The Truth Behind the Fall of the 25-Year-Old AI Stock Market Prodigy: Not Just a Leverage Death Story

On July 29th to 30th, Citadel, through a large block trade, acquired the entire public stock book of Situational Awareness. Both long and short positions were liquidated overnight.

Now, the fund led by the "AI Stock God" Leopold only holds its private positions, along with a small amount still available for secondary trading. Among these private positions, Anthropic, valued at around $5 billion, has become the fund's most valuable holding.

From a 400% gain in six months to almost zero in secondary positions, Situational Awareness was returned in one night to what it was originally meant to be—a venture capital fund.

The market buzzed with activity, and even Chen Bin came to a conclusion: stay away from leverage. It seems like a classic story—an old tale of a trader unable to control their greed, ultimately perishing due to high leverage.

But that's not the truth.

In Wall Street, this Silicon Valley elite created a miracle in the Silicon Valley way, initiating an AI bull market through a paper and then, in the Silicon Valley manner, sticking to his views, only to end up in ruins.

Position Branded with the "Silicon Valley Stamp"

Leopold's legend stems from the collision of his tech elite background with his Wall Street trader identity. A paper titled "Situational Awareness" was seen as the starting point of this AI bull market and was also the name he used when he founded the fund in 2024.

The entire position structure of this fund seemed like an expression of a viewpoint established based on the "Situational Awareness" paper, leaving a deep, Silicon Valley-style "growth stamp."

In June 2024, Leopold published a 165-page long essay titled "Situational Awareness." The core judgment was that by around 2027, models would be capable of replacing AI researchers and engineers. This article later became a public text for Silicon Valley discussions on AGI and was one of the key narrative engines of this AI market.

He infused his faith into his positions, making them a tool to express his views. But the Wall Street game wasn't played that way.

A while back, Situational Awareness disclosed that they held several put options. These included $20.4 billion in semiconductor ETF put options, $15.7 billion in NVIDIA, as well as Oracle, Broadcom, AMD, and further down the line Micron, TSMC, ASML, and Intel.

Many people thought they were hedging, but in reality, that was not the case.

This structure directly stemmed from their whitepaper. Leopold wrote in "Situational Awareness" that chips are likely a smaller constraint than power. He believed that the real bottleneck is electricity, the generator and turbines, HBM, and advanced packaging, rather than the GPU itself. He even calculated the number of shale gas drilling rigs, with one rig being able to drill three wells per month, meaning forty rigs could power a gigawatt cluster in one year.

Following this assessment, both the long and short sides made sense.

They were long on companies involved in chip power supply and chip integration. Bloom Energy focused on fuel cell power generation, CoreWeave, IREN, Core Scientific were involved in data center and mining farm transformation, and T1 Energy dealt with power infrastructure.

They were short on the companies manufacturing chips. The market's valuation of this chain was based on the premise that "chips are the scarcest thing." If the scarcest resources are actually electricity and space, then this premise falls apart. So, their short position was not based on the semiconductor's fundamentals but on the scarcity premium of semiconductors.

This was not actually about downside protection but about using their own expectations to do Beta hedging. Their long and short structure conveyed a viewpoint: the dividend of computing power, semiconductors, and other AI infrastructure has already been priced in, while Neocloud and energy have not. Therefore, in the future, the latter will see a greater increase.

This is actually Silicon Valley's growth logic, aiming for higher Alpha rather than correct downside protection. So these past few days, the market began to teach the "AI Stock Godfather" a lesson.

The semiconductor sector did indeed decline. During that week, the semiconductor ETF dropped by 5.63%, AMD by 12.9%, Intel by 6.52%, and NVIDIA by 4.75%. Looking solely at the short side, they did make a profit.

However, the things they were long on dropped even harder.

In July, their core holdings experienced downturns ranging from 27% to 54%, with their top position, Bloom Energy, retracing by about 43% in just one month.

