Leopold Aschenbrenner, regarded as one of the world's most aggressive AI investors, has established nominal short positions of approximately $9 billion in public markets against NVIDIA, ASML, and Oracle, while reallocating capital towards deeper layers of AI infrastructure and model assets such as power, memory, data center networks, and companies like Anthropic.
The two hosts believe this does not signal the bursting of the AI bubble, but rather indicates a rotation within infrastructure trades from a "chip-first" approach to one prioritizing "energy, networks, and data center construction." Especially following NVIDIA's recent $25 billion bond financing and the soaring valuation of Anthropic, the market implications of this judgment are rapidly amplifying.
Key Takeaways
Leopold's Core Trading Logic
· "The classic 'selling picks and shovels' trade in AI has become too crowded, and Leopold's recent portfolio changes convey precisely this signal."
· "His judgment is not that AI infrastructure has peaked, but that certain tiers within the infrastructure stack, particularly semiconductors and traditional popular stocks, have become excessively crowded."
· "If the question becomes where the money will flow next, there are two answers. The first and most direct is flow into the next genuine infrastructure bottlenecks, namely power, memory, and data center network segments. The second answer is that mysterious investment exposed just a few weeks ago."
· "His bets have always been very infrastructure-oriented, investing in both these optical companies and power-related companies."
· "If he is cautious on NVIDIA, then capital will move to areas like power and memory; at the same time, he also wants to invest directly in the 'mine' itself, rather than continuing to only buy the 'shovels.' Anthropic is his favorite 'mine.'"
The Signal from NVIDIA's Financing
· "The issue isn't whether NVIDIA will continue to make money, but why a company with extremely high profit margins and already large cash reserves would borrow an additional $25 billion externally."
· "If a company simultaneously engages in massive stock buybacks and significantly increases dividends while also borrowing money in the same month, it's clearly not borrowing due to a cash shortage. A more reasonable explanation is that this is cheap capital, and the financing methods within this AI cycle are undergoing a slight shift."
The Next Wave of AI Infrastructure Benefits
· "The real bottlenecks are no longer just GPUs, but power, memory, data center networks, and the actual capability to build these things."
· "Even if you raise endless amounts of money, you cannot build data centers fast enough, expand memory chip capacity sufficiently, or immediately expand the power grid, transmission lines, and related infrastructure. There aren't enough people on the ground, and permitting, regulation, and various procedures are also hindering progress."
· "Whoever can build the data centers will take the profits."
Optical Modules, Copper, and Fiber Optics
· "As GPU clusters grow larger, copper wires will get hotter, energy losses will increase, and efficiency will suffer. In this scenario, fiber optics become the next upgrade direction."
· "For many high-bandwidth, short-distance transmission scenarios, copper is almost the only material everyone truly wants to use. Only when it becomes unsuitable, such as over long distances or with excessive heat, do they switch to fiber optics. Therefore, the combined market demand for copper and fiber optics is currently very strong."
· "The recent strength in copper futures is essentially because everyone needs it; it's the most critical foundational material for short-distance, high-bandwidth transmission, and fiber optics is the next step."
· "Copper remains the most critical material for short-distance, high-bandwidth transmission, but once distances increase or heat becomes too high, a switch to fiber optics is necessary."
· "The next wave of capital will land on infrastructure companies that don't sound particularly sexy."
Why Energy is the Safest Bet
· "I've always been bullish on energy because, even if AI demand slows, energy itself remains a global necessity, and this demand will only increase."
· "The single trend that continues to rise regardless of the scenario is our demand for energy, electricity, and power. These companies are the ones I am most willing to be long-term long on."
· "The companies I most want to follow are those Jensen is investing in that also intersect with Leopold's logic. So, the company I'm currently closest to tailing is Marvell."
· "The best long-term positions are not necessarily the hottest chip companies, but the power infrastructure companies that are indispensable in any macroeconomic scenario."
