Hedge Funds Wipe Out Two Years of AI Positions in a Week as TSMC and Jensen Huang Respond to Compute-Demand Doubts

This wave of concentrated selling in global tech stocks wiped out hedge funds' net buying in AI over the past two years within five trading sessions, leaving whether deleveraging will continue as the market's biggest question; at the same time, TSMC's expansion plan with three-year capex far exceeding history, Nvidia CEO Jensen Huang's public breakdown of compute demand structure, and BlackRock's record $15.49 trillion in assets under management offered different answers from fundamentals and flows. Asian markets opened higher on October 16, with innovative drugs and defense strong, showing that capital is looking for new footholds outside tech.

Goldman Sachs Prime Book: Five Days Erase Two Years of AI Net Buying

According to a latest report by Vincent Lin, head of Goldman Sachs Prime Brokerage, hedge funds net sold global AI stocks for five consecutive trading sessions from October 8 to 14, completely reversing all cumulative net buying since the start of 2025—every net purchase over the past two years was zeroed out within a week. The Goldman Sachs Prime Book shows global equities saw large net selling, with single-day net selling reaching -2.1 standard deviations on a one-year lookback basis; North America and emerging Asia led the decline, with South Korea and Taiwan the most concentrated markets for net selling. The shock has already hit specific markets: South Korea's KOSPI fell into a bear market, Seoul triggered a circuit breaker, Taiwan saw a record single-day foreign net outflow of about $9.3 billion, and central banks in several Asian countries were forced to intervene in currency markets. TMT directional long/short funds lost 2.7% in a single day, an extreme move of -2.6 standard deviations, with October losses widening to 4.5%.^[1]

In terms of trading structure, this selloff was mainly driven by active long-position reduction. Goldman Sachs data shows that the day's net selling in global AI stocks was the second-largest in the past year, behind only the crash in July this year, with the ratio of long selling to short selling at about 3.5 to 1. Just shortly before, Goldman Sachs noted that hedge funds' net exposure to the Mag7 in its Prime Book accounted for nearly 22% of total U.S. exposure, a historical high; five days later, the same group of funds rushed for the exits en masse. Rich Privorotsky of Goldman Sachs' Delta-One business summed it up: the pain trade used to be long AI, short bonds; now both legs are being closed simultaneously, and the market is discovering that compute demand can grow exponentially, but the capital that can finance it cannot.^[1]

Positions Remain Heavy, Deleveraging Room Persists

Despite the extremely heavy selling this week, AI positioning remains elevated from a longer-term perspective. Goldman Sachs data shows that net allocation to global AI stocks as a share of global single-stock net market value is currently 26.1%, down from a peak of 31.7% in early June, in the 49th percentile over the past year, but still in the 83rd percentile over the past three years; the long/short ratio of AI portfolios has fallen from a June peak of 4.34 to 3.27, in the 7th percentile on a one-year basis and still in the 64th percentile on a three-year basis. Leverage data better explains why the decline was so violent: overall gross leverage in the Goldman Sachs Prime Book fell only slightly by 0.4 percentage points to 310.3%, in the 72nd, 91st, and 94th percentiles on one-, three-, and five-year bases respectively; net leverage fell 1.5 percentage points to 73.3%, down 4.3 percentage points in October, in the 6th percentile over the past year. JPMorgan's market intelligence TPM positioning indicator is currently in a neutral range around the 46th percentile, well above the 19th percentile at the late-July low, meaning there is still room for deleveraging before key tech catalysts arrive.^[1]

The defensive rotation is equally clear. The Goldman Sachs Prime Book shows that 8 of the 11 U.S. sectors saw net selling, with information technology and industrials leading the decline; within IT, software, semiconductors and semiconductor equipment, and IT services were all net sold; buying was concentrated in health care and consumer staples, with both sectors up more than 2% in Thursday morning trading, supporting the Dow, while the Nasdaq slid to a near one-month low, CoreWeave fell another 8.5%, and the Philadelphia Semiconductor Index declined for a seventh straight day. Amid this collective bleeding among fund managers, global systematic long/short funds rose 0.5% on the day, up 1.9% in October, leading all strategies with a year-to-date return of 21.7%.^[1]

