Finding Opportunities in the Holdings of Big Players — Q2 2026 13F deep dive · what fourteen funds bought, what the filings proved, and the

Fourteen elite managers filed their June-quarter books on 14 August 2026. Ten of the names they were accumulating have since reported, and every single one beat — then five recovered their quarter-end price and five are still 18–33% below it. Nothing in the fundamentals separates the two groups.

SpaceX priced 638.9 million Class A shares at $135.00, over-allotment included, on Nasdaq as SPCX. Eighteen trading days later the quarter closed and the position became reportable. Across the roster, eight of fourteen managers held $25.07B of it — every one of those positions opened in the quarter. Nobody trimmed.

It was not only the hedge funds. Nvidia disclosed 122.8 million shares worth roughly $21B, AMD 3.31 million worth $565M, Harvard's endowment $2.2B, and Saudi Arabia's Public Investment Fund 154,146,835 shares worth $26.34B. A second new listing, Cerebras Systems, pulled in another $2.26B from four roster managers — again all opened, none trimmed.

A move means nothing until you know whose move it is. Bridgewater trimming a name is portfolio mechanics across 997 lines; Hohn trimming one of eleven is a decision. All sizes are disclosed US-listed equity at 30 June 2026.

Two filing notes changed what the roster looks like. Ackman's holdings arrived under a new EDGAR filer identity — Pershing Square Capital Management filed a notice saying its securities are now reported by Pershing Square Inc., the listed parent. And Einhorn's book files as DME Capital Management; the Greenlight Capital filer has been dormant since 2023.

A 13F is a photograph of long, US-listed equity positions on the last day of a quarter, published up to forty-five days later. It shows no shorts, no bonds, no cash, no non-US lines and no private positions, and it gives no reasons. It is also stale on arrival: between 30 June and 14 August TTM fell 47%, recovered, and finished 25% below its quarter-end mark.

What it does show is unfalsifiable and worth a great deal: share counts, at a date, from people who had to sign them. The discipline that follows is to read the diff in shares rather than in value, and to weigh agreement across unaffiliated books over any single fund's move.

Third Point is the cleanest test, because its filing was read everywhere as a repudiation of AI. Loeb exited Nvidia, Broadcom, KLA, Lam Research and the semiconductor ETF. Those five exits totalled $96.1M out of a book that ended the quarter at $4.65B, having more than doubled from $2.08B. In the same filing he added 67.3% to TSMC and 150% to ASML, and opened Keysight at $204.8M, Flex at $166.9M and TTM at $94.4M.

The shape repeats. Tiger Global cut Alphabet 45.4% and TSMC 12.3%, and opened Cerebras at $662.8M, AMD at $392.0M, Seagate at $275.1M and Visa at $274.0M. Coatue trimmed Applied Materials 19.7% while lifting Micron 1,794% and opening Intel at $1.69B. Selling the most-owned line to fund a less-owned one is not a change of view about artificial intelligence.

Then July happened. Between 21 and 29 July the complex fell in near-lockstep and bottomed within three days of itself — Astera −49%, Nebius −47%, TTM −47%, Sterling −41%, Flex −38%. Names do not correlate that tightly on their own news. The widely reported cause was the forced unwind of a large leveraged AI fund up 439% through June that lost roughly two-thirds in days. That account is press and commentary, not a filing, so treat the cause as reported and the effect as measured.

The counter-evidence worth stating: Nvidia was genuinely net-sold. Five managers held it, one bought, six sold, roster net −4,285,427 shares. That is a real reduction in the most crowded position in the market — and entirely consistent with the funding story, since Nvidia is exactly what you sell first when you need $25B.

An accelerator on its own computes nothing. It has to be fed — power in, heat out, and data across to the thousands of other accelerators sharing the job. As rack designs went from air-cooled boxes to liquid-cooled systems drawing tens of kilowatts, the parts carrying those signals stopped being commodities.

The second engine at TTM has nothing to do with AI. Aerospace and defence carries the Golden Dome programme and multiple munition programmes, and the company is expanding existing sites into it — brownfield expansion inside buildings it already owns, which is why capacity arrives in quarters rather than years. Two acquisitions under $220M are expected to close in Q3.

At Astera the transition that matters is a product one. CEO Jitendra Mohan said the Scorpio fabric switch family — a 320-lane part — becomes the largest product family in Q3, a quarter earlier than previously expected. Astera began selling one retimer into one socket; a fabric switch is a larger and stickier position in the rack. Seventy-three per cent gross margins are what a company earns when the alternative to its part is a board redesign.