Leopold's AI Investment Portfolio
Josh Kale:Leopold Aschenbrenner, this 24-year-old focused on AI investing, is now almost regarded by the market as the world's strongest AI investor. External rumors suggest the nominal size of his fund's positions now exceeds $20 billion. When we looked at Ejaaz's post a month ago, the fund size was only $13.7 billion, essentially doubling every quarter.
This time, we've obtained several significant new changes in his recent investment moves. Last episode, we discussed his portfolio, and the most surprising point then was that he was actually shorting a company almost everyone knows—NVIDIA, the world's highest market cap and hottest AI stock. Many couldn't understand why he placed over $9 billion in short exposure against such a company.
Now we have a new clue that might explain this. NVIDIA is actually raising capital, and through debt issuance. On the surface, this seems illogical. Why would a company of NVIDIA's immense size and extremely high profit margins take on an additional $25 billion in cash from a recently completed financing? Today, we want to combine this with Leopold's portfolio to discuss why he has made so much money, what he's looking at next, and what NVIDIA's financing truly means.
Ejaaz Ahamadeen: First, some background. Leopold Aschenbrenner was a former OpenAI researcher who raised a fund about a year and a half to two years ago. The initial size wasn't large, I recall around $200 million. But judging from his latest 13F filing, the fund's public holdings are now worth $13.7 billion.
So naturally, the market wants to know which positions he's taken, what his core investment logic is, and where his next big trade will land.
To understand this, you must first know that until about a month ago, Leopold was very optimistic about the entire AI sector, particularly bullish on the "selling picks and shovels" logic, i.e., GPU and upstream hardware suppliers like NVIDIA.
But about a month ago, the market discovered he wasn't as bullish on the semiconductor line. He remains bullish on the real bottleneck areas like memory and power, and likely also on new types of cloud providers, but he is notably not bullish on the world's most valuable company, NVIDIA. More specifically, he has placed a total of roughly $9 billion in bearish exposure against several companies seen as core AI infrastructure beneficiaries, including NVIDIA, ASML, and Oracle.
The Logic Behind Shorting NVIDIA
Ejaaz Ahamadeen: When this emerged, many started worrying, thinking the AI bubble might be about to burst. On the surface, NVIDIA's GPUs are still selling massively, and demand hasn't noticeably weakened. So where's the problem?
Later, we dug up several new clues, the most important being that NVIDIA just raised $25 billion externally through bond financing. This means it's not simply using its own cash but adding external leverage. So the question arises: Why would the world's most profitable, highest-margin, strongest-cash-flow company go out and borrow $25 billion?
Josh Kale: Furthermore, they initially intended to raise only $20 billion but ended up expanding to $25 billion, with subscriptions exceeding three times the amount. Last episode, when discussing this portfolio, we said not to worry about a bubble just yet because although these companies have massive capital expenditures, their revenues are high enough to theoretically support expansion with their own balance sheets.
But this is NVIDIA's first significant move to finance off its balance sheet since 2021, rather than directly using its cash. I recall it currently has around $12 billion in cash on hand. Putting all this together creates a strange tension: on one side, Leopold is shorting; on the other, NVIDIA, seemingly with infinite cash and profits, is issuing debt. So what's happening?
Breaking Down NVIDIA's Bond Financing
Josh Kale: Ejaaz, can you break down this transaction itself for us? Because this isn't ordinary financing; it's a bond issuance. Ultimately, NVIDIA's balance sheet now has an additional $25 billion, and the interest rates likely appear very low.
Ejaaz Ahamadeen: I'll present both explanations. NVIDIA originally had about $13.7 billion in cash, meaning it could have just spent its own money. So why raise external capital? The simplest analogy is buying a house. Many people choose a mortgage even if they have the full amount because their own capital can be used elsewhere, and if the borrowing cost is low enough, it's actually more economical.
The interest rate environment hasn't been friendly in recent years, but if you're NVIDIA, one of the world's most valuable and sought-after companies, you can borrow on very favorable terms. This $25 billion bond issuance has maturities ranging from 2 to 30 years and can almost be considered very cheap money, with interest rates approaching US Treasury yields.
