Unpacking the Two-Week AI Sell-Off

Investors are expressing apprehensions over the high degree of borrowing and leverage to build out data center capacity ahead of an uncertain ROI for AI. Ongoing negative FCF may have no end in sight as capital spending plans from the hyperscalers, neocloud vendors and others have, in aggregate, increased every quarter for the past 3–4 years. Inference demand continues to outstrip supply, so demand uncertainty is not so much the issue here, in our view, at least over the next few years (industry sources).

Two weeks, two different moves

From October 9th to the close on Friday, October 23rd, our Datacenter Enablers segment fell 18.2% and Silicon & Semis 11.9% (Exhibit 1). For Datacenter Enablers it was the worst ten-session stretch in our data, which begin in December 2022. The Hyperscalers fell 5.9%. Software stocks followed a different pattern. Cybersecurity and Data & Dev Infra, which had both closed at records on October 9th, fell 14.2% and 12.7% in the first week, then rose 10.3% and 7.6% in the second, ending the two weeks down 5.5% and 6.1%. Applications, the only one of our six segments down for the year going in, slipped 0.8% in the first week and rose 5.4% in the second, gaining 4.6% over the two-week period.

The sell-off was concentrated in the first week, October 9th to 16th. Negative catalysts included September consumer prices, released before the open on October 14th, which put headline inflation at 3.7% (BLS) and raised the prospect of a second Federal Reserve rate increase (Reuters) in December. Larry Fink from BlackRock was quoted on CNBC noting that he was not sure the market could continue to accommodate all the supply of debt capital required in the market.

On October 14th shares of debt-heavy AI companies such as Nebius and CoreWeave tumbled, and the cost of insuring the debt of some AI companies against default (credit default swaps), including Oracle and CoreWeave, rose to records (Reuters). On October 15th, CoreWeave’s credit default swaps implied a nearly 63% chance of default within five years, and swaps on Amazon, Alphabet, Meta and Nvidia, though at lower levels, were also at records (Reuters). Also on October 15th, Reuters reported that some Anthropic investors were weighing a lower IPO valuation after its latest financial disclosures implied slower sequential revenue growth (public filing expected this week).

The second week, October 16th to 23rd, opened with a rebound. On October 19th AI stocks partially recovered from the previous week’s selloff; Goldman Sachs upgraded CoreWeave, saying the selloff offered a buying opportunity, and Morgan Stanley named Amazon its top pick (Reuters). On October 21st the Nasdaq led a higher close after Super Micro, late on October 20th, put preliminary quarterly revenue at $18.5–19.5 billion, against guidance of $14.5–15.5 billion (Super Micro), and SpaceX reported its first profit as a public company. Reuters said the reports “appeared to provide some reassurance to investors who have been worried about the sustainability of massive spending on AI buildouts and their profitability” (Reuters). Optimism was, however, further dampened on Thursday, October 22nd, with Silicon & Semis falling 5.4%, its worst day since June 23rd, and Hyperscalers falling 3.7%, after results from Meta and Lam Research, discussed below, while Applications rose 1.2%. On October 23rd big semiconductor stocks were mostly higher (Reuters), and Cybersecurity rose 5.6%, led by CrowdStrike (+8.2%) and Palo Alto Networks (+7.2%) (Exhibit 2).

Cumulative total return from the 10/9 close; S&P 500 dashed. Source: FactSet, Sherlund Partners.

October 9th to 16th: a crowded trade unwinding

We ranked the stocks in our universe by year-to-date return as of October 9th and split them into quartiles. In the first week the top quartile, whose median stock had gained 129% this year, fell a median 17.2%. The bottom quartile, down a median 26% for the year, rose a median 1.4% (Exhibit 3). The pattern held within each of the five layers with enough stocks to test: in each one, the half of stocks with the bigger gains this year was worse than the other half. Because the year’s biggest winners did worse in software as well as in hardware, this was an unwind of crowded positions across the AI stack, not simply a rotation out of AI hardware into software.

Two other patterns show up within layers in the first week. The stocks with the most short interest going in fell more than others in their layer, and the stocks with the highest free cash flow relative to their market value held up better.

Prime-brokerage data point the same way. Goldman Sachs’ data showed hedge funds cutting technology stocks at the fastest rate in at least a decade in the five sessions to October 15th, and their net leverage fell to a one-year low of 48.6%; even so, net allocation to technology was still 23.5% of Goldman’s global prime book, in the 90th percentile of the past five years (Investing.com).

Goldman, commenting on aggregate hedge fund positioning, indicated that long-short managers had sold global AI stocks for six straight sessions. Bruno Schneller of Erlen Capital Management called it “less a fundamental rejection of AI and more a classic, high-velocity unwind of a crowded trade meeting a macro reality check” (Reuters). Fund investors bought as hedge funds sold: global technology funds took in $5.05 billion in the week to October 21st, reversing $4.77 billion of net sales the week before, according to LSEG Lipper data (Reuters).

October 16th to 23rd: laggards and shorted stocks led software’s rebound

In the second week the pattern changed. The bottom quartile rose a median 10.1%, while the top quartile was roughly flat, up a median 0.7% (Exhibit 3). The stocks with the most short interest, which had fallen more than their peers in the first week, rose more than others in their layer in the second. In application software, the third of the segment with the highest short interest rose a median 12.0% over the week, against 6.0% for the third with the lowest short interest.

