Every time someone makes a lot of money in the US stock market, the first thing the onlookers do is always the same: look through his position report to find the next stock to buy.

The most frequently read report recently belongs to a 24-year-old German named Leopold Aschenbrenner.

In March this year, domestic media focused on him, with similar titles; for example, the genius who was fired from OpenAI wrote a 165-page paper to predict the trend of AI, opened a hedge fund, and managed US$5.5 billion...

But tags are just tags. What’s really striking about this fund is that it doesn’t buy NVIDIA, it doesn’t buy OpenAI, and it doesn’t buy any companies that make AI models. It only buys things that AI cannot live without, including power generation, chip manufacturing, optical communications, data centers...

In the words of his own paper,The bottleneck of AI is not the algorithm, but electricity and computing power. The entire fund is betting that this statement is correct.

The investment blogger on social media calls him "the son of the US stock market version in the AI era", or the AI version of the stock god Buffett. Recently, this title has been revealed again, because the degree of his correct bets has begun to become a bit outrageous.

According to data released by the copywriting platform Autopilot on May 1, the investment portfolio that simulated his holdings rose by 61% in two months. Based on this calculation, the size of his fund is approaching US$9 billion.

Where did you make the money? Mainly relying on two heavyweight stocks. Bloom Energy, a fuel cell company that provides off-grid power to AI data centers, is up 239% year to date.

According to a public position report at the end of last year, he holds $875 million in stocks and options in the company, and its market value has now swelled to nearly $3 billion.

Also Intel. The same position report shows that he bought 20.2 million Intel call options in the first quarter of 2025. At that time, Intel's stock price was around $20, and the mainstream judgment on Wall Street believed that Intel was not a good company.

Last week, Intel rose to $113, a 25-year high. It has nearly quintupled in less than a year, and this young man’s option return multiple is much more exaggerated than that of stocks.

I can understand the impulse of the onlookers. The American investment website Motley Fool published four articles a day dismantling his positions, and the overseas Reddit investment forum was discussing whether to copy his work. Everyone is trying to find the next Intel from his position report.

But you should know that position reports are generally delayed by 45 days. By the time you see what he bought, the market is already halfway gone.

More importantly, even if you know his position in real time, you can't replicate the reason why he continues to bet.

Circle is the greatest Alpha

First of all, the most amazing thing about Leopold Aschenbrenner is the paper he wrote on AI in 2024, which almost predicted the current development direction and investment context of AI.

The core argument can be summarized in one sentence: the training computing power of AI models increases by about half an order of magnitude every year. At this rate, general artificial intelligence (AGI) with capabilities close to humans will appear around 2027.

But to maintain this growth rate, the key constraints are not at the algorithm level, but at the power, chip production capacity and physical space. The power consumption of a single training cluster jumps from megawatts to gigawatts, approaching the output of a large nuclear power plant.

This is the underlying logic of his entire fund. The speed of AI development is determined by physical bottlenecks, so you should invest in the bottlenecks themselves.

This judgment sounds like a conclusion deduced by a smart man after doing a lot of homework in the study; but in fact, I think it was the circle that made him form this judgment.

Before writing his thesis, he worked in OpenAI's Superalignment team for a year. This team specializes in how to control AI that is smarter than humans and reports directly to chief scientist Ilya Sutskever.

In that year, what he saw was the internal training plan, the actual computing power consumption, and the specific needs of the next generation model for power and chips. When he wrote the judgment of "gigawatt-level electricity consumption" in his paper, he may have based it on the internal road map in the laboratory.

He was fired from OpenAI in April 2024. The trigger was that he wrote an internal memo to the OpenAI board of directors, warning that the company's security measures were insufficient and that it might face the risk of infiltration by foreign intelligence agencies.

The memo sparked tensions between management and the board of directors, and OpenAI later fired him for "leaking information."

Two months later, the paper was published. This paper is not so much an independent study as it is a public version of his knowledge within OpenAI.

The AI paper solves the problem of "what direction to look in". But when investing, just knowing the direction is not enough.

AI needs more power, and many analysts are saying this in 2024. What’s really valuable is timing and positioning. For example, do you dare to invest 20 million call options when Intel’s stock price is $20?

This kind of confidence comes not only from believing in the general trend of AI, but also from knowing specifically which company is signing a large power purchase contract, which data center is expanding, and what the magnitude of the demand is.

And the investors of Situational Awareness, the fund founded by Leopold Aschenbrenner, happen to be sitting in the front row of these decisions.