Moreover, the offering was roughly oversubscribed by 4 times. In other words, there was $85 billion in market demand chasing the $25 billion offering; NVIDIA could practically pick its investors. If you only look at the official statement, NVIDIA's explanation is that this is primarily for financial arrangements, to repay and refinance some existing debt. Google did something very similar a few weeks ago and also in February this year. So you can certainly accept this explanation, viewing it as financial optimization.
But the other side is hard to ignore: Over the past month and a half, NVIDIA, Amazon, Google, and several other hyperscale cloud providers have almost all been increasing external financing. Some via debt, some via equity sales. Leopold's view might not be entirely without merit—could this be a sign of the bubble starting to loosen, the house of cards beginning to wobble? However, if you look solely at the financial structure, it doesn't yet clearly point to danger.
Josh Kale: I see it similarly. $9 billion short on NVIDIA is a truly massive position. But during our research, we saw something else: on May 18th, NVIDIA's board authorized an additional $80 billion for share buybacks and increased the dividend from $0.01 per share to $0.25, a 25-fold increase.
If a company simultaneously engages in massive stock buybacks, significantly increases dividends, and also borrows money in the same month, it's clearly not borrowing due to a cash shortage. A more reasonable explanation is that this is cheap capital, and the financing methods within this AI cycle are undergoing a slight shift. Everyone wants to participate in these capital moves, and NVIDIA realizes borrowing via debt is even cheaper than other financing methods, so it just goes ahead. For now, at least, NVIDIA itself is still doing very well.
Why He Adjusted the Portfolio
Josh Kale: This brings us to another question. What exactly is Leopold thinking? Why has his judgment changed? The stock chart you just showed also indicates NVIDIA's recent performance hasn't been particularly strong, but it's not terrible either. It's still the world's largest company by market cap, nearing $5 trillion, down only about 7% in a month—nothing significant amidst the surge of other AI stocks.
Ejaaz Ahamadeen: I don't think NVIDIA will disappear. Its GPUs, including the CPU product line launched a few weeks ago, I believe will perform very well. AI product demand is currently in exponential surplus, and NVIDIA remains the primary core machine supplier capable of meeting this demand.
But I do think the classic 'selling picks and shovels' trade in AI has become too crowded, and Leopold's recent portfolio changes convey precisely this signal. Looking at his recent 13F, his bearish exposure is clearly skewed against the semiconductor line, such as NVIDIA, ASML, Oracle, and other infrastructure-level companies.
Yet simultaneously, he is heavily invested in directions like memory, power, and new types of cloud. This indicates his judgment is not that AI infrastructure has peaked, but that certain tiers within the infrastructure stack, particularly semiconductors and traditional popular stocks, have become excessively crowded.
If the question becomes where the money will flow next, there are two answers. The first is the most direct: flow into the next genuine infrastructure bottlenecks, namely power, memory, and data center network segments. The second answer is that mysterious investment exposed just a few weeks ago.
The Unexpectedly Exposed Anthropic Position
Josh Kale: This was the most surprising to me. I only learned about it yesterday from you, and my first reaction was disbelief. Could it be that Leopold's fund, 'Situational Awareness,' has approximately 20% allocated to Anthropic equity? Current external rumors suggest this company constitutes about one-fifth of Leopold's fund. The Wall Street Journal and several other media outlets have reported this, and sources very close to the transaction have confirmed it.
This becomes a completely unexpected card in his portfolio.
Because 13F filings only disclose public market holdings, not private equity, and Anthropic happens to be a large piece of non-public equity. This is also why people have begun to understand why external estimates value his portfolio at $20 billion.
If 20% of the fund is Anthropic, and he invested around early 2025, then the returns on Anthropic over that year have been akin to seven years' worth. This change forces a significant revision in our understanding of his entire investment portfolio.
Ejaaz Ahamadeen: Yes. His first investment in Anthropic via private channels or the fund was around March 2025, when Anthropic's valuation was approximately $60 billion. Now, based on the latest funding round, it's valued at $96.5 billion.