He never bet on semiconductors dropping, but on electricity and data centers rising more than semiconductors. The answer given in July was exactly the opposite. The direction was right, but the relative relationship completely flipped. The little money made from hedging couldn't fill the hole in the long position. And under four times leverage, not being able to fill it was no longer a retracement issue.

In Silicon Valley, you can be wrong nine out of ten times, as long as one investment increases a hundredfold. But on Wall Street, being wrong just once can wipe you out completely.

Silicon Valley's 2027 Hasn't Arrived Yet, Wall Street's Wednesday Has

When Silicon Valley talks about the future, they often mention specific timelines. Leopold, in "Situational Awareness," was specific up to the year 2027.

From GPT-2 to GPT-4 took four years, and he recalculated the next four years: computing power will continue to increase, algorithms will keep saving on computing power. Chatbots will acquire tools, plans, and the ability to take action. Adding up the first two, he believes that effective computing power could increase by about a hundred thousand times. By 2027, models will be able to do the work of AI researchers and engineers without needing to believe in science fiction first.

The front half of this chart is more urgent. In his paper, he wrote that from 2025 to 2026, machines will surpass ordinary college graduates. Looking further ahead, giant training clusters will cost upwards of billions of dollars, and items like power, land, permits, and data center construction will slow down compared to the chips themselves. Chips are no longer a monolithic block. Advanced packaging and HBM memory will be the initial bottlenecks, and power will block the road further down the line.

This timeline was integrated into his positions.

Leopold's investments were never just in "AI"; he was chasing the milestones that the computing power curve in papers had to pass through. GPUs, memory, advanced packaging, data centers, power. Those companies may seem to belong to different sectors, but in reality, they have all charged tolls for the same thing. On the other hand, he doesn't quite believe that software companies that haven't had time to turn AI into profit deserve their original valuations. Leopold pushed his positions up layer by layer as if preparing ahead for 2027.

That's also why every reshuffling of storage, computing power, and software makes him uneasy.

The logic in the papers is lengthy, but bottlenecks in reality take turns to be emphasized. Today the market focuses on HBM, tomorrow on GPUs, and the day after it's discovered that capital expenditures are squeezing cloud providers' profits first. Each phase may only last a few weeks, but it's enough to chop up a large and complete future into many fragmented losses.

Compared to Silicon Valley, Wall Street's calendar is much thinner.

It only has the next earnings report, the next interest rate cut, the next supply chain meeting, and the last few hours before market close today. A fund manager isn't unaware of the importance of 2027. He just has to get through the phone call tonight first. Client redemptions won't wait for the model's capability curve. Stop-loss levels also lack historical context.

When these two calendars collide, accidents are bound to happen.

Leopold's advantage is to view the future as a line. The market's skill is to constantly plant flags along this line. It doesn't need to deny the endpoint, as long as it sets up several toll booths in between, it can allow the person heading in the right direction to park their car first.

A Moment's Dividend

Silicon Valley's patience for returns has always been longer than that of the trading floor.

Money invested by a venture capitalist that only takes shape five years later is not considered a mishap. No one is surprised when it becomes a public company after ten years. Even twenty years later, Silicon Valley is still willing to assess what a company has truly left behind. A good project must first withstand the tests of product development, talent acquisition, and cash flow before valuation can celebrate it.

Leopold has compressed this calendar significantly. He doesn't wait for ten years. In 2024, he writes 2027, leaving himself only four years in the future. In Silicon Valley terms, this is almost considered sprinting.

He has tried using an even shorter calendar, but he still hasn't switched to the shortest one.

In the past, a bull market cycle would leave some time for learning. But today's Wall Street only gives you half a year. One quarter is for belief, the next quarter expects to see the money. If the money is not visible, the price will make the judgment for you.

Back in the 90s when the Internet was just beginning to heat up, news had to make its rounds between magazines, TV, and brokerage reports before money caught up. Bubbles still existed, but after making a mistake, there was occasionally a chance to reflect. There is no such gap now. While 2027 is still in the distance, the market is already trying to squeeze it into the second half of the year.

