A Lack of Situational Awareness

Every desk and group chat was consumed by Leopold Aschenbrenner and the Situational Awareness blowout. Fair enough, as a spectacle, it is hard to beat. A former OpenAI researcher in his mid-twenties, no prior investment experience, raises $225 million in September 2024, turns it into roughly $45 billion by early July, up 439% net in the first half of this year alone. Then July happens- down about 67%, prime brokers calling for collateral, and by Wednesday the 29th the whole public equity book sold to Citadel in a single block before the open. Assets decreased from $45 billion to about $10 billion.

Roughly 4x leverage on the public portfolio, expressed as a long AI hardware and infrastructure position against short software. Core positions appear to have fallen on the order of 30–50% through July, set against a Philadelphia Semiconductor Index down 28.6% from its 22 June peak. The hedge was no hedge: the software names he was short (Adobe among them) moved sharply against him, so the short book became a second source of loss rather than an offset, and the collateral behind both legs went at once.

At 4x, a 20% drawdown in the underlying takes 80% of your equity. There is no recovery from that, no matter how right you eventually turn out to be. Whilst leverage may amplify returns, it also costs you time on the position. Unfortunately, often time is exactly the one thing you need.

The profile of the book is worth a look too. Of about $13.7 billion in reported US equity and options, around 62% by notional was in put options. That isn't a portfolio so much as a statement of certainty.

Much has been made of Griffin as the shadowy operator who engineered all of this, and we think that is, to put it plainly, nonsense. Millennium reviewed the book and passed. Jane Street reviewed it and passed. Citadel bought the lot, every long and every short, in a single block. Absent that bid, the portfolio is zeroed out into a thin tape the following morning, at a materially worse outcome for investors.

Markets are a blood sport. Positioning this badly wrong carries a price, and whoever has the balance sheet and the nerve to catch it when nobody else will is entitled to be paid for doing so.

The accountability question, however, remains, and few seem inclined to ask it. Who wrote the cheques? Professional allocators with fiduciary obligations and due diligence budgets committed tens of billions to somebody who had never managed money. Not managed it badly – just simply never managed it before. Managers with twenty-year records and the scar tissue to match struggle to raise a fraction of that. That says more about how capital is being allocated in 2026 than any post-mortem on the trade.

The same week all of that was unfolding, the Fed met and held rates at 3.50%–3.75%. The 30-year sold off anyway, up 11 basis points on Wednesday 29 July to 5.20% (its biggest single-day move in more than a year) before finishing at 5.28%, a level last seen in 2007.

Compared to recent market history, this isn't a direct inflation story. Real yields account for roughly 93% of it since mid-June, which makes this a story about term premium and supply rather than inflation expectations. Investors want to be paid more to fund the United States for thirty years, and that is a harder problem for the Fed than a breakeven problem would be.

The 2007 comparison flatters nobody either. Back then the 30-year sat at 5.28% with Fed funds at 5.25%, more or less level with each other. Today the long bond is running about 150 basis points above the policy rate, which is not a curve that believes the central bank is in control of the outcome.

The AI buildout has turned into a fixed-income event, and it feeds straight into that supply picture. The big five are running toward $750 billion of capex this year and something close to $1.2 trillion in 2027. Internal cash flow stopped covering that a while ago, so they borrow: Alphabet, Amazon, Meta and Oracle issued about $194 billion through 7 July, against $108 billion across the whole of 2025.

And the market is starting to gag on it. Of 91 hyperscaler bonds issued this year, 78 were trading above their issue yield by 28 July, and Amazon's $25 billion deal was enough on its own to push its existing 30-year paper 20 basis points wider. So while equities are still arguing about whether this capex will earn its cost, credit has already answered it, in basis points.

Being right about direction is worth nothing if your structure can't survive the path. Aschenbrenner may well be vindicated on AI infrastructure over a decade, but he won't be there to collect, because 4x leverage turned a drawdown into a liquidation. Sizing a position properly buys the time for a thesis to work.

Moreover, the AI complex has commanded almost all the available attention for two years, and attention is finite. Everything outside that frame has been developing unobserved.

If there is any concluding observation to take away from here, it is this: the US long bond just printed its highest yield in nineteen years, in the face of a Fed on hold, driven by real yields and supply rather than inflation expectations. That is a repricing of the risk-free rate that sits underneath every asset on earth, including the AI trade itself. If long yields keep grinding higher on real rates and supply, the curve will have hiked for them.

Portfolio figures reference SEC Form 13F-HR (Q1 2026, as of 31 March 2026) and a 13G filing for NBIS (May 2026); positions may have changed between filing dates and 30 July.