A fund up 439 per cent went looking for money
Leopold Aschenbrenner left OpenAI and started a fund built on a single argument: that the market was systematically underpricing how much capital artificial intelligence would absorb. Through 30 June 2026 the argument had worked to a degree that is rare in the industry. Situational Awareness, running roughly 20 billion dollars, was up 439 per cent on a net basis, a multiple of about 5.4 times. Its backers were the sort of names that do not need the publicity: Patrick and John Collison of Stripe, Nat Friedman, Daniel Gross, and the proprietary trading firm Jane Street.
By the end of July the fund was in conversations with existing investors and lenders about raising fresh capital after heavy losses in the selloff in artificial-intelligence shares, and some investors had been offered the option of buying assets out of the portfolio. The invitation to commit new money runs from 1 August. Aschenbrenner told investors the portfolio had been affected by the volatility and framed the correction as an opportunity rather than a verdict. Both of those things can be true. What matters for anyone outside the fund is the timing: the reversal from a 439 per cent gain to a capital call sits inside roughly four weeks.
What the position disclosure actually showed
The fund's quarterly position disclosure, covering the end of March and made public in May, is the part worth reading closely. It listed a notional portfolio of about 137 billion dollars against roughly 20 billion dollars of assets, a ratio of about 6.9 to one. More than 60 per cent of the disclosed value sat in put options on artificial-intelligence hardware. The named positions included new puts of more than 1.5 billion dollars tied to Nvidia and more than 2 billion dollars tied to the VanEck Semiconductor exchange-traded fund, alongside Broadcom, Oracle, Advanced Micro Devices and Taiwan Semiconductor.
At the same time the fund increased long positions in a different set of companies entirely: CleanSpark, Riot Platforms, Applied Digital, IREN and CoreWeave. That is not the chip layer. That is the layer that houses chips, powers them and rents them out. Four investment professionals were running the whole thing, which works out at about 5 billion dollars of assets each.
Two legs, one bet
Read those two lists together and the structure is clear. The fund bought crash insurance on the companies that make artificial-intelligence silicon and went long the companies that supply power, buildings and rented capacity to run it. That reads like a hedge. It is not one. Both legs are positions on the same underlying variable, which is the pace at which capital keeps flowing into artificial-intelligence build-out. The trade expressed a view about where inside the theme the money would land, not whether the theme would hold.
July tested the difference. The companies reported to have come under pressure in the portfolio include Oracle, Advanced Micro Devices, Nebius, Sharon AI, Bloom Energy and memory-related businesses. Power names and compute-landlord names are on that list next to chip names. When the market repriced the pace of the build-out, it repriced the whole stack at once, and the split between silicon and everything that surrounds silicon did not provide the separation the structure implied. That is the finding, and it generalises well beyond one fund.
What is not disclosed, and why it matters here
Precision is worth more than drama on a story like this, so the limits are worth stating plainly. The size of the loss has not been disclosed. The position disclosure is a quarter-end snapshot from the end of March, published in May, and positions can have changed entirely since then. That filing type shows long equity and option positions and does not show short stock, so it is a partial view by construction. Notional value on an options book is not capital at risk, and the 6.9 to one ratio describes exposure rather than borrowing.
None of that softens the central observation, because the central observation does not depend on the loss figure. The fund's own disclosure is enough. It shows a portfolio that was simultaneously bearish on chip makers and bullish on the infrastructure those chips sit in, and reporting on the July drawdown shows both categories under pressure in the same weeks. You do not need to know the drawdown to draw the conclusion about correlation.
If you buy compute from this cohort, price it in
The practical consequence lands on procurement rather than on anyone's trading book. A European business renting graphics-processing capacity from a listed specialist provider, or signing a multi-year colocation agreement with one, has been told for two years that this diversifies away from vendor concentration. It does not. Those providers finance their build-out in public markets, and their equity moves with the same sentiment that moves the chip makers they buy from. A move against Nvidia is not a move in your supplier's favour. It is the same move arriving one layer down, usually with more leverage attached.
The question to put to a compute supplier before signing is narrow and answerable. What happens to your financing if the semiconductor complex falls 20 per cent, how much of your capacity is contracted to named tenants, and what is your debt maturity schedule. A provider with committed tenants and long maturities is a different counterparty from one funding construction against a rising share price. Contract terms should follow that difference: shorter commitments, break rights, or a second supplier where a single failure would stop the work.
Read next: Renting Frontier AI Is Now a Choice, Not a Must | Amazon Borrowed $25 Billion for AI Capacity
