Slapped in the face from both sides! "Insufficient power" vs. "excess computing power" either way, the result is "selling AI"

US hyperscalers have just delivered their sixth consecutive quarter of accelerating growth, while Intel suffered its worst single-day plunge in half a century (-29%). Caught in the middle, hedge funds spent all of October selling anything with "AI" in the name. What exactly is the market afraid of? Rich Privorotsky, head of EMEA Delta One trading at Goldman Sachs, said in his Monday morning weekly note that no one really knows what to worry about. He wrote: The great AI debate is becoming somewhat circular. Half the market believes we simply cannot build enough power and data centers to support the 2027 capex plans, so growth in "picks and shovels" stocks must peak and valuation multiples will compress. The other half believes we are about to massively overproduce compute, which will destroy rental economics and likewise demand lower valuation multiples. I think the bigger issue is that no one really knows what to worry about. Strong demand is undeniable, but the return on the next trillion dollars of investment is far from certain. Insufficient compute, or excess compute? Privorotsky's observation highlights the factual mutual exclusivity of the two bear-case logics. If the bottleneck is power, then compute will remain scarce, rental prices will stay firm, the problem is not enough power to connect the next GPU, and the growth curve for "picks and shovels" stocks flattens. Morgan Stanley analyst Stephen Byrd estimates the US net power shortfall will still reach 38GW by 2028. If the industry is about to massively overproduce compute, then rental economics will collapse and the returns on existing chip inventories will disintegrate. Privorotsky warned of this risk as early as June this year, when he noted that the market's core premise is built on compute scarcity. These two scenarios cannot both be the true core constraint at the same time, but both can be reasons to sell. Privorotsky wrote: "When positioning is crowded, every new worry becomes another reason to reduce risk exposure." The market does not need to pick one bear-case logic; it only needs to find the most popular explanation of the day. Demand is real, and so are the bills Last week's hyperscaler earnings were stellar: Azure grew 48% for the quarter, AWS grew 39%, and Google Cloud grew 103%. Privorotsky called it "extraordinary demand." Combined cloud revenue growth for Amazon, Alphabet and Microsoft reached 54% year-on-year in Q3, accelerating for a sixth consecutive quarter. But then comes the "that said." Privorotsky acknowledged that "negative free cash flow, equity financing and growing leverage requirements have raised legitimate questions about returns on capital." His Goldman colleague Lee Coppersmith provided the specific numbers: hyperscaler capex is expected to reach US$1.2 trillion in 2027, while combined free cash flow already turned negative in Q3, burning about US$31 billion in a single quarter. Simple accounting logic: US$1.2 trillion of capex plus negative free cash flow means the "next trillion dollars" that Privorotsky worries about will be financed mainly through the bond market rather than operating cash flow. And the bond market itself is not comfortable either. AI-related off-balance-sheet debt has swelled to US$4.1 trillion, and financing demand is becoming increasingly crowded. Intel plunges 29% Last Friday was a typical example. Earlier in the week, the market was still digesting a series of relatively positive news, but Intel then delivered an "unexpectedly severe PC demand correction," sending its shares down 29% in a single day. Privorotsky believes Intel's warning is "more about consumer PCs, margins and the manufacturing business than about AI demand itself." But the market was clearly less patient. Investors still traded it as another victim of the AI narrative. Meanwhile, Amazon which Privorotsky called a "crowded position" narrowed its premarket decline of 7% to close down just 1.4%. Even though concerns about its free cash flow have not disappeared, the stock showed some resilience. Selective pricing is beginning to emerge in the market. However, Kioxia's 15% plunge on Monday also reminded investors that this is not an indiscriminate broad rally. Hedge funds see largest deleveraging since 2015 Is it deteriorating AI fundamentals, or forced liquidation of crowded positions? Privorotsky's positioning data leans toward the latter. Goldman-tracked systematic long/short strategies fell 2.17% for the week, "the worst weekly performance since December 2023." Fundamental long/short net exposure recovered only to the seventh percentile historically, "achieved mainly through index short covering." Single stocks saw net selling for a seventh consecutive week, while macro products saw net buying for a second straight week. The AAII bull-bear sentiment gauge returned to negative territory, and CNN's Fear & Greed Index fell to just above 20. Morgan Stanley also found that hedge funds' AI-related positioning had fallen to an underweight for the first time since early 2023. This followed one of the largest sustained deleveraging episodes in more than 15 years. Privorotsky's conclusion: "Sentiment and positioning have been reset." But his Goldman colleague Coppersmith is more cautious. He noted that global risk appetite in October saw its largest monthly decline since November 2021, but "we are still not at the risk-appetite index level that would trigger a contrarian buy signal." More importantly, although net exposure is already low, gross leverage remains as high as 314% (97th percentile of the past five years). In other words, fund managers are cutting positions, but overall leverage remains very high, and the market still has pent-up energy for further volatility. "Almost unmanageable," but a year-end rally may already be under way Privorotsky had an unusually candid passage in the report: It feels almost unmanageable. Almost every client I have spoken with is struggling. Last week felt like broad-based deleveraging... momentum factor unwinds, the most-shorted stocks squeezed, leverage withdrawn from the system. Defensive stocks and high-quality assets surged, AI valuation multiples compressed. The market is effectively pricing an abrupt growth slowdown, a stall in AI capex, lower nominal growth and rates, and a premium for high-quality balance sheets. He then noted that "there was nothing in last week's hyperscaler earnings to suggest any weakening in willingness to spend... if anything, the opposite." Depending on which bear camp you belong to, that is either a reason to be long or precisely the problem. After key events including the midterm elections, the US Treasury's quarterly refunding announcement (QRA) and nonfarm payrolls data, Privorotsky gave his judgment: My gut tells me we are setting up for a painful year-end rally... underpositioned, low participation, hard to chase. Friday did feel like capitulation, but I think we need to get through these events first before risk is added back. UBS also warned of year-end short-squeeze risk, with short positions near record levels. Goldman derivatives strategist Brian Garrett also sees signs of "healing" in seven markets but added that "we just need to get through 2026 first." The conditions for a positioning squeeze do exist: net exposure at the seventh historical percentile, sentiment at rock bottom, and seasonal tailwinds approaching. But as Privorotsky himself asked: demand is real, but the return on the next trillion dollars is uncertain. With hyperscaler free cash flow already negative, gross leverage at the 97th historical percentile, and the 2027 capex bill about to land on the bond market at scale, any year-end rally could be a selling window. This article is reprinted from the "Zhuifeng Trading Desk" WeChat public account. GMTEight editor: Song Zhiying.