Situational Awareness: Being Right and Wrong in the AI Trade

Last Thursday, before the US market open, Situational Awareness LP sold the bulk of its public equity book to Citadel in a single block. Longs and shorts together, in one trade, at a discount. Millennium and Jane Street had also bid, Jane Street being one of the fund’s own investors. Just six days earlier the fund had written to investors reporting a net return of 439% for the first half of 2026, describing the sell-off as one of the best buying opportunities since early last year, and inviting fresh money from 1 August.

The fund launched in late 2024 with about $225m. Assets were reported north of $20bn through June, and CNBC put the peak at $45bn earlier in July. What is left is roughly $10bn, around half of which is a single private position in Anthropic.

The irony is that, despite all the turmoil, the fund’s investment thesis remains correct today.

What the fund got right

Its founder published “Situational Awareness: The Decade Ahead” in June 2024, a few months after leaving OpenAI’s superalignment team. The essay argued that machine intelligence scales with compute, that compute scales with capital and electricity, and that the constraint would end up being physical. Power, land, memory, cooling, interconnection.

That call was early and it was audacious. In mid-2024 the consensus AI trade was Nvidia. Underwriting gas turbines, fuel cells, NAND makers and data centre operators as artificial intelligence exposure was a minority position, and the essay carried enough authority to raise $225m from Patrick and John Collison, Nat Friedman and Daniel Gross. Jane Street came later, which is a rare thing for a firm that normally has no reason to back an outside manager.

Then the thesis kept arriving. The fund returned 47% in the first half of 2025 against roughly 6% for the S&P 500. Since then, hyperscaler capital expenditure commitments for 2026 reached around $700bn, up 77%. The DRAM supply-demand gap opened to its widest in fifteen years, contract prices rose more than 50% in a quarter, and SK Hynix printed a first-quarter operating margin above 70% with the year sold out. Interconnection queues lengthened. Every physical bottleneck named in the essay tightened ahead of the schedule the essay had set for it.

By the end of June the fund was up 439% for the year and more than 1,000% since inception. This was not luck. The success was built on professional knowledge of the field and an accurate roadmap of the AI build-out for the years to come.

What the book actually was

The 13F for the first quarter of 2026, filed 18 May, disclosed $13.68bn of reportable securities, more than double the $5.52bn three months earlier. The long equity book inside it ran to about $3.86bn, with the top five positions accounting for more than three quarters of that: Bloom Energy at 22.8%, SanDisk 18.8%, CoreWeave 14.4%, IREN 10.4%, Core Scientific 10.1%. Around the edges sat bitcoin miners, Nebius, Micron, AMD, Oracle and SK Hynix.

Power, storage, GPU rental and mining capacity all respond to one variable. Each is priced off the expected growth rate of AI infrastructure spending. Five tickers, one bet, and in Core Scientific’s case around 8% of the float.

The filing also showed $8.46bn in put options, about 62% of the reportable total, including more than $2bn against the VanEck semiconductor ETF, more than $1.5bn against Nvidia, and roughly $1bn each against Oracle, Broadcom and AMD, with smaller lines down through Micron, TSMC, ASML and Intel. That figure is the notional value of the underlying shares, which is how 13F reports options. The premium actually at risk was a fraction of it, and so was the protection.

On paper that is a hedge. In practice it was mismatched, which renders it insufficient. This is because the longs were high-beta infrastructure with second-order exposure to credit and bitcoin. The puts were written on the lowest-beta part of the same complex. When the group sold off together, the longs fell two to three times as far as the instruments hedging them, so the payout covered a fraction of the loss it was sized against.

The rest of the risk was expressed in swaps, options and offshore lines, with gross exposure reported at roughly four times capital.

What July did

On 1 July, Bloomberg reported that Meta was building a cloud division to sell surplus compute. That is a direct threat to the neocloud business model, and CoreWeave and Nebius are neoclouds.

From there bearish sentiment compounded. More than $1 trillion came off chip stocks in the month. Nvidia lost $238bn of market value, SK Hynix $176bn, Samsung $173bn, Micron $113bn. The Nasdaq 100 fell about 10%. Korea’s Kospi fell roughly a third, and SK Hynix’s US-listed shares were down about 47% from their June peak within weeks. Nebius fell 43% on the month, CoreWeave 36%, with the credit default swap market pricing CoreWeave at close to a 50% five-year default probability by 29 July.

At four times gross, a 30% drawdown in the long book is more than the whole equity. The hedges did not cover it, because semiconductor and Nvidia puts do not pay enough when the loss is in fuel cells, miners and GPU landlords. The software shorts, Adobe among them, went up while everything else went down, which is what happens when money rotates out of the AI build-out and into things that had already been written off.

Bank of America, Goldman Sachs and JPMorgan were the prime brokers marking the book. By the end of the month they were showing the positions to buyers, and on 30 July Citadel took the bulk of it.

The three usual suspects

Set against the stellar figures, the causes of the downfall are ordinary. It took leverage, concentration, and a crowd.

Leverage decides who chooses when you sell. At four times, the choice moves to a margin department once the drawdown passes a certain depth, and no amount of conviction about AI-in-2027 buys back that authority.

The concentration was hidden behind a diversified-looking exposure report. Fuel cells, flash memory, GPU rental and bitcoin mining look like four industries. They are four beneficiaries of the same forward capex assumption, and they correlate to one in the only week that matters.

The crowd is the part we wrote about last week. In “The Crowded Kospi” we argued that the Korean memory trade had been correct on fundamentals and dangerous in construction, because the positioning had migrated into leveraged wrappers held by people who all faced the same margin call in the same window. When the trade is crowded, the holdings become a position in liquidity rather than in fundamentals.

In April we argued that the next crowded trade would be assembled faster than the old ones, because published reasoning now propagates through machines as well as people. The essay that gave the fund its name was the seed document for a great deal of capital that never met its author. And in “Where Does the AI Build-Out Top First” we made the point that a build-out this large tops in stages rather than on a day, which is precisely the environment in which a levered long of the entire complex bleeds without ever getting a clean signal to leave.

What is left

The fund is not shut. It keeps its private positions, and the largest is a stake in Anthropic reported at around $5bn, roughly half of remaining assets. Anthropic was last valued at $965bn in its Series H in May, filed confidentially on 1 June, and could list within months. Stakes in MatX and Fluidstack are still there. The fund has approached existing investors for capital, described as ad hoc rather than a coordinated raise.

If the essay is right about the coming decade, the private book is the part of the thesis still standing.

The analogy

Situational awareness is an aviation term. It means holding an accurate model of where you are, where the aircraft is going, and what the machine is doing, all at once. Its failure mode is called task saturation, and it does not present as confusion. It presents as complete focus on one instrument while the others drift out of tolerance.

It is the drift out of risk tolerance that proved the culprit for the book closure.