TSMC: Three-Year Capex Far Exceeds $100 Billion

TSMC's third-quarter results released on October 15 showed quarterly revenue of $46.6 billion, up 16% quarter over quarter, about 2% above the Bloomberg consensus; fourth-quarter revenue guidance midpoint was $53.3 billion, about 6% above consensus, with full-year growth guidance of 44%. More closely watched was the investment signal: management confirmed 2026 capex will remain in the $60 billion to $64 billion range, and made clear that combined capex for 2026 to 2028 will far exceed the historical total of about $101 billion for 2023 to 2025. A JPMorgan report by Mio Shikanai's team on October 16 argued that management not only maintained near-term guidance but extended capex certainty to 2028 and even around 2030, providing semiconductor equipment makers with long-term order visibility beyond a single fiscal year; equipment makers with a higher share of revenue from foundry customers benefit most directly.^[5]

The expansion roadmap is equally specific. In Taiwan, TSMC plans to build 13 advanced chip and advanced packaging fabs, while overseas it plans to add five to six fabs. By node, A14 technology is expected to enter risk production in 2027 and volume production in 2028, N2 is planned to double capacity in 2027, and N3 expansion is stronger than previously expected; the original new fab in Tainan is scheduled to start production in the first half of 2027, and due to strong customer demand, management decided to build an additional Tainan fab, expected to enter volume production in the first half of 2028. Overseas, construction of the second fab in Arizona, U.S., is complete and in the equipment installation stage, with volume production planned for the third quarter of 2027, slightly ahead of the original schedule; volume production of N3 at the second Kumamoto fab in Japan is expected to begin in 2028. Management also mentioned prospectively that it plans to introduce high-NA EUV lithography equipment around 2030, at which point capex levels may be significantly higher than current levels.^[5]

Jensen Huang Responds to the Question of Who Pays for Compute

The world's top 20 tech stocks shed more than $1 trillion in combined market value this week, as doubts about the return on AI infrastructure investment heated up sharply. One pressure point is CoreWeave: according to ICE Data Services, the five-year credit default swap spread insuring against the company's debt default risk surged as much as 74.5 basis points intraday, hitting an intraday record high of 1,194.5 basis points, implying a five-year default probability of more than 60%. Nvidia is both its equipment supplier and, with about a 10% stake, its second-largest shareholder; this circular financing structure has raised questions about the financial sustainability of its expansion. In an interview with Bloomberg in Stockholm, Jensen Huang said CoreWeave does not need Nvidia to provide backstop support and that its cost of capital is actually falling; CoreWeave closed up 5.94% against the trend that day, halting a two-day cumulative plunge of more than 18%, and the Nasdaq subsequently recovered all its losses.^[6]

In a nearly 140-minute interview on the BG2 podcast, Huang broke Nvidia's compute demand into three roughly equal parts: about one-third is basic computing entirely unrelated to AI labs, including molecular dynamics, physics simulation, and classical machine learning, a business growing 30% to 40% annually; about one-quarter to one-third comes from frontier labs, and he sees five or six institutions rather than the two the market usually recognizes; another roughly one-third comes from thousands of startups renting compute, 80% of which use open-source models. He said betting trillions of dollars of construction on the revenue of two labs would be a big mistake. He also disclosed that Nvidia's market value per gigawatt of compute has risen from about $18 billion to more than $20 billion, now exceeding $40 billion, and still rising; it accounts for about 40% of global cloud service providers' capex and about 80% on the non-cloud side. His view for 2027 is that demand will only be greater, with bottlenecks in land, permits, power, and construction labor, while power scarcity actually benefits platforms with the highest compute output per watt; by the end of this decade, annual construction scale will be between $3 trillion and $4 trillion, closer to $4 trillion, of which pure computing is about $1.5 trillion.^[6]

BlackRock Assets Under Management Rise to $15.49 Trillion

BlackRock released third-quarter results before the U.S. market open on October 14, with assets under management reaching $15.49 trillion, a record high, equivalent to half of U.S. annual nominal GDP (about $31 trillion to $32.5 trillion). The company's quarterly net inflows reached $205 billion, with cumulative net inflows of $869 billion over the past year; revenue rose 19% year over year to $7.75 billion, GAAP earnings per share were $10.57, and adjusted earnings per share were $15.93. The flagship iShares ETF lineup attracted nearly $135 billion in net inflows this quarter, while the company repurchased $550 million of stock in the same period. Chairman and CEO Larry Fink said the company is entering the seasonally strongest fourth quarter with accelerating momentum, strong market fundamentals, and continued improvement in corporate margins and earnings momentum. The stock spiked and then fell back on the day results were released, finishing flat, but buying returned the next day and the stock closed up more than 4%.^[3]