This represents nearly a 15x increase. According to the algorithm presented in our show today, his latest 13F disclosed liquid portfolio value is $13.7 billion. If we add the Anthropic portion reported by The Wall Street Journal, roughly an additional $7 billion, the total fund management size reaches $20 billion.
How exaggerated is this? Bill Ackman, a top investor with three to four decades in the market, runs Pershing Capital at a similar size of around $20 billion. Leopold has been in this game for only a year and a half, he's 24, and has virtually no real investment experience.
Yet he has made some incredibly prescient calls. What's crazy is that he practically wrote all of this out in advance. When he launched the fund a year and a half ago, he published a 65-page AI essay titled 'Situational Awareness,' almost completely laying out the entire logic, including how capital would rotate from semiconductors and certain infrastructure segments to other bottleneck constraints. The market is now developing along these lines, which is truly astonishing.
The Next Wave of Infrastructure Trends
Ejaaz Ahamadeen: So this also tells me where the next money will flow. If he is cautious on NVIDIA, then capital will move to areas like power and memory; at the same time, he also wants to invest directly in the 'mine' itself, rather than continuing to only buy the 'shovels.' Anthropic is his favorite 'mine.'
Josh Kale: This indeed looks like a new trend, and again, he is earlier than most. For the past 12 months, everyone has been searching for AI's bottlenecks—rare metals, memory, RAM, etc.—and the market chased each wave. Those judgments weren't wrong because that wave of gains did happen.
But now, the valuations of directions seen as bottlenecks are gradually becoming rationalized. People already understand these companies' business models, market potential, and future revenues better, so much of the value is already priced in. In the next round, we care more about where the subsequent money will continue to flow.
The direction you mentioned—land, power, enclosures, physical infrastructure—seems correct. Because if we think about what's truly most important for AI, the answer increasingly seems to be physical construction capability. Look at xAI, or more accurately, look at SpaceX, which is now public. Its revenue core isn't rockets themselves but AI infrastructure construction.
Consider its recent deals with Anthropic and Google, which have created value exceeding the combined sum of Starlink, Starship, and the entire satellite business. There is clearly immense demand and value here. So the question becomes, who can actually build these things?
SpaceX is clearly one answer. Last night after hours, its stock was around $230, implying a valuation of about $3.1 trillion. We'll dedicate an episode to SpaceX this week because its recent run is simply too extreme. It just completed the acquisition of Cursor, its valuation has reached $3 trillion, and Elon Musk made more money in one day than Warren Buffett earned in his entire career.
Who Will Capture the Next Round of Profits
Josh Kale: We care about which companies are best at this hardware infrastructure, at developing those 'machines that build machines.' Combining Leopold's direction and the broader trend, we think the next wave of capital will move here. So Ejaaz, in reality, which companies will this rotation land on?
Ejaaz Ahamadeen: Many will be those unsexy-sounding infrastructure companies. A name frequently mentioned recently is Marvell. A few weeks ago at Computex in Taiwan, Jensen Huang directly stated on stage that this would be the next trillion-dollar company.
And just three months before he made that statement, NVIDIA had invested $1.5 billion in Marvell. I'm starting to wonder if this counts as insider trading or market manipulation, because after he said that, the stock rose another 70%.
I think it's easy now to directly claim AI infrastructure has peaked. But if you compare it to historical financial crises, like 2008, that flavor of high leverage, financial engineering, and systemic manipulation hasn't fully emerged this time.
There are two key differences. First, the products these companies are making today are genuinely being purchased. There wasn't such solid real demand during the dot-com bubble or the financial crisis. Second, constrained by physical laws, we can't infinitely add leverage now because the entire system is constrained by manpower and construction capability.
Even if you raise endless amounts of money, you cannot build data centers fast enough, expand memory chip capacity sufficiently, or immediately expand the power grid, transmission lines, and related infrastructure. There aren't enough people on the ground, and permitting, regulation, and various procedures are also hindering progress.
