The number that caught attention on the Goldman Sachs trading desk was not a price. It was a z-score, below minus three for global AI stocks, measured across six consecutive sessions of net selling, a reading the bank’s prime brokerage describes as the most extreme reduction in positioning since ChatGPT was released in late 2022.
The returns explain the anxiety. Fundamental long-short funds focused on Asia are down 7 percent this month, according to the same prime brokerage data, and their gain for the year has fallen from roughly 40 percent at the June peak to 16 percent. Four months have undone more than half a year of performance, and the bulk of the loss arrived in the past few weeks.
Mark Wilson, a partner at the bank, described the shift in the vocabulary of a pendulum. Sentiment, pricing and positioning momentum have all swung sharply negative, he said, and the swing has very likely carried too far in the other direction.
That phrasing does real work for a desk that wants to discuss a rebound without promising one. Wilson attached a condition that undercuts the optimism: until the junk-rated assets at the speculative end of the AI trade stop falling, any rally in AI shares reflects repaired positioning rather than a fresh judgment about earnings.
He named CoreWeave as the reference point. The neocloud’s stock fell 15 percent in a single session last week, its five-year credit default swaps reached a record, and traders began marking the July low as the level that would decide whether the selloff had found a floor.
The composition of the trade explains why a memory price in Seoul can move a fund in Hong Kong. The crowded version of the AI bet was never a single stock. It was a package: the accelerator maker, the three memory suppliers stacked behind it, the power and cooling contractors, and the rented-compute companies that convert contracts into debt.
Asia’s long-short books were built heavily around that supply chain. A manager who wanted exposure to AI computing without paying a premium for American software multiples could buy the memory complex and short the laggards. When memory pricing and hyperscaler spending plans drive both legs, the book behaves like one position held in several accounts.
De-grossing then does the rest. Rising volatility prompts prime brokers to lift margin requirements, and funds cut gross exposure before they cut conviction. Correlations converge toward one, winners and losers are sold in the same week, and a portfolio designed as a set of independent bets is revealed as a single bet.
The fundamental case has not collapsed. Memory contract prices are still rising, advanced packaging capacity remains sold out, and the largest cloud operators have not withdrawn the capital spending plans they published this year. Analysts who cover the sector note that nothing in the order books deteriorated during the six days of selling.
What deteriorated is the price of financing those order books. Credit spreads on the hyperscalers reached their widest level of the year, according to Goldman, and the high-yield market has grown selective about lending to technology companies without operating cash flow. Equity investors are now discounting a cost of capital that the bond market set first.
Wilson’s sequencing argument follows from that observation. Credit leads equity, because lenders reprice before shareholders do, and the weakest borrower in a complex reprices first. Junk-rated AI assets are therefore the place to watch for a bottom, not the shares of companies with investment-grade balance sheets and cash flow from other businesses.
For allocators the arithmetic is less philosophical. Most Asia-focused funds offer monthly or quarterly liquidity, which means an investor who asks for money in October is paid from sales made in October and November. Redemptions force selling into weakness regardless of what the manager believes about 2027.
The year is not a loss. A 16 percent gain would rank well against most of the past decade for this strategy, and funds that entered June with less concentration have kept more of it. The shape of the year has changed, though, and the change is instructive. Returns that come from being early to a theme can leave faster than the thesis takes to play out.
There is also a mechanical reason the damage clustered. Position data from prime brokers is published with a lag, so by the time a crowded trade is visible in the numbers, the crowding has already been paid for. The z-score measures an exit, not an entry.
What remains unresolved is whether the June peak marked the top of a cycle in positioning or the top of a cycle in earnings. The first is a trading problem that heals with time and lower gross exposure. The second would require memory prices and capital spending plans to turn, and neither has turned yet.