The private credit business showed relative resilience. BlackRock's net inflows in that business increased by $5 billion quarter over quarter, and redemption requests for corporate loan funds fell from 13.3% in the second quarter to about 11.5% in the latest quarter, easing redemption pressure at the margin. Chief Financial Officer Martin Small said on the analyst call that investment demand for private credit is broadly solid and that there is a larger, more durable opportunity in providing private high-grade credit across the platform. However, rating agency Moody's warned in an October 14 report that, dragged down by a combination of persistently higher interest rates, refinancing difficulties, redemption pressure, and insufficient transparency, asset quality risks are building for weaker middle-market borrowers in the U.S. private credit market.^[3]

Asian Markets Open Higher on October 16; Defensives and Tech Rebound Together

On October 16, A-shares opened broadly higher in morning trading; as of publication, the Shanghai Composite was up 0.49%, the Shenzhen Component up 0.65%, and the ChiNext Index up 0.51%, with ChiNext briefly up nearly 1.8% at the open before turning negative; Hong Kong stocks opened higher and extended gains, with the Hang Seng Index up 1.31%, the Hang Seng Tech Index up 2.19%, MINIMAX up more than 6%, Kuaishou up more than 5%, and Tencent, Bilibili, Horizon Robotics and others up more than 4%. At the sector level, defense and military industries rose strongly, with Chenxi Aviation and Great Wall Military Industry hitting the daily limit first, while Shengnan Technology, Sinan Navigation, Tongyi Aerospace and others surged; innovative drugs and biotech rose in tandem, with the Hang Seng Biotech Index up more than 3%, and WuXi AppTec, Asymchem, Joinn Laboratories and others climbing together; in the U.S. pharma camp, AbbVie and Eli Lilly approached record highs, seen as a signal that investors are rotating to defensive assets amid turmoil in AI concept stocks. JAC Motors hit the daily limit after the Ministry of Industry and Information Technology and two other departments set up a joint investigation team to verify online reports of a broken brake pedal bracket on the Zunjie V800 vehicle.^[4]

AI hardware also recovered, with optical module and optical communication concept stocks rising; Zhongji Innolight opened nearly 3% higher, and the company plans to change the use of A-shares repurchased in September 2026 and cancel all of them; printed circuit boards, copper-clad laminates, and lithium battery sectors fell into correction, with T&S Communications down nearly 4%, while Victory Giant Technology, Eoptolink Technology, and Lens Technology weakened. On the news front, AstraZeneca signed a $1.86 billion AI drug discovery agreement with Insitro on October 15, aimed at finding new drugs to treat liver and heart diseases, driving activity in A-share AI pharma concept stocks, with Hongbo Pharmaceuticals hitting the 20cm daily limit. In the bond market, government bond futures rebounded across the board, with the 30-year main contract up 0.04%, the 10-year up 0.04%, the 5-year up 0.05%, and the 2-year up 0.02%; most domestic commodity futures rose, with palladium, Shanghai silver, rubber, industrial silicon and others up more than 1%.^[4]

Two Main Threads in the Same Week

Putting these threads together, capital is doing two things: continuing to reduce crowded AI long positions on one hand, and rotating allocations toward defensives such as health care, consumer staples, and innovative drugs on the other; the October 16 rally in defense, innovative drugs, and biotech in Asian markets is a continuation of that preference. The disagreement is over where the correction stands: AI net exposure is still in the 83rd percentile on a three-year basis and gross leverage in the 94th percentile on a five-year basis, with Goldman Sachs arguing deleveraging is closer to the beginning than the end; JPMorgan, by contrast, leans toward buying the dip, with its TPM positioning indicator still in neutral territory. On fundamentals, TSMC has extended capex visibility to 2028 and even around 2030, while Jensen Huang is trying to prove that compute demand does not depend on the revenue of two labs; whether capital accepts these two propositions will determine whether the selling pressure on tech stocks is temporary or trend-like. BlackRock's record assets under management and the marginal easing of private credit redemption pressure offer another angle, while Moody's warning on middle-market borrowers suggests credit risk under persistently high rates still needs monitoring.^[1][3][5][6]