So I actually think this gives investors an advantage. Since you already know the hottest chip and 'pick and shovel' trades are too crowded, then the next money will flow into power, data networks (companies like Astera Labs), and other related segments.
What you really need to think about is when these contracts start to materialize, when these fabs are actually built, when SpaceX rockets can launch AI satellites, and even when they can start using solar energy to train AI models.
The timeline determines the betting rhythm. At least that's the framework I'm investing by, though this is not investment advice. I see it this way because over the past year and a half, we've witnessed firsthand how capital flowed from general AI stocks to semiconductors and infrastructure trades.
Josh Kale: If you continue looking at this portfolio chart, you'll see this story is clearly written into his holdings structure. By category, what is his largest allocation? Power and energy. Next is memory, followed by cloud and GPU miners—the most tangible infrastructure.
He wants to hold new cloud providers like CoreWeave and also miners who have pivoted to cloud compute. What he wants to own is this physical infrastructure because he believes this is where the real bottlenecks are. As you mentioned, there are many finer details, like the immense difficulty of actual construction, hardware manufacturing, and data center building itself.
If we ask where the biggest bottlenecks are, even permitting might be one. Who is solving these problems? SpaceX wants to move data centers to space, Tesla wants humanoid robots to address labor shortages. But both are far off. In the short to medium term, there exist vast blank opportunities, and this is precisely where Leopold is betting.
The Advantages of Optical Modules and Fiber Optics
Josh Kale: I want to add one detail we didn't expand on earlier. For those wanting to dig deeper and find more alpha, many of his clues lie in optics and deeper layers of the tech stack. Ejaaz, you've been researching this lately. Can you explain his thinking?
Ejaaz Ahamadeen: If you look at the positions on his screen, CoreWeave and Iron are essentially top-tier new cloud service providers. Simply put, they are somewhat like Amazon Web Services, except AWS provides cloud services to internet companies, while these companies provide ready-made GPU infrastructure to AI companies.
They handle setting up GPUs, networks, deployment—all so AI companies don't have to worry about underlying infrastructure and can directly train models and access compute. CoreWeave and Iron have been among his largest concentrated positions since inception and have delivered the highest returns.
Notably, he still holds these two companies among his largest positions today. This also indicates one thing: in his view, this trade is far from over. Furthermore, he has privately invested in Core Scientific, a company that can help unlock CoreWeave's infrastructure supply capacity. In a sense, he's adding another layer of leverage to CoreWeave.
Beyond these, look at companies like Coherent and Lumentum; they are essentially suppliers related to fiber optic and optical connectivity. To explain in the simplest terms, for semiconductors and GPUs to communicate with each other, traditional methods often rely heavily on copper wires.
The issue is, as GPU clusters grow larger, copper wires will get hotter, energy losses will increase, and efficiency will suffer. In this scenario, fiber optics become the next upgrade direction. It enables faster data transfer, is more cost-efficient, and allows companies providing inference and training compute to earn more. So you'll notice his bets have always been very infrastructure-oriented, investing in both these optical companies and power-related companies. It may not sound sexy, but in my view, this is where the money is truly flowing now.
Josh Kale: The copper aspect is also interesting to me because I only recently realized how critical it is for short-distance data transmission. For many high-bandwidth, short-distance transmission scenarios, copper is almost the only material everyone truly wants to use. Only when it becomes unsuitable, such as over long distances or with excessive heat, do they switch to fiber optics. So the combined market demand for copper and fiber optics is currently very strong, which is why observing the copper trade is interesting.
The recent strength in copper futures is essentially because everyone needs it; it's the most critical foundational material for short-distance, high-bandwidth transmission, and fiber optics is the next step.
Thinking at an even more fundamental level, the materials angle has always been interesting. The absolute bottom layer beneath all layers is essentially the raw materials most core to achieving intelligence. Copper is one, lithium is another, and many others. We really should do a dedicated episode on materials. Perhaps Leopold hasn't reached that layer yet, and we might see the next rotation before he does.
