What the best 13F funds owned before the AI crash, and whether they saw it coming

Form 13F books are dated June 30. That turned out to be the top: SOXX, the iShares semiconductor ETF, made its post-quarter high on June 30 itself and fell 27.4% to a low on July 29, 2026. QQQ fell 10.1% over the same days and SPY 2.3%. This study uses two windows, June 30 to July 29 and June 30 to September 24, 2026, and measures every fund's book over both. By September 24, 2026 SOXX was −12.4% versus June 30, QQQ +0.5% and SPY +2.9%.

It was a semiconductor drawdown, not a market one. Among the industries where the top-rated funds had $5bn or more, the largest falls to the trough were electrical equipment & parts (−34%), hardware, equipment & parts (−29%), auto - manufacturers (−28%), computer hardware (−27%) and semiconductors (−22%, $635.5bn of elite positions). Internet, software, drugs and banks were flat to slightly down.

Add up the top-50 positions of the 763 top-rated funds and the book is $2.70tn. The largest names are the ones every 13F reader expects: Nvidia ($133.0bn, 383 holders, 4.9% of elite dollars), Broadcom ($115.6bn), Micron ($96.0bn, 243 holders), then Microsoft, Amazon, Alphabet, Apple, Meta and TSMC ($66.1bn). Define an AI / semis basket by industry label, with no hand-picking: semiconductors, computer hardware and communication equipment. That basket was $710.2bn, 26.3% of the elite book, held in 3,840 positions at 591 of the 763 funds. Semiconductors alone were 23.5%, more than twice the next industry (internet content & information, 9.1%).

To the trough, the basket fell 22.1% and the rest of the book 1.6%; the whole elite book, held unchanged, fell 7.0%, or $187.9bn, of which $156.6bn was the basket. Inside the basket the damage was uneven. Nvidia fell 5% and Broadcom 2%; Micron fell 36%, Intel 41%, KLA 44%, TSMC 22%. The two largest positions were the two that held up. By September 24, 2026 Nvidia was +12% versus June 30 and Micron −8%; KLA (−39%) and Intel (−11%) had not recovered.

Take the 564 top-rated funds whose March 31 book can be reconstructed (Q1 shares of the June 30 top 50, plus the Q1 top-50 positions they exited). At March 31 prices the basket was 19.5% of that book. Reprice the same March 31 book at June 30 prices and it is 26.6%: the second quarter's rally in memory and semiconductor equipment did that on its own. The actual June 30 book is 26.4%. A quarter of trading moved the elite's basket weight by 0.15 points. Per fund, the median basket weight went from 15.2% at March 31 to 20.1% at June 30; 38% of funds cut their weight by more than half a point relative to the no-trade book and 27% raised it.

In dollars the elite sold $128.0bn of basket stocks in Q2 and bought $88.6bn, a net $39.5bn out of a $710.2bn position. 380 funds were net sellers of the basket and 197 net buyers. The selling was concentrated in memory: Micron, net −$18.6bn across 157 sellers, then Seagate (−$6.9bn), Sandisk (−$5.5bn), Nvidia (−$5.3bn), Lam Research and Lumentum. The buying went to Intel (net $4.2bn), Western Digital, Cisco and KLA. Sandisk then fell 55% to the trough, Micron 36%, Lam 42%; Intel, the largest net buy, fell 41%.

Multiply each sale by the stock's subsequent fall and the elite's basket selling avoided $37.5bn of drawdown to the trough; multiply each purchase the same way and the basket buying walked into $24.7bn. Net, the quarter's basket trades were worth $12.8bn against the $156.6bn the basket held at June 30 then lost. The largest single sales were Capital World Investors in MU ($14.5bn, which then fell 36%) and Sanders Capital, LLC in STX ($9.1bn, −21%); the largest purchases were Capital International Investors in INTC ($4.4bn, −41%) and Coatue Management LLC in MU ($3.4bn, −36%).

For each fund, hold the June 30 top-50 book with its June 30 weights and value it at the July 29 close and at the September 24, 2026 close. 569 top-rated funds have at least half their disclosed portfolio priced this way (median coverage 87%). The median book fell 5.8% to the trough; the middle half runs from −11.3% to −2.5%. 23.7% of books fell less than SPY's 2.3%, 71.9% fell less than QQQ's 10.1%, and 96.5% fell less than SOXX's 27.4%. 29.2% fell 10% or more; 13.4% rose.

The foil, the 1,278 bottom-rated funds, had a median book of +0.9% to the trough, and 76% of them beat SPY. That is not a fair fight and it is worth saying why: a low rating means the fund did not own the last three years' winners, which are the stocks that fell. The median bottom-rated fund had 4.9% of its book in the basket against 16.2% at the median top-rated fund.

Basket weight explains most of the spread. Top-rated funds with no basket position at all had a median book of −2.2% to the trough and 53% of them beat SPY. Under 10%: −4.6%. Ten to 20%: −5.2%. Twenty to 40%: −7.8%, with 8% beating SPY. Over 40% of the book in the basket: a median of −20.3%, 61 funds, 11% of the elite. By September 24, 2026 the order had changed: the 10-to-20% and 20-to-40% buckets had recovered to +0.4% and −0.2%, while the no-basket funds sat at −4.5%, because what they owned instead did not rally either.

