a discount, and the long-infrastructure, short-software strategy.
• Bloomberg, "Situational Awareness Assets Drop to $10 Billion After Liquidations", 30 July 2026. Source for the decline from roughly 45 billion dollars to about 10 billion.
• The Next Web, "Leopold Aschenbrenner's AI fund lost most of its value. Now it's back with $400M." Source for the fund's return at roughly 400 million dollars.
• Named positions as of 31 March and the leverage figure of up to 400 percent reach us through the coverage cited rather than from filings read directly, and are reported as such. The Economist's issue of 8 to 14 August 2026 carries a first-half 2026 return figure for the fund that we could not corroborate in contemporaneous reporting; it is not used here.