The LPs of this fund include the two founders of Stripe. The company handles the payment flow of most of the technology companies in Silicon Valley and can directly feel the acceleration of infrastructure spending;

Another investor is Nat Friedman, former GitHub CEO and current Meta AI product leader, who is involved in computing power procurement decisions every day.

What they bring to the fund, in addition to initial capital, is a continuously updated information pipeline.

In addition, the research director of his fund is also a key role in this chain. Carl Shulman, a veteran in the field of AI security, previously worked at Peter Thiel’s hedge fund Clarium Capital, where he was responsible for transforming the knowledge of the AI ​​​​circle into executable trading strategies.

In his holdings, there is another crypto corner that is easily overlooked.

The position report at the end of last year shows that he established new positions in CleanSpark and Bitfarms, both of which are Bitcoin mining companies that are transforming BTC mining facilities into AI computing power centers.

Crypto mining farms naturally have large-scale power access and cooling systems, which happen to be the scarcest resources in AI data centers.

Interestingly, he is no stranger to the encryption industry. In 2022, he worked for nine months at the Future Fund, the FTX charity fund founded by SBF, and left just before the FTX thunderstorm.

Whether this experience directly affected his judgment on mining companies, outsiders have no way of knowing. But what can be confirmed is that he is one of the very few people who has had in-depth contact with the encryption industry and AI cutting-edge laboratories at the same time. This intersection itself is also a rare cognitive position and the possibility of human connection.

There is another detail, his fiancée Avital Balwit is the chief of staff to Anthropic CEO Dario Amodei. Anthropic is Claude’s parent company and OpenAI’s most direct competitor.

He worked at OpenAI and his fiancée is with the CEO of Anthropic. One of the two companies at the forefront of the AGI competition has practical experience, and the other has daily contact with them.

American Fortune magazine interviewed more than a dozen insiders who had contact with him last year and concluded that he is good at "packaging the ideas brewing in Silicon Valley labs into narratives."

The author thinks this statement is too polite. What he did was more direct, that was to place bets on the public market based on the knowledge gained from his private circle. The AI ​​paper sent out is the declassified version, and your own investment fund is the complete version.

A positive feedback loop that outsiders cannot enter

Looking back, Leopold Aschenbrenner's fund chose a less common structure.

Most funds in the AI field take the venture capital route, investing in early-stage companies and betting on who can become the next OpenAI. He didn't take this path. According to Fortune, he explicitly rejected the VC model when he founded the fund on the grounds that the impact of AGI was too great. Investment judgment can only be fully expressed in the most liquid public market.

This choice itself exposed a consensus in his circle: the biggest investment opportunities in the AI era may be hidden in old companies that already have physical infrastructure.

It can be a fuel cell company with ready power access, a chip giant with a foundry production line, or a Bitcoin mining company with mines and cooling systems. These companies have been listed for many years and have good liquidity, but most analysts are still pricing them using the old valuation framework, and have not seriously incorporated the variable of "AI infrastructure rigid need" into the model.

This is his arbitrage space.

People in the circle already know the pace and scale of AI infrastructure expansion, and the public market is still pricing using old logic. The price difference in the middle is the source of profit.

This information advantage also has another characteristic: it is self-reinforcing.

The better the fund returns, the more people at the core level of the industry are willing to become LPs. The more LPs there are, the more intensive decision-making information the fund has access to. The denser the information, the more accurate your bets will be. This is a positive feedback loop, and the barriers to entry for outsiders will only get higher and higher.

Of course, this cycle also has a fragile side. Highly concentrated holdings coupled with significant leverage mean that the entire fund is extremely reliant on a single narrative. As long as the premise of "AI infrastructure continues to expand" remains true, everything will go smoothly.

But if the pace of AI development slows down, or the energy bottleneck is bypassed by some technological breakthrough, the retracement speed of concentrated positions will be much faster than the speed of position opening. He bet not only on the direction, but also on the rhythm. Once the rhythm is misaligned, the consensus in the circle may become a collective blind spot.

Back to the original question.

Everyone is studying his positions and trying to copy his operations. But behind the stock god-level returns, there are structural conditions.

The paper is public, the position report is public, and his investment logic is also clearly explained in podcasts and interviews. But even if you fully understood every one of his judgments, you couldn't replicate the position he was in when he made them.

Positions can be traced back, and the profits are enviable, but the source of knowledge cannot be shared. This is probably the most expensive asymmetry in this era.