Below the silicon, an AI data centre is a construction project with an electricity problem. Sterling builds the site before anything is installed; Flex builds the power and cooling that goes into it.

The Flex gap is the entire thesis, because the market currently prices the blended average. Flex intends to separate Cloud and Power Infrastructure — data-centre power systems, thermal management, liquid cooling, rack infrastructure and power conversion — with completion targeted for the first quarter of calendar 2027. When it stands alone it gets compared to AI-infrastructure peers rather than to contract manufacturers. Third Point opened the position in the quarter that segment crossed $2.2B.

SpaceX is not a story stock. In the June quarter it reported revenue of $7,814M against $4,071M, up 92%, and $12,508M across the half against $8,138M — a company earning at roughly a $31B annual run rate and roughly doubling. Revenue splits across launch services, connectivity (where all product revenue sits, meaning Starlink hardware) and an enterprise and government line that includes Starlink Mobile.

The institutional response was the most one-sided move of the quarter. Add the non-roster holders visible in the same filing round — Nvidia's 122.8 million shares, the Public Investment Fund's $26.34B, Harvard's $2.2B, AMD's $565M, Thiel and Founders Fund's 427.3 million shares — and this was less a stock-picking decision than an index-like scramble to own a company unavailable for twenty years. That is also the risk: a position everyone had to establish in eighteen days has no natural marginal buyer once established, and the lock-up expiry adds supply into that vacuum.

Cerebras is the smaller, harder version of the same trade. It drew $2.26B from four roster managers — Coatue $1.55B, Tiger $662.8M — all opened, plus a reported new $2B position from Altimeter and a new stake from Citadel. It sells wafer-scale accelerators, which is to say it competes directly with the most formidable company in the market rather than supplying it. That is a materially different risk profile from every other name here.

Two names on the roster were not part of the funding event at all, and that is precisely why they belong here. S&P Global's Q2 revenue was $4,146M against $3,755M, up 10.4% — but operating profit rose 16.8% to $1,812M and diluted earnings per share went from $3.50 to $4.12, helped by a share count falling from 306.1 million to 295.5 million. Ratings, indices and analytics compound at low incremental cost, which is why a 10% revenue line becomes a 17% profit line.

Alphabet is the roster's single largest conviction: sixteen positions across two share classes, ten buyers, and Berkshire lifting its stake 83% to 105,979,600 shares worth $37.77B, making it a top-three holding. Third Point added 485.7%, Hohn 12.3%, Coatue 12.6%. The disagreement is real and worth naming — Tiger cut 45.4%, Viking exited entirely and Ackman sold out. That split is a genuine argument about whether Alphabet's capital expenditure earns its return, had between managers reading the same disclosures.

Ranking by holder count and dollars sorts for the companies you have already read about. A position one specialist opened in a $3B business sits at the bottom of that table and disappears under any "held by 3+ funds" filter. So the same fourteen filings were re-screened on the opposite axis — conviction relative to obscurity: weight inside the manager's own book, whether it was newly opened, and fewer holders scoring better rather than worse.

TeraWulf is the one worth the work, and a growth screen would throw it away. Revenue fell 6%, which on that line alone reads as a shrinking business. The composition says otherwise: high-performance-compute lease revenue was 71% of the total against 0% a year earlier, with mining down from 100% to 29%. This is a bitcoin miner that has substantially finished converting into a data-centre landlord, and the transition shows up as flat revenue precisely because one business is replacing another rather than adding to it.

The honest caveat: these are one-fund positions, carrying none of the cross-manager corroboration that makes the main basket worth anything. A single manager can be wrong in a way five unaffiliated managers usually are not. What they carry instead is the thing the main basket has already lost — nobody has written them up, and the price has not been bid by people reading the same filings. Size them accordingly.

Nebius fell 47% — as deep as TTM — and is now above its 30 June mark. Same drawdown, same beat, opposite outcome. No fundamental variable separates the two groups, which is the strongest available argument that what is left in the first group is positioning residue rather than information.

TTM is the position to hold in size: three-fund accumulation across different styles, a fully validated print, defence revenue that ignores AI capex cycles, and the tightest technical box in the group. Flex is the second call and the cheaper entry, with the Cloud and Power separation ahead of it. Astera and Sterling are the higher-beta expressions of the same idea and belong in smaller size. SpaceX is a genuine starter kept small for the lock-up. S&P Global and Alphabet earn their slots precisely because they sat out the funding event. Nebius, Seagate and Cerebras have already made the round trip.