The claim
Tae Kim, on 30 July, read it as right positions and wrong leverage. At four times gross, a drawdown became a liquidation. That is the proximate cause. But leverage is judged against net exposure, and net exposure is a model. The arithmetic says the hedge was not sized wrong. It was built from the wrong instruments.
What the filing shows
The long leg is small and mid cap AI infrastructure, high beta. The disclosed short leg is puts on mega cap semiconductors and Oracle, low beta. One capital cycle, held twice.
| Item | What the filing says |
|---|---|
| Positions disclosed | Forty-two, totalling $13.68bn |
| Long leg | $3.86bn, small and mid cap AI infrastructure |
| Eight largest longs, about 91% of that leg | Bloom Energy, Sandisk, CoreWeave, IREN, Core Scientific, Applied Digital, Riot, CleanSpark |
| Disclosed short leg, puts | $8.46bn notional. VanEck semiconductor ETF, Nvidia, Oracle, Broadcom, AMD, Micron, TSMC, ASML, Intel |
| Calls also held | $1.36bn |
The hedging ceiling
From adjusted closes, June 2025 through July 2026, I built daily return baskets for the two sides. The ceiling is the most risk the short basket could ever remove from the long one. No sizing beats it.
| Window | Correlation | Ceiling: most any hedge can remove | Dollar neutral achieved |
|---|---|---|---|
| Before July | 0.70 | 28.6% | 27.2% |
| July | 0.95 | 67.3% | 42.8% |
The dollar-neutral version, equal dollars long and short, captured almost the whole ceiling. Sizing was near optimal and the outcome still poor. The pair of baskets is the problem.
Why 0.70 is the worst number
It is high enough that both sides fall together when the theme unwinds, and in July they did. It is too low for the short to do real work. Half the long book's variance is unexplained by the short book, and you keep that half however the hedge is sized. That is the instruments, not the risk management.
| Measure | Value |
|---|---|
| Share of the ceiling the dollar-neutral hedge captured | 95% |
| Baskets weighted by filing value | Correlation 0.76, 27.6% removed |
| Long variance unexplained by the short book | About half |
| Volatility retained versus outright long of the high beta leg | About three quarters |
| Four times gross at roughly flat net | About two dollars long against two dollars short |
| Volatility of that book | About 1.5 times an unlevered outright long of the high beta leg |
The cheap check
Regress the long book on the short book, on public prices, before sizing. Near 0.7 a book looks hedged on the exposure report and is not hedged in the returns.
What would overturn this
The disclosed short book is partial, and this caveat matters most. A 13F captures long positions and options, not short stock. If the undisclosed part was large and uncorrelated with the long leg, my correlations overstate the problem.
They held puts, not linear shorts. Treating put notional as linear short exposure runs against the fund. A put's short exposure rises as the underlying falls, so the real hedge probably worked better in July than my proxy.
The filing was already stale, and the fund's size, leverage and returns come from press reporting, not from me. The leverage arithmetic depends on that reported gross.
| Item | What the record says |
|---|---|
| Nebius | 5.6% position, May event date, absent from the March filing |
| A micro cap name | Position near 20%, June event date, absent from the March filing |
| Method | Baskets of daily returns of the disclosed long names and of the names underlying the disclosed puts, weighted equally and by filing value. Correlations and betas by ordinary least squares on daily returns. Volatility is the standard deviation of daily returns annualised at 252 days. The ceiling is the minimum variance result. A hedged book cannot fall below unhedged volatility times the square root of one minus the squared correlation. |
Educational research only. Not investment advice. This note is about a portfolio structure, not about any individual's conduct or ability.
Educational research only. Not investment advice. This note is about a portfolio structure. It makes no claim about any individual's conduct or ability.