Just as a long article is published, podcasts immediately start reading the harshest lines from it. After a while, someone chops it into a forty-second video. And then, someone on a trading app has already translated it into code, options, and a triple-leveraged ETF. Before the viewpoints even become outdated, positions are already full.

“Situational Awareness” holds a delicate position in this round of the market. It is certainly not the only driving force. Without chip orders, without cloud provider spending, without real model advancements, an article wouldn't ignite such a fire. But it has given a voice to many scattered excitements. Suddenly, people know why they should be excited and where to invest their money.

This ability used to be rare, but today, it's too easy to replicate. Screens will find readers for an article and buyers for a stock. By the time search interest spikes, the trading volume often beats it to the punch. Many people didn't even finish reading that report, let alone understand its reasoning. They only see one thing: everyone else seems to be buying.

That's already enough to send a stock soaring.

The barrier to entry for buying is so low it's almost invisible. In the past, to research a company, you would have to be willing to open an account, pick a stock, and acknowledge that you might be wrong. Later, you could just buy an industry index. Then, one click could amplify your judgment two or threefold. Believing in a story became as simple as a finger swipe.

This eroded the consensus's wear and tear, so it arrived exceptionally quickly.

When prices rise, everyone will find themselves a beautiful set of reasons. The models become stronger. The demand grows larger. The world is forever changed. But when prices fall, the reasons remain, just in reverse order. The models are strong, so costs will decrease. The demand is high, so competition will be fierce. The world indeed is different, but it doesn't necessarily only reward you holding those few stocks.

This whole thing is useful for FOMO and equally useful for fear.

In the past, it took a long time for fear to spread; now, you don't even need to leave a building. An incomprehensible earnings call, a context-stripped image, a viral short video can make everyone simultaneously remember why they should run. ETFs will gather the runners together. Leverage will urge them to run faster. By the time everyone is finally willing to carefully read that report, the numbers on the screen have already read it for them.

There are more and more Stock Gods, but fewer who can stay at the table

A side effect of a shorter cycle is the rising output of Stock Gods.

The market likes to deify people because deification is easy. It doesn't have to understand a person's entire judgment, just pick out a screenshot from their most beautiful trades. If a young person hits on AI, chips, or a software stock, and their account curve turns upward, their name will spread in the group. Some see him as a teacher. Some fear they can't keep up. And some have already begun to write about his next victory.

Communication tools have sped this up, and buy-in tools have made it cheaper. Deification no longer needs a bull market cycle; a single quarter is enough.

But the same things that send people up will also bring them down.

Serenity made 4502% in half a year, with a 49% drawdown in July. Leopold achieved over 1000% return from inception to establishment in less than two years; from over 1000% return to the entire public ledger being sold off took just one month.

They are not greedier than anyone else, and may even be more diligent. Many retail investing legends are like this, spending a long time reading materials, memorizing company names, waiting for a moment when they can finally speak up. The market rarely gives them a chance to be right, so when they are, it looks like a stroke of genius.

This is also why Buffett repeatedly appears like an antique. He talks about margin of safety, discusses return on capital, and likes to keep money for a long time. In a market where consensus can be rewritten in three minutes, these words sound a bit like asking a high-speed train to wait for a porter. In the first half of this year, there were countless people who outperformed him. As of July 29, he lagged behind the S&P 500 by six point six percentage points, still holding $397.4 billion in cash.

Leopold is more qualified than most to believe in his own deductions. He has spent time, written arguments, and knows how those dull components in data centers are connected to national competition and corporate profits.

But the more someone is like this, the harder it is for them to accept that sometimes the market simply doesn't want to see the endgame.

And every person who goes from hero to zero in two years is conducting an experiment for Buffett. The shorter the cycle, the more times this question is asked. So the more investing legends there are, the higher the value of Buffett. Not because he has done something new, but because others have tried all the faster routes for him.

After Leopold exits, the market will continue to tell stories to AI. It will find new names, new investing legends.

The public market positions taken by Citadel will become a set of new numbers in other people's models. While the private shares in Leopold's hands will still remain silent.

They are waiting for the companies to grow larger.