There is a particular species of investor who is convinced that the hard part is seeing the future. Get the thesis right, the reasoning goes, and the returns arrive as a matter of arithmetic. It is a flattering idea, because it locates the difficulty in the one place the ambitious investor is confident of his own superiority: insight.

The last week of July offered an expensive correction to that view.

Leopold Aschenbrenner, formerly of OpenAI, author of a 165 page essay in 2024 that argued more capable artificial intelligence would drive enormous demand for semiconductors, memory, data centres and power, was correct. Broadly, structurally, and in a way that has made a great many people a great deal of money, he was correct. Chips did run hot. Memory did tighten. Power did become the bottleneck. The trade worked.

His fund still lost roughly three quarters of its assets in a matter of days.

Situational Awareness, launched in late 2024 with reported seed capital of $225 million, grew to something in the order of $45 billion at its peak. It reportedly returned 439 per cent net in the first half of 2026 and more than 1,000 per cent since inception. Then the semiconductor complex pulled back, the margin calls arrived, and the prime brokers did what prime brokers do. The entire leveraged public equity book, including positions in names such as SK Hynix and CoreWeave, was sold to Citadel at a discount. Assets fell to around $10 billion. Reported leverage was in the region of four times.

Being right, it turns out, is roughly half the job. The other half is being solvent on the day you are proved right.

What the market is focused on

The commentary since has largely divided into two camps, both of them slightly beside the point.

The first camp reads the collapse as a verdict on artificial intelligence. The bubble has popped, the emperor is unclothed, and the sensible investor should now retreat to tinned goods and dividend stocks. This is the same reflex that declared the internet finished in 2001, shortly before it ate the global economy.

The second camp reads it as a story about a 25 year old with no risk management scar tissue, which is closer to true but still too comfortable. It permits everyone else to feel professionally superior and change nothing.

Neither camp is looking at the mechanism. And the mechanism is where the lesson lives.

What actually happened

Situational Awareness ran what had become the most crowded expression of the AI trade: long the hardware, short the software. Long the physical bottleneck, the chips and memory and infrastructure that AI demand cannot proceed without. Short the application layer companies whose economics AI might quietly dismantle.

It is, as a thesis, defensible. Arguably it is right. We have written before about the point at which every digital revolution discovers concrete, copper, steel and lead times.

The problem is not the direction. The problem is that a correct trade held at four times leverage is not the same instrument as a correct trade held unlevered. It is a different asset with a different risk profile and a different failure mode. Unlevered, a 30 per cent drawdown in semiconductors is an uncomfortable quarter and an opportunity to add. Levered four times, the same move is an existential event, and the decision about whether you survive it is taken by someone else, at a time of their choosing, at a price you do not negotiate.

That is the part worth sitting with. In a forced liquidation, the investor stops being an investor. Judgement, conviction, time horizon, the entire apparatus of thinking that justifies an active fee, all of it becomes irrelevant. A risk officer at a bank presses a button. The counterparties on the other side, in this case Citadel and reportedly Millennium, buy the assets at discounts of 20 to 50 per cent and book the difference.

This is not a market failure. It is the market working precisely as designed, transferring assets from the person who could no longer hold them to the person who could.

The second act is more interesting than the collapse

Here is where it becomes genuinely instructive.

Despite losing most of its assets, the fund is reported to be up something like 80 per cent for the year. The reason is that roughly a quarter of the portfolio sat in private holdings, dominated by a large stake in Anthropic, which has reportedly been marked up around 620 per cent. That single private mark is said to have offset the catastrophic public equity losses.

Read that again slowly.

The publicly listed portion of the portfolio was marked every second of every trading day, and those marks triggered the margin calls that destroyed the fund. The private portion is marked when a financing round happens, which is to say, when someone chooses to mark it.

One book was liquidated by its own transparency. The other book was rescued, on paper, by its opacity.

We do not say this to imply anything improper. Private valuation practice is what it is, and a large funding round is a real transaction with real capital behind it. But an investor looking at a headline return of 80 per cent and an investor looking at a forced fire sale of the entire public book are looking at the same fund. Only one of those pictures involved a price anyone was compelled to accept.

There is a broader point here that extends well beyond one fund. A great deal of capital has migrated in recent years toward assets that are marked infrequently, and a portion of that migration has been sold on the basis of lower volatility. Lower observed volatility is not the same thing as lower risk. It is often the same risk with the reporting turned down.

Why this matters beyond one blow-up

The AI infrastructure trade is not discredited by this episode. The demand for power, memory, cooling, transmission and land has not changed because one leveraged fund was liquidated. If anything, the businesses at the physical end of the chain remain the least speculative way to hold the theme, because their revenue arrives under contract and their assets exist in the ground rather than in a forecast.

What the episode does tell us is that the trade had become crowded enough, and levered enough, that a moderate pullback in semiconductors could produce a violent unwind. Crowding is not visible in a company’s accounts. It is visible in how the market behaves on a bad day. Late July was a bad day, and the behaviour was informative.

For long term investors, the useful reframing is this. The question is rarely whether a theme is real. Most of the popular themes are real. The question is what price is being paid, what structure is being used to hold it, and whether the holder can survive the interval between being right and being proved right. That interval is where most capital is lost, and it is almost never shortened by conviction.

Australian investors are not immune to this simply because the ASX has no meaningful AI hardware exposure of its own. Superannuation funds, global equity allocations and passive index holdings carry the exposure regardless. The transmission is indirect, but it is present.

Risks and counterpoints

Several things could make this analysis look foolish.

The pullback may already be over. Semiconductor names have recovered from sharper falls than this within a quarter, and a fund being liquidated is frequently a sign that the selling is finished rather than beginning. Forced sellers mark lows more often than they mark tops.

The reported figures deserve scepticism. Peak assets have been reported variously at $45 billion and around $20 billion, leverage at four times and at 400 per cent, and the Anthropic mark up at 620 per cent. These are sourced from anonymous briefings and unaudited estimates. The direction of the story is not in dispute. The decimal places very much are.

There is also a risk in drawing too tidy a moral. Plenty of leveraged funds have been correct and survived. The counterfactual in which semiconductors rose 15 per cent in July instead of falling produces a fund manager on magazine covers rather than in obituaries, and the underlying decisions would have been identical. Survivorship shapes which lessons get taught.

Finally, treating this as a signal to reduce AI exposure generally is its own error. A leverage accident is a leverage accident. It is not an earnings downgrade.

TAMIM Takeaway

A fund built on a correct thesis about artificial intelligence lost most of its assets in days, not because the thesis failed but because four times leverage removed the option to wait. The public book was liquidated by its own daily marks. The private book was cushioned, at least on paper, by marks that only move when someone chooses to move them.

For long term investors, three things are worth carrying away. First, the difficulty in investing is rarely identifying the trend. It is holding the position through the part where you look wrong. Second, leverage does not amplify returns so much as it transfers the decision about when you sell to a stranger. Third, an asset that is not marked often is not thereby less risky, it is merely quieter.

The AI infrastructure theme remains intact. Power, grid, cooling and memory are still bottlenecks with contracted revenue behind them. What changed last week was not the theme. It was a reminder that surviving long enough to collect on a good idea is the whole game, and that the market charges its highest fees to those who forget it.