Being Right About AI Is Not Enough

The most painful investment mistakes are not always made by people who fail to see the future.

Sometimes they are made by people who see it too early, size the bet too aggressively, and lose the position before the future arrives.

That is the lesson sitting behind two of the strangest AI investment stories of the past five years: Leopold Aschenbrenner’s Situational Awareness fund and Sam Bankman-Fried’s FTX estate. The characters could hardly be more different. One is a former OpenAI researcher who turned an AI infrastructure thesis into one of the fastest-growing funds in recent memory. The other is a convicted fraudster whose crypto empire collapsed in 2022.

Yet both stories bend toward the same point.

Both identified Anthropic before the market fully understood what frontier AI would become. Both were directionally right. Both lost access to key positions before the thesis finished paying off.

Aschenbrenner’s route to finance began inside AI research. After leaving OpenAI, he published a series of essays arguing that the next phase of artificial intelligence would require a vast buildout of compute, memory, semiconductors, energy and data-centre infrastructure. He then built a hedge fund, Situational Awareness LP, around that thesis.

For a time, the results were extraordinary. The fund reportedly grew from $225 million to as much as $45 billion in under two years, helped by large positions in AI infrastructure beneficiaries such as compute providers, energy companies, memory makers and Bitcoin miners repositioning themselves as data-centre operators.

The thesis was not frivolous. AI is increasingly a physical infrastructure trade. Models need GPUs. GPUs need high bandwidth memory. Data centres need power, cooling, land and financing. As demand for compute exploded, the companies enabling that buildout became some of the market’s most crowded winners.

The problem was not the idea. It was the structure.

Situational Awareness reportedly ran its public equity book at roughly four times leverage. That magnified gains on the way up, but it also left the fund vulnerable to a sharp reversal. In July 2026, AI-related momentum stocks sold off hard. Semiconductor and technology indices fell sharply from their peaks. With losses compounding through leverage, prime brokers issued margin calls. Fresh capital did not arrive in time. The fund sold much of its public equity portfolio to Citadel.

Aschenbrenner reportedly retained private holdings, including a large Anthropic stake. But the public equity liquidation showed the brutal mechanics of leverage. A lender does not wait for a five-year thesis to mature. A margin desk looks at today’s collateral.

Four years earlier, FTX had made an eerily prescient AI bet of its own. In 2021, FTX and Alameda Research invested $500 million for an estimated 8% stake in Anthropic, long before the company became one of the central players in frontier AI. Alameda also invested early in Anysphere, the company behind Cursor, the AI coding tool that later became one of software’s most watched growth stories.

Those were excellent investment calls.

They were held inside a disastrous structure.

FTX collapsed in November 2022 after customer funds were misused and commingled with Alameda’s trading operations. Bankman-Fried was later convicted and sentenced to prison. The bankruptcy estate that took control of FTX’s assets had a different job from a venture investor. It had to recover money for creditors on a legal timetable, not hold speculative positions for maximum future upside.

The estate sold the Anthropic stake for $1.3 billion in 2024. It also sold the Cursor-linked stake before its later valuation became clear. By 2026, those positions were estimated to be worth far more than the estate received.

The investment judgment was not the failing. The holding structure was.

That is the uncomfortable overlap between Aschenbrenner and FTX. One failed through leverage. The other failed through fraud and bankruptcy. Those are not morally equivalent. But in both cases, the holder of a correct AI position lost control of the exit timeline.

A position sold by a margin desk and a position sold by a bankruptcy estate have the same practical flaw: neither gets to wait for the thesis.

This matters because AI and crypto investors often confuse conviction with durability. They believe a strong thesis justifies a larger position. Sometimes it does. More often, it demands a better structure.

A correct idea can still fall 50%. A transformative company can remain illiquid for years. A crowded trade can unwind before fundamentals deteriorate. A token, stock or fund can be right about the future and still fail in the present.

The old Kelly criterion, used by gamblers and investors to size bets, captures the principle neatly. The size of a position should be related not only to the expected payoff, but to the probability and severity of loss. Bet too much, even with an edge, and ordinary volatility can become ruin.

That is the real lesson of the AI conviction trade.

Spotting the next Anthropic is rare. Holding it through volatility, financing pressure, legal risk, liquidity stress and human overconfidence may be rarer.

The future does reward vision.

But only if the position survives long enough to collect.