What if the biggest names in the investing world already knew the AI party was ending weeks before the market finally noticed? That question kept looping in my head while I went through the latest round of 13F filings. These documents are usually boring paperwork. This time they felt more like a crime scene photo taken just before the lights went out.
The Quiet Exit Before The Storm
Every quarter the same reports land, showing what large funds held 45 days earlier. This batch was different. It captured positions as of September 30, only 36 days before the Nasdaq delivered its ugliest session since the early pandemic days. October alone wiped an estimated three percent from the broader hedge fund industry. Then came that brutal Thursday when the semiconductor index collapsed more than twelve percent in a single session. The AI trade, which had carried so many books higher, suddenly looked like a trap.
Tech-focused long-short funds that had been up roughly twenty-four percent at their mid-year peak were essentially flat by early November. High-beta tech and media names that had rocketed higher earlier in the year gave back more than half their gains in a matter of weeks. Some managers were still buying the same hyperscaler names the day before the selling pressure intensified. That kind of late rotation rarely ends well.
I decided to dig into ten of the most watched books: Duquesne, Appaloosa, Lone Pine, the Soros vehicles, BlueCrest, Situational Awareness, Pershing Square, Greenlight, Berkshire, and even Nvidia’s own corporate portfolio. The goal was simple. Who was already reducing risk while the rest of the market was still celebrating?
Three Sellers And One Stubborn Buyer
The semiconductor and related equipment names told the clearest story. Three of the four heavy chip books were cutting exposure. Only one was adding. That split already said a lot about internal conviction levels heading into the fourth quarter.
Stanley Druckenmiller did what he has done many times before. He left early. His Duquesne fund slashed its TSMC position by ninety-four percent, turning a position worth nearly three hundred million into a small residual stake. He halved the STMicro holding and completely exited AMD, Sandisk, Lam Research, Arm, Entegris, Rambus, and a couple of smaller names. Chip and storage exposure fell from sixteen percent of the book to roughly ten percent. What remained was dominated by one surprising name: Seagate. He nearly tripled that position and layered on call options. He also opened a fresh basket of gold miners. Hard drives and gold. Not exactly the classic AI portfolio you would expect from a manager who had ridden the previous wave.
Lone Pine’s moves were more complicated and, in hindsight, more painful. The flagship fund had been up sharply through June, then suffered a steep third-quarter drawdown. The 13F shows the internal rearranging. Applied Materials was cut by ninety-two percent. ASML and Teradyne saw meaningful reductions. Seagate, Corning, and TTM Technologies were sold completely. Several AI-power and bitcoin-miner related names were also trimmed hard. At the same time the fund built a large new Intel position and dramatically increased its TSMC holding. Core chip exposure stayed roughly flat in dollar terms, just shifted into different names. That rotation did not protect the book when the broader semiconductor complex sold off in October and early November.
David Tepper took the opposite side. Appaloosa increased its Micron stake by twenty-eight percent, making it the largest position in the fund. Nvidia was raised another thirty-one percent. Lam Research and ASML saw additions, and a new Applied Materials line appeared. Overall semiconductor exposure climbed from thirty-five to thirty-nine percent of the book. He did cut Alphabet and Meta, and he multiplied his SpaceX holding many times over. Still, the core bet remained the chips. Five weeks later Micron and Lam each dropped about seventeen percent in a single session. Timing is rarely kind in these situations.
The Vanishing Eighteen Billion
No filing drew more attention than the one from Situational Awareness. The reported 13F book collapsed from more than twenty billion at the end of June to just under 1.6 billion by the end of September. That is an eighteen-point-seven billion dollar reduction, roughly ninety-two percent of the previous reported long book. Twenty-one of twenty-six positions disappeared entirely. The biggest exits included multi-billion dollar stakes in Sandisk, Micron, Bloom Energy, TSMC, and Nebius.
Those who followed the story in real time already knew what had happened. Heavy leverage on a concentrated group of AI momentum names produced an abrupt unwind in late July. A large block of the public book was reportedly sold at a discount. By September the fund was back, but the new book looked very different. Roughly two-thirds of the remaining reported value sat in call options on many of the same names that had caused the earlier damage. Bloom Energy calls exploded higher in notional size. Fresh calls appeared on Nebius, Sandisk, and a memory ETF. The largest outright stock position was a smaller AI-related name that had been increased by fifty percent. In other words, the fund that had been forced to liquidate because of excessive leverage returned with a fresh layer of options exposure just in time for the next leg lower.