To isolate what the quarter's trading did, value each fund's March 31 book over the same window and subtract. For 564 top-rated funds with a visible March 31 book, the June 30 book fell less than the March 31 book at 56.9% of funds, and the median effect was +0.2 pts (mean +0.7 pts). In aggregate dollars the June 30 books fell 6.7% and the March 31 books 6.8%: a difference of 0.04 points. The median March 31 book fell 6.4%; had no fund traded in Q2, 20.9% would have beaten SPY instead of 23.7%. To September 24, 2026 the picture is the same: trades helped at 56% of funds by a median of +0.1 pts.

The tails are where the decisions show. Madison Avenue Partners LP (+21.0 pts: sold SNDK, ICLR · bought INVH, CPNG), Ardmore Road Asset Management LP (+16.2 pts: sold SNDK, GLW · bought GWRE, AMT) and Westerly Capital Management LLC (+14.4 pts: sold MRVL, AIP · bought COUR, ANET) sold memory, optical or semiconductor names in the quarter and would otherwise have been among the hardest hit. Ratan Capital Management LP (−14.6 pts: sold NVDA, META · bought SIMO, QCOM), Mizuho Markets Cayman LP (−11.4 pts: sold GEHC, ADI · bought MRVL, PANW) and Berkshire Capital Holdings INC (−7.1 pts: sold NVDA, COHR · bought MU, AMD) bought into the quarter's rally. The bottom-rated funds, measured on history alone (their exits are not reconstructed), were helped at 62% of funds against 64% for the top-rated on the same basis.

Restrict the list to investable funds: at least $250m disclosed, eight or more positions, at least 60% of the book priced, no fund with more than half the book in one stock, and no venture, charitable, endowment or pension filers. 398 top-rated funds qualify; 22% of them beat SPY to the trough. The twenty whose frozen June 30 book moved most favourably are led by Gqg Partners LLC (+7.1%, PM 15%, CB 8%, PBR 6%), Pale Fire Capital Se (+6.9%), Ruffer LLP (+5.9%), Miller Value Partners LLC and Keenan Capital LLC. Of the twenty, 17 had under 1% of the book in the basket. These are tobacco, energy, banks, emerging markets, small-cap value and private-equity sponsor stakes: books built to own something other than the AI trade.

The other end of the same list is a roll-call of semiconductor and memory specialists. Berkshire Capital Holdings INC (−37.7%, 76% in the basket), AIGH Capital Management LLC (−36.5%, SNDK 31%, MXL 9%, INTC 8%), Oaktop Capital Management II L P (−34.6%, KLAC 39%, LRCX 38%, ADI 8%), WT Asset Management Ltd (−33.1%, 91% in the basket), Trivest Advisors Ltd (−31.9%, STX 22%, MU 22%, WDC 20%). Whale Rock Capital Management LLC, the largest fund on the list at $12.5bn, was −29.9% with 43% in the basket. Every one of the fifteen had at least 28% of the book in the basket at June 30.

Two things about this list. First, several of these funds did sell in Q2: Oaktop Capital Management II L P sold LRCX, KLAC · bought PINS, WT Asset Management Ltd sold SNDK, COHR · bought STM, AMAT, Trivest Advisors Ltd sold TER, MU · bought STX, WDC. Their Q2 trades helped by +3.3 pts and +1.7 pts, but a specialist rotating within the basket cannot trade its way out of the basket. Second, these are the funds that earned their rating by owning this trade for three years; the cost-basis study found the median top-rated Micron holder far above its estimated purchase price at June 30. A 36% fall from there is a drawdown, not a loss.

Three numbers answer the question. 76.3% of top-rated books fell more than the S&P 500 to the trough, and 29.2% fell 10% or more. The basket weight the elite carried into the crash was a product of price, not purchases: 19.5% at March 31, 26.6% by June 30 with no trades, 26.4% after the quarter's trades. And the quarter's trading, taken as a whole, was worth 0.04 points of drawdown in aggregate and +0.2 pts at the median fund, with 57% helped and 43% hurt. The basket selling avoided $37.5bn; the basket buying gave $24.7bn of it back.

What the elite did have was a book that was never all semis. 96.5% of top-rated books fell less than SOXX and 71.9% less than QQQ, because the median fund held 16% in the basket, not 100%, and the two largest basket positions (Nvidia, Broadcom) fell 5% and 2% against the basket's 22%. By September 24, 2026 the median top-rated book was −1.7% against June 30, with SPY at +2.9%, QQQ at +0.5% and SOXX at −12.4%; 41% of books were above their June 30 value. Still, 41% of top-rated funds held 20% or more of the book in the basket at June 30, and 64% held 10% or more. If the question is whether the best-rated 13F filers sold the top, the answer in the filings is no.

Search the 569 top-rated funds with a priced June 30 book. Columns: the frozen June 30 book's move to July 29 and to September 24, 2026, the effect of Q2 trades in points (June 30 book minus March 31 book, to the trough), the basket weight at June 30, and the share of the disclosed portfolio covered by priced top-50 positions.