Josh Kale: If you keep going down the stack, you could even look directly at copper mines to see how these materials are produced. But returning to the core judgment, I think the next rotation indeed shifts from seemingly smaller bottlenecks to the truly difficult things: hardware and large-scale data center construction.
Whoever can build the data centers will take the profits. We've already seen how much SpaceX is earning due to immense data center demand. Whoever can bring more data centers online faster, provide sufficient power and GPUs, will earn the most. This is essentially what Leopold is betting on now.
Is a Bubble Forming?
Josh Kale: To summarize, we don't believe we are in a bubble-bursting phase yet. Leopold's portfolio seems more like a rotation than a full-scale retreat. So, should we still follow his lead?
Ejaaz Ahamadeen: I admit, when I first saw his 13F, my initial reaction was: this guy is shorting the world's most valuable company with demand booked until 2029? That's outrageous. But now, seeing this financing, I'm starting to think if NVIDIA continues to add external debt in the future, or even potentially sells equity, if this trend persists, then Leopold might be right once again.
If that happens, his fund could ultimately surpass the world's top traders and best investment funds. He has been consistently winning; it's hard not to respect that.
Josh Kale: However, another important point: his life so far has almost always been long-only; he hasn't truly been tested by large-scale selling. We mentioned Bill Ackman earlier—achieving 30x returns and surviving in the market for 30 years are two different things.
If he can maintain this growth trajectory and also learn when to press the sell button, how to manage risk, how to use hedges for protection, that would be even more formidable. We are starting to see the embryonic form of this capability. That $9 billion short isn't achieved by directly shorting $9 billion in cash but through options and leverage, not a one-to-one naked short. Regardless, this is all worth continued observation.
Energy is the Core Bet
Josh Kale: If you had to pick one stock from his entire portfolio that you most want to buy yourself, which would it be?
My own answer is energy stocks. I've always been bullish on energy because, even if AI demand slows, energy itself remains a global necessity, and this demand will only increase.
Even completely ignoring AI, we need more energy, more electricity. Companies like Bloom Energy that can enhance power supply and transmission capacity excite me most because they resemble a hedge-like bet. The single trend that continues to rise regardless of the scenario is our demand for energy, electricity, and power. These companies are the ones I am most willing to be long-term long on.
Ejaaz Ahamadeen: My answer is a bit of a cheat. The companies I most want to follow are those Jensen is investing in that also intersect with Leopold's logic. So the company I'm currently closest to tailing is Marvell. While not a public holding of Leopold's, it aligns very well with his bets on fiber optics and power, and Jensen has already invested $1.5 billion of real money.
I've observed a phenomenon: whenever Jensen invests in a company through NVIDIA, whether it's Intel, CoreWeave, or others, it mostly keeps rising afterward. So my current positioning is roughly here. I also hold some CoreWeave myself because both Jensen and Leopold are extremely bullish on it.
Josh Kale: Marvell is up 270% over the past 6 months. This might genuinely be a good rule of thumb: when influential figures like Jensen, or even someone with enormous sway like Trump, publicly say to buy a certain stock, you probably should take a serious look.
Past instances have repeatedly shown such signals often have substantial realization potential. Whether Intel or Marvell, these cases show that, on one hand, they truly understand what they're talking about, and on the other, they have the ability to influence these companies' outcomes. So this market run is truly insane.
I hope it continues. Based on current evidence, it likely will. At least for now, we remain bullish and optimistic, and will continue making judgments daily based on changes.
Josh Kale: Any final thoughts you want to add regarding Leopold's portfolio update?
Ejaaz Ahamadeen: I'd really like to hear what skeptics think. If after listening to our analysis, you think we're completely wrong or have misinterpreted something, feel free to point it out directly.
Yesterday, I stared at NVIDIA's $25 billion financing news for a long time, intending to find flaws. But if you look purely at financial logic, it does make sense.
Why not borrow this almost risk-free cheap money? Using borrowed money for expansion is clearly more rational than selling your own equity because you retain more future earnings.