The fund that was liquidated for running too much leverage in a handful of AI momentum names came back with options on a handful of AI momentum names.
That sequence is worth sitting with for a moment. Markets sometimes reward resilience. They more often punish the same concentrated risk-taking when it is repeated too quickly.
Everyone Wanted Intel Except The Company That Knows Demand Best
If one name attracted broad buying interest in the third quarter, it was Intel. Several large funds either initiated or dramatically increased positions. One vehicle multiplied its stake nearly thirty-fold. Another opened a brand-new position worth more than seven hundred million. In retrospect the timing looks unfortunate given the subsequent price action, but the consensus was clear at the time.
The other side of that trade included Nvidia itself. The company’s own 13F showed a meaningful reduction in its Intel holding, amounting to roughly 6.5 billion dollars of stock sold at quarter-end prices. Nvidia also cut its Nokia and Synopsys positions and added further to SpaceX, which became its largest reported holding. The overall size of Nvidia’s 13F book declined by about a quarter. While public commentary still celebrated the AI infrastructure build-out, the company’s own portfolio was quietly monetizing one of its best-performing equity stakes.
SpaceX itself produced a split decision among the group. One manager increased the position seventeen-fold. Another opened a small new line. Nvidia added modestly. Meanwhile a different large fund sold ninety-four percent of its stake. Consensus is rarely as clean as the narrative suggests.
What The Macro Books Were Really Saying
The two more traditional macro-oriented books offered a different window into market expectations. One fund made its largest reported line a substantial put position on the energy sector, later supplemented by additional puts on an oil and gas exploration ETF. Combined, the notional short exposure approached a billion dollars against a much smaller overall book. That positioning aged better than many pure long energy bets once political and geopolitical premiums faded.
The same book also bought calls on long-duration Treasuries and utilities. Owning those instruments while the Federal Reserve was still in a restrictive stance required conviction. The other macro vehicle took a different but equally interesting approach. It sold its equity index ETF holdings almost entirely and replaced them with a much larger call position on the S&P 500. At the same time it increased puts on high-yield credit, financials, and small-cap indexes, while also adding Treasury calls. Long the broad market through options, short credit and smaller companies. That structure provided a measure of protection when volatility finally arrived.
The Quiet Books That Owned Almost No Chips
Three of the ten funds examined held essentially zero semiconductor exposure. Their third-quarter activity looked calmer, which is hardly surprising. Berkshire made one new purchase in an Italian software roll-up that had recently gone public and nearly doubled its position in a large homebuilder. It reduced its Bank of America holding by a sizable amount while leaving the massive Apple stake untouched. With a cash pile reported in the hundreds of billions, the overall posture remained one of patience rather than aggressive risk-taking.
Another well-known activist fund initiated two new billion-dollar positions and paid for them by trimming Microsoft and several other large holdings. The jump in the overall reported book size was driven largely by the appreciation of a single internal holding rather than a broad expansion of risk. A third value-oriented manager increased a California utility position to a round ten million shares, more than doubled a large media holding, and exited several other names. Interestingly, that same utility attracted fresh capital from three additional funds in the broader sample. It was the closest thing to a non-AI consensus trade that appeared across the filings.
Widening The Lens Changes The Picture
When the sample expands beyond the original ten names to include additional well-followed vehicles, the sense of a coordinated exit softens. Amazon saw an equal number of buyers and sellers. Meta attracted more buyers than sellers. Alphabet was the most sold of the mega-cap technology names. The cleaner directional signals sat in Intel, where buyers clearly outnumbered sellers, and in Micron, where most funds that already owned it added while only the previously mentioned leveraged vehicle exited. Broadcom and Nebius saw no new buyers and several outright sales within the group.
Those mixed signals matter for what comes next. Recent prime brokerage data showed the long-short ratio in U.S. information technology at the lowest level in records going back nearly a decade. Hedge funds as a group stood more than twenty points underweight the sector relative to the broader market. The de-risking that many traders described as the largest gross exposure reduction since the early pandemic period did not begin in October. It was already visible in the September 30 snapshots. Some managers had reduced chip exposure months earlier. Others had merely rotated. A few had increased. The sharp moves in late October and early November simply sorted those groups more publicly.
Bulls will argue that a record-low long-short ratio creates the perfect setup for a powerful squeeze higher. That view is not crazy. History contains plenty of examples where extreme underweight positioning preceded strong sector rebounds. At the same time the filings also show less comforting details. The fund that suffered a severe margin-driven unwind in July was already rebuilding leveraged option exposure by September. The largest dedicated chip long among the original sample was still adding into the weakness. And the corporate investor with arguably the best real-time view of AI hardware demand was itself a net seller of a major semiconductor name.
