Situational Awareness: The 25 year old taking on Wall St

A deep dive into Leopold Aschenbrenner's L/S Equity Portfolio

Leopold Aschenbrenner’s AGI hedge fund has hit ~$5.5bn in equity exposure after ~18 months of operations. The fund famously beat the S&P500 by 47% in its first 6 months.

If you weren’t aware already, Leopold’s track record is ridiculous:

Timeline:

1. Age 19: Graduates Columbia as Valedictorian

2. Age 22: Joins OpenAI’s Superalignment team

3. Age 22: Fired for raising AI safety concerns

4. Age 23: Publishes viral manifesto on AGI

5. Age 23: Launches hedge fund with $225M

6. Age 24: Returns 47% in six months

7. Age 25: $5.5B in equity exposure

But even better - Situational Awareness LP’s 13F filing just dropped - so we can see what he’s holding.

So how is a 25 year old taking on the Titans of Wall St? And will his strategy succeed if the AI bubble pops?

Let’s begin.

---

Fund Strategy

So, why did an AI researcher start a hedge fund?

At first glance, it feels brash to assume that an economics whiz with no track record in finance could start outperforming armies of Wall St analysts trained at places like Blackstone. After all, even Peter Thiel couldn’t get Clarium Capital Management to beat the markets for an extended period of time, going bust in 2008-09 during the GFC.

But the reality is that investors get paid for asymmetric information, and it’s easy to have asymmetric information when you have a different context. Where his competitors are watching the AI race unfold from afar in Manhattan, Leopold is situated in the beating heart of San-Francisco, with his finger on the pulse of the labs. Where venture funds are tracking METR for latest benchmarks, your average Hedge Fund investor probably hasn’t heard of Cursor.

More broadly, it feels like many of the forewarnings of the AI labs are starting to shape out as themes impacting the public markets. In a world where software stocks drop ~20% on announcement of an Anthropic release, being an AI first equities investor is an incredible advantage. Block recently announced a 40% reduction of their workforce, in part due to automation which is enabled by language models. More and more, just as the internet, and mobile were driving forces in the economy over the previous decades, AI will be the tidal wave that starts to impact every company.

The capital markets have been lapping up AI in recent years as the new source of growth amidst a largely stagnant economic backdrop. Capex spending amongst Hyperscalers is set to reach ~$650bn in 2026 (Bloomberg), exceeding the cost of the Apollo Program in inflation adjusted terms. A similar narrative has been true in private markets, where the vast majority of venture dollars have been allocated to AI companies. OpenAI’s recent $110bn fundraise would represent roughly 25%-30% of total venture capital dollars raised in 2025, a staggering figure.

If something is clear about Leopold’s investment strategy - it’s that he is not a value investor. Situational Awareness’s average multiple across their 5 largest positions is ~14x EV / Revenue, a far cry from the disciplined strategies of Buffett or Marks. In a world where stocks are trading on momentum, this strategy works incredibly well.

But what happens if the merry-go-round stops turning?

---

Portfolio Concentration Risk

The first risk to consider is the high level of correlation in the Situational Awareness portfolio.

Just as Amaranth Capital blew up in 2007 having placed ~80% of the portfolio into energy and natural gas bets, Leopold runs a huge risk given SA’s concentration. The fund is one large bet that compute resources are too scarce to serve AI demand.

In part, this is due to the lack of other AI investment options in the public markets - you can’t buy lab equity on the NASDAQ. As Charlie Songhurst noted in his interview on Cheeky Pint, one of the peculiarities about the market bubbles is that they often manifest by investors pouring capital into markets in “ways that they know how”. The Internet Bubble of 1999 was really a Telecom Bubble, with a dark fibre buildout to support the infrastructure of the internet.

Similarly, in the recent AI wave, with public equity investors locked out of the venture rounds of large model providers, a large pool of dollars have been forced into derivative AI bets backing suppliers of compute at steep valuations.

It could be argued that the ‘AI bubble’ so to speak exists largely outside of venture capital, in the data centre market.

The saving grace for Leopold here is his investor lock-up period, forcing LPs to multi-year hold periods before they can redeem capital. Unlike private markets, where LPs commit to 10+ year fund lives and assets aren’t marked until subsequent fundraises, public market portfolios show incredible volatility because stock prices move everyday. Historically, many hedge funds have struggled with liquidity crunches, forced to liquidate positions after short-term downturns when LPs get nervous, right at the point in time where you should be ploughing more dollars into your positions.

Given the concentration in his portfolio, the investor lockup should enable him to weather the storm (for a while at least).

---

Correct Thesis, Incorrect Timing

Situational Awareness LP is very much a bet on accelerated AI timelines being mispriced in the public markets. The opening chapter of Leopold’s Manifesto breaks this down clearly:

“You can see the future first in San Francisco … The AGI race has begun.

We are building machines that can think and reason. By 2025/26, these machines will outpace many college graduates. By the end of the decade, they will be smarter than you or I; we will have superintelligence, in the true sense of the word.”