The 129 stocks in our universe with a year-to-date return, grouped into quartiles by year-to-date return at the 10/9 close. Source: FactSet, Sherlund Partners.

Short positions had been building. It was reported on October 22nd that short sales had outpaced long purchases across global equities by about five to one in October and that short leverage in Goldman’s prime-brokerage book was at the 99th percentile of its five-year history, raising the risk of a short squeeze (IBTimes). On October 22nd all three hedge fund strategies Goldman tracks lost more than 1%, the first time since the pandemic, and Goldman said performance was weighed down in part by “a squeeze in crowded short positions” (Reuters). FINRA short-interest data, which will show whether short positions fell during these two weeks, will be published on October 26th (for the October 15th settlement date) and November 10th (for October 30th) (FINRA).

Takeovers added to the gains in application software. Salesforce agreed on October 20th to buy Braze for $47.50 a share in cash (Salesforce), a premium of about 58% (Reuters), and Braze rose 58% over the two weeks. In the first week Proofpoint agreed to buy Varonis in a deal valued at about $6.3 billion (Reuters), a GTCR-led consortium including Elliott agreed to buy CCC Intelligent Solutions (CCC), and Intuit agreed to buy Sequence (Intuit).

AI hardware: financing concerns and supply constraints

AI hardware did not take part in the second week’s rebound. Investors grew more concerned about how the build-out will be paid for. Among the companies building AI data center capacity, the three with the most negative free cash flow relative to revenue, Nebius, CoreWeave and Oracle, fell furthest (Exhibit 4). CoreWeave fell 34% and Nebius 28%; in Datacenter Enablers, the four stocks with negative consensus free cash flow for 2026 fell a median 30.6%, against 20.8% for the other ten. CoreWeave has about $50 billion of debt, and JPMorgan expects its free cash flow to stay negative until 2030 (Reuters). Some of CoreWeave’s corporate bonds have slid to around 80 cents on the dollar, pushing yields to roughly 15% (Bloomberg), and Reuters wrote on October 23rd that the recent performance of neocloud companies such as CoreWeave had “raised talk on Wall Street that capital is growing scarce” (Reuters).

Meta fell 7.0% on October 22nd after its third-quarter results. Revenue beat estimates and the fourth-quarter revenue forecast was above estimates (Reuters), but Reuters reported that investors “balked at another increase in its annual AI spending forecast and a plan to sell new stock to help pay for it” (Reuters). Meta narrowed its 2026 capital spending forecast to $135–145 billion from $130–145 billion, raising the low end, and reported free cash flow of negative $2.12 billion (Meta), which Reuters called “the first cash burn at Meta in more than a decade” (Reuters). Meta created a new class of publicly listed, non-voting shares, expected to begin trading in late January 2027, which its chief financial officer said could be used for capital raising, employee compensation and acquisitions. She also said that because its Muse AI product is compute-intensive, expanding its user base will require significant investment and total serving costs will rise. Meta is not providing a full-year 2027 revenue outlook (Meta earnings call).

Lam Research, which makes chip manufacturing equipment, fell 11.8% on October 22nd, even though Lam described demand as strong; the constraint it cited is supply. Lam said supply chain constraints held back its revenue guidance and that limits on clean room space and components will keep chipmakers and equipment makers from fully meeting demand in 2027 (Lam earnings call). Its quarterly profit and revenue were slightly ahead of analysts’ expectations, though the midpoint of its profit forecast for the current quarter was slightly below (AP), and it was reported that chip stocks fell amid concern that such limits could hurt chip sales (Yahoo Finance).

Within our Hyperscalers segment, Microsoft, the only one with positive consensus free cash flow for 2026, fell least (2.0%), and Oracle, which has the most negative free cash flow and carries more debt relative to its size than the others, fell most (10.8%). Amazon, Alphabet and Meta, all expected to have negative free cash flow this year, fell 5.5%, 8.9% and 10.6%.

Valuations fell; earnings estimates have barely moved

Consensus calendar-2027 earnings estimates changed by less than 1% in every segment over the two weeks: up 0.7% for Silicon & Semis and down 0.2% for Applications. Forward price-to-earnings multiples fell in every segment except Applications between October 9th and 23rd (Exhibit 5). Estimates often lag prices, but so far the selling has come through valuations, not earnings forecasts.

On October 23rd our Silicon & Semis segment traded at 12.7 times forward earnings, its lowest multiple in our data, down from 20.7 times at the end of June. The low multiple partly reflects how far earnings forecasts have risen: as of October 22nd, consensus expected the segment’s earnings to grow about 71% over the next 12 months. Cybersecurity, even after its decline, traded at a higher multiple on October 23rd than on 97% of trading days in our data.

Next week’s catalysts

Microsoft and Alphabet report third-quarter results on October 28th and Amazon on October 29th; investors will be watching for 2027 capital spending plans and how they will be funded. Consensus expects Alphabet’s free cash flow to fall from about −$6 billion this year to about −$34 billion in 2027 and Microsoft’s from about $46 billion to about $33 billion, while Amazon’s stays negative at about −$24 billion (Exhibit 4). The Federal Reserve announces its rate decision on October 28th. Late on October 22nd, fed funds futures implied only a 5% chance of an increase at that meeting but about an 80% chance of one in December, Reuters reported (Reuters).

We welcome your thoughts and a follow-up conversation.