Hard Drives, Gold, And The Value Of Leaving Early
Perhaps the most interesting portfolio construction choice belonged to the manager who cut nearly every traditional AI-related semiconductor name and concentrated the remaining exposure in Seagate while adding a basket of gold miners. Hard drives and precious metals. That combination does not appear in many momentum-oriented presentations. Yet the same investor has a long history of stepping aside before the crowd recognizes the risk. Following that approach in real time is difficult. Reading about it forty-five days later is easier and far less profitable.
The next round of filings will arrive in February and will capture the positions held through the end of December. By then the market will have delivered its own verdict on how many of the remaining AI-related longs survived the volatility. I suspect we will see a longer list of red columns than green ones. Markets have a way of enforcing humility on even the most confident books when concentration meets a sudden change in liquidity.
In the end the 13Fs do not tell us who will be right over the next twelve months. They do show, with unusual clarity, who was already adjusting risk while the narrative remained overwhelmingly bullish. That information is rarely available in real time. When it does appear, it deserves careful attention rather than reflexive dismissal. The managers who reduced exposure in the third quarter did not need perfect foresight. They only needed to recognize that the easy part of the trade had already occurred. That recognition alone separated them from those still adding into the final stretch of the rally.
Looking across the full set of books, a few practical observations stand out. First, concentration risk remains underappreciated until the moment it is not. Second, the difference between rotating within a sector and actually reducing sector exposure is often larger than it appears on paper. Third, corporate investors with superior information sometimes act more cautiously than the funds that follow their public commentary. Finally, the managers who have survived multiple cycles tend to leave earlier than feels comfortable at the time. Those lessons are not new, yet they are rediscovered with surprising regularity whenever a dominant narrative begins to fray.
The AI infrastructure build-out is real. Capital spending numbers remain elevated. Demand for advanced chips has not disappeared overnight. None of that changes the fact that valuations, positioning, and leverage had reached levels that left little room for disappointment. When the inevitable pause arrived, the funds that had already lightened exposure were better positioned to take advantage of the resulting dislocations. Those still fully invested, or worse, still adding, faced a more difficult path.
Investors reading these filings months later can extract useful signals without needing to copy any single manager. The common thread among the more defensive books was a willingness to sell strength rather than wait for confirmation of weakness. That discipline is harder to practice than it sounds, especially when the surrounding narrative remains powerful and the short-term performance pressure is intense. Yet the managers who applied it in the third quarter avoided at least some of the damage that followed.
As the calendar moves forward, attention will shift to the next set of numbers. Until then, the September 30 snapshot remains the clearest available picture of how the most closely watched portfolios were positioned when the AI trade still looked relatively intact. Some were already walking toward the exits. Others were still ordering another round. The market eventually made the distinction clear. The filings simply recorded the choices in advance.
One additional pattern deserves mention. Across both the original sample and the expanded group, the funds that avoided semiconductors entirely or kept exposure modest experienced far less drama in their quarterly changes. Their activity centered on utilities, housing-related names, selective industrials, and the occasional new public company. That quieter approach produced smaller absolute returns during the peak of the AI rally, yet it also produced fewer sleepless nights once the correction began. In a year when many high-profile books swung from large gains to flat or negative territory in a matter of weeks, the value of simply not participating in the most crowded trades became apparent once again.
The temptation to chase the dominant theme of the moment is always strong. Performance rankings, investor pressure, and media attention all pull in the same direction. The managers who resisted that pull in the third quarter did not necessarily forecast the exact timing of the November session. They simply decided that the risk-reward in the most popular names no longer justified the capital at risk. That judgment call, more than any single stock selection, separated the early reducers from the late holders.
Future market historians will likely look back at this period as another example of how quickly a powerful narrative can reverse when positioning becomes one-sided. The 13F data provides the contemporaneous evidence. It is imperfect, lagged, and incomplete. It still offers one of the better windows available into the thinking of the largest and most sophisticated pools of capital just before the weather changed. For anyone trying to understand how the AI trade evolved from consensus long to source of forced selling, those filings remain essential reading.
The story is not finished. New capital will enter the sector. Existing holders will reassess. Corporate spending plans will be revised. Through all of it, the discipline of reducing exposure when the easy gains have already been captured will remain rare and valuable. The managers who demonstrated that discipline in the third quarter have given the rest of the market a useful case study. Whether the broader community absorbs the lesson is, as always, a separate question.