Q2 2026 13F REVIEW — 53 MANAGERS, 19 READ OFF PRIMARY FILINGS
Graham Advisors / LC / 15 Aug
Three things in this quarter's filings are not visible on a standard 13F screen, and all three cut against the obvious read.
Start with SpaceX, because the marks are wrong. Divide value by shares on any filer holding SPCX and you get the same number: $170.86. D1 at $21.54bn over 126,042,232 shares, Nvidia at $20.98bn over 122,764,805, Atreides and Coatue identical. That is the 30 June mark. The stock priced at $135 on 12 June, ran to $225.64 intraday on the 16th, broke back through issue on 15 July and closed at $108.27 on 5 August, an all-time low. It trades near $140 today. So every SPCX line in these filings is carried roughly 18% above market, and D1's disclosed $34.78bn book — 61.9% of which is this one position — is overstated by something like $3.9bn. Anyone ranking managers by disclosed AUM off
these filings is ranking them on a stale print.
The lock-up matters more than the mark. SpaceX staggered its release
rather than running a single cliff. The 6 August tranche freed 911.5m
shares and took the float from 4.9% to 11.8%. A 10% bonus tranche was
written to release early if the stock held 30% above $135 into
earnings; it did not trigger, because the stock was between $108 and
$135. Roughly 7% tranches drip through to October. Then the largest
single release, about 28% or 1.3bn shares, lands at Q3 earnings in late
October or November, with full 180-day expiry on 8 December. Musk's
6.4bn shares are locked to 12 June 2027 with no early release. Only 5%
of the company traded at IPO. The point for positioning is simple: a 30
June ownership snapshot tells you very little about who owns this by
year-end, and the worst of the supply is still in front of us.
Second, the crowded trade already had its forced seller. Situational
Awareness held $20.24bn at 30 June with 55% in SanDisk and Micron,
levered around four times, up roughly 439% in the first half. In July
it lost about two thirds, took margin calls from Goldman, JPMorgan and
Bank of America, and sold a $16bn equity book to Citadel inside 24
hours. Read its holdings against the consensus accumulate list and they
are the same list — MU, SNDK, CORZ, APLD, NBIS, CRWV, RIOT, IREN. The
Q2 filings are a photograph of a portfolio that was liquidated three
weeks later. Treat any Q2 accumulation signal in those names with that
in mind.
Third, the most committed books are hedged, and you cannot see it long
only. Atreides carries a QQQ put of $2.36bn notional, 16.4% of the fund
and its second-largest line, sitting directly against its AI supply
chain positions. TOMS Capital's single largest position is a SPY put,
$1.19bn against a $4.67bn book. Appaloosa runs an AAPL put. Separately,
Millennium shows an unpaired NVDA put of $5.1bn. Citadel's much larger
MU and SNDK put lines are paired against comparable calls and are
dealer inventory, not a view — do not read those directionally.
Two smaller correctives. The Nvidia exit story is overstated: four
funds sold out, but Coatue still holds $1.21bn, Appaloosa $305m and
Atreides $250m, so the three most AI-literate books in the sample
stayed long. And ignore the Berkshire homebuilder headline — the DHI
line is 3,564 shares, $580,504, inside a $299bn book. The real housing
commitment is the $6.8bn Taylor Morrison acquisition in July, which
does not appear in the 13F at all.
One trade idea rather than an observation: the interconnect layer is
the least crowded part of this rotation. Astera Labs is held by both
Gavin Baker ($588m) and Cliff Sacerdote ($757m). Coherent is held by
both, plus Nvidia itself at $3.07bn. Nvidia's own eight-position book
reads as a supply-chain map — Intel $29.99bn, CRWV $4.70bn, COHR
$3.07bn, NOK $2.21bn, SNPS $2.15bn — and the funds selling Nvidia
bought into names Nvidia owns.
Full note attached. All figures aggregated per issuer from primary
13F-HR filings, not top-holdings summaries.