But what if timelines take longer than we think?

The narrative espoused by the AI labs is that scaling laws will hold; and provided that they hold, we will see incredible productivity growth and a world of abundance. In many ways, this is a self-serving narrative. To justify the valuations of their recent fundraises, companies like Anthropic and OpenAI need to portray a view of the world where AI will be a sweeping force that changes everything.

But the observed reality is very much disconnected from the benchmarks. Inside organisations, ~90% of businesses have not adopted AI in a meaningful way (RBA). The failure of improved model performance to have directly observable impacts on the economy is encapsulated in Sam Altman’s blog post The Gentle Singularity, with the goal posts for AGI shifting with every model release.

Anecdotally, the reality of AI adoption being slow, and largely incremental feels true. Very few of my friends have adopted AI in a way that meaningfully improves their productivity at work, and the possibility of becoming fully automated feels further away than the narratives of Dario or Sam would suggest.

Timing the revenue ramp will be critical for Situational Awareness. Underpinning the compute shortage thesis behind SA’s positions in the Neoclouds, and HPC component manufacturers is the implicit assumption that demand for AI related products will be significantly higher in the future than it is today. But will we start seeing revenues inflect soon enough for this thesis to be correct? OpenAI’s revenue growth rates are starting to slow, only growing ~17% from November 2025 to March 2026, and making their 2030 target of $280bn look increasingly ambitious.

The risk is that if progress isn’t fast enough, there won’t be enough profitable demand for tokens to justify continuing to spend on data centres

---

Quantitative Funds Eating Returns

The last thing to note is just the dominance of high frequency trading firms has eaten away the alpha in public equity markets that built the hedge fund industry. In Q3 of 2025, Jane Street generated ~$6.8bn in revenue, rivalling the revenues of JP Morgan and Goldman Sachs despite only having 3000 employees. Jim Simons’ medallion fund is famously the best performing public equity investment firm ever, generating 39.9% net of fees since 1988.

Whilst these investors trade on a significantly shorter time horizon than typical L/S equity managers, they iron out much of the pricing inefficiency in the market, and scalp equity investors on their way into large positions by identifying when whales start to move.

It’s never been harder to be a successful public markets investor. But perhaps if anyone can do it, Leopold can.

---

Investment Themes

Okay, okay, enough on the strategy.

What are these guys actually holding?

We can break down Leopold’s portfolio down according to a few different investment themes:

1. Data Centre Energy Production (Bloom Energy, EQT)

2. Neoclouds (Coreweave / IREN / Cipher Mining / Applied Digital / Core Scientific)

3. HPC supply chain: Memory manufacturers & Optical communications

---

1.1 Onsite Data Centre Energy Thesis - Bloom Energy

Situational Awareness’s largest position is in Bloom Energy (15% fund), a company focused on the design, manufacturing and deployment of power systems, providing on-site renewable electricity generation across a range of applications (the fastest growing segment of which is data centres). This echoes the direction of Blackstone, monetising AI through data centres and supporting the grid, rather than investing in the models or software layer.

In a world where data centre capacity coming online is struggling to find access to cheap power through the grid (Credit Semianalysis above), on-site power solutions become critical. One of the reasons why the X-AI colossus data centre was able to come online so quickly was their energy strategy, able to source a fleet diesel turbines to solve the power crunch, where other energy solutions can take 2+ years to come online.

The lead times on gas turbines in the US have increased to 5 years, and one of the primary reasons supporting the SpaceX / X-AI push into Orbital Data centres. Given these supply constraints, companies like Bloom that provide an immediate solution to the energy crunch become very interesting.

There are 4 important things to note about Bloom’s technology:

1. Flexible Fuel Options: Their fuel cells run primarily on natural gas, but are future-proofed to use hydrogen or blends.

2. Lower Emissions: Compared to combustion, BE’s systems emit significantly less SOx and NOx, with a much higher fuel efficiency, especially in their combined heat and power (CHP) setups.

3. Resilient and Reliable: BE’s solutions offer high uptime (>99.999%)—crucial for industries like data centers.

4. Rapid Deployment: Systems can be installed in as little as 90 days, far outpacing traditional power solutions.

Bloom is currently trading at a ~70x EV / 1-yr forward EBITDA multiple and only grew ~35% last year, so it’s very expensive. Equity researchers are also relatively bearish on the business, with most recommending a hold given the current valuation. Looking forward, the business is expecting revenue growth to accelerate next year to >50%, and gross margins to increase, both of which should help the company improve free-cashflow, but is it enough?

The willingness to invest in a ~70x earnings energy manufacturer gets to the heart of Leopold’s background, and the SA investment strategy, focusing more on a mispriced view of Bloom’s future demand rather than grounding investments in fundamental analysis.

1.2 Data Centre Energy: EQT (Natural Gas Provider)

Situational Awareness also increased their stake in EQT over the last quarter, the largest producer of natural gas in the US (not to be confused with Swedish Private Equity firm EQT Group). The business operates as a vertically integrated company from exploration, to the drilling and extraction of natural gas, and through the acquisition of Equitrans, now involved in the midstream (gas pipelines) segment. EQT announced a $15bn data centre natural gas supply deal Pittsburgh in July 2025, an early harbinger of how important natural gas will be in powering data centres.

---

2.0 Neoclouds: (Coreweave / IREN / Cipher Mining / Applied Digital / Core Scientific)

The second key investment thesis driving SA’s portfolio is a positive outlook on Neoclouds, increasing holdings in Coreweave, IREN, Applied Digital, Core Scientific and Cipher Mining (all of which are at various stages of transitioning their data centre portfolios away from Bitcoin mining ASICS towards serving GPU workloads).

Many investors have been incredibly skeptical of the Neoclouds, and for good reason. Their product offering appears to be relatively commoditised, they are forced to front large capital requirements upfront, and the long term demand of their customers is still uncertain. Hyperscalers have been leasing significant capacity from these companies for the next 5 years in anticipation of profitable consumer demand for models, but what happens beyond this? Will these Neoclouds be sitting on commoditised data centre capacity with obsolete chips? Or will the hardware retain value. We have not yet seen the inflection in profitable use cases for models necessary to justify the compute buildout, something David Cahn explores in his piece AI’s $600bn question.

More recently, there has been a racket amongst value investors over a shift in accounting principles adopted by the Hyperscalers and many of the Neoclouds, adopting aggressive assumptions around GPU depreciation that assume a useful chip life of 6 years, where for traditional ASICs, a 3 year useful life was assumed. This helps to juice short term earnings by reducing D&A on the income statement, and could be argued was used as a lever to justify the incredible capex buildout.

But recent GPU pricing data is optimistic. Dylan Patel pointed out on Dwarkesh recently that:

“an H100 today is worth more than a H100 yesterday”.

Let’s unpack this.

If we get to an actual human on a server, the elusive ‘drop in remote worker’, these chips will become incredibly valuable as a constraint on economic performance output in the same way that population growth is today. The utility of models, and the economic value of tasks that we are seeing them handle is increasing over time. Software engineering was first, but what happens when these models start to eat into the work of lawyers, doctors, consultants, and more of the white collar economy?

We could see a big spike in the utility, and therefore price of older chips if this case holds true.

But the data on actual chip pricing doesn’t really stack up here. There appears to have been a general downward trend in H100 pricing since 2023 from $7 - $10 USD per GPU hour in 2023 (Silicon Data) to $2 - $4 per GPU hour in late 2025.

Overall, I have the least conviction in SA’s neocloud positions across the key investment themes. But perhaps we need to go deeper in the stack?

It is possible that semiconductor fabrication capacity constraints will help improve GPU pricing over the next 5+ years. Despite the significant increase in demand for high performance chips from the Hyperscalers and leading labs, and subsequent increases in NVIDIA’s planned production capacity, TSMC has not made the capex requirements necessary to meet these huge demand increases (Credit Semianalysis).

---

3.0 HPC Upstream Suppliers: Lumentum, SanDisk

SA also has meaningful stakes in SanDisk (4% portfolio) and Lumentum Holdings, producers of components and systems used in High Performance Computing (HPC) clusters. SanDisk is a designer and manufacturer of flash memory storage products, and has significant exposure to AI given the importance of SSD (Solid State Drive) memory in data storage for AI training / inference, used as a component inside of data-centres. They compete with players like Micron and SK Hynix in their core SSD market. AI has caused dramatic increases in the demand for DRAM, with OpenAI reporting in October 2025 that they alone plan to consume up to 40% of global DRAM capacity, or 900k wafers per month (Contrary Research), inking contracts with SK Hynix and Samsung. This is because subsequent generations of NVIDIA GPUs are using more and more DRAM (Bloomberg), meaning that memory is increasingly becoming a bottleneck to scaling compute, and as a result, these companies should see strong revenue and margin improvements. These themes have certainly held true over the last 6 months, with memory stocks ripping on the back of rising memory prices (and associated margin improvements for these businesses), with SanDisk’s gross margins hitting 51% in Q2 2026 (up 29% YoY).

---

---

Overall Thoughts

Where most Wall St investors are obsessing over multiples for industrial companies started in the 1960’s, Situational Awareness is chasing mispriced growth.

Like many of the best technology investors - Leopold is betting on exponentials and non-linearity. If Marc Andreessen’s coining was “Software is Eating the World”, then Leopold’s would be “AI will eat everything”.

Overall, I like the investment strategy. Wall St tends to be a few years behind Silicon Valley, so it will be interesting to see if the returns stack up in the long term.

The next Druckenmiller? Or is it just AI hype?

Only time will tell.

---

Thanks for reading!

Until next time.

Wrucky. :))

Also make sure to subscribe and drop your comments down below (I read every reply)