AI Guru Leopold's High-Stakes Comeback: Inside the Situational Awareness Fund's Risky Bet on AI Infrastructure Bottlenecks

The numbers don't lie. A fund that controlled $45 billion at its peak just months ago now sits at roughly $100 billion — an 80% collapse driven by leverage so aggressive it makes most hedge fund managers wince. And now, according to sources close to the matter, this same fund is back in the game, loading up on hundreds of millions of dollars in AI infrastructure and storage-related options. The play? A concentrated bet that the AI boom's next bottleneck won't be GPUs — it will be everything else. SK Hynix. SanDisk. AMD. Bloom Energy. CoreWeave. The Roundhill Memory ETF. This isn't diversification. This is a thesis wrapped in ticker symbols, and it's either brilliant or reckless. Probably both.

Let me break down what's actually happening here, because the mainstream coverage is missing the real story. The collapse wasn't just bad luck — it was structural. And the comeback isn't just confidence — it's a high-wire act with no safety net.

What Actually Happened to the Situational Awareness Fund

The Situational Awareness fund, run by the former OpenAI researcher known in financial circles as Leopold, represents one of the most aggressive expressions of AI infrastructure bullishness I've seen in my years covering these markets. The fund doesn't just believe in AI — it believes in the physical infrastructure that makes AI possible, and it was willing to lever up to unprecedented levels to prove it.

At its July peak, the fund commanded over $45 billion in assets under management. That's not a typo. Forty-five billion dollars concentrated in a strategy built on high-leverage positions in AI-adjacent equities. For context, that's larger than many established crypto protocols' fully-diluted valuations during peak cycles. The scale alone should have raised eyebrows, but when the trade works, nobody asks questions.

Then the music stopped.

The summer of 2024 delivered a harsh reality check for AI-adjacent positions. Rate expectations shifted, profit-taking hit momentum names hard, and suddenly the leverage that amplified gains started amplifying losses in the exact same proportion. The fund found itself in a liquidity spiral — the kind of situation where you're forced to sell what you can, not what you want. According to sources familiar with the matter, the fund liquidated nearly its entire public equity portfolio, selling the positions at discounted prices to a firm called Castle Securities.

Castle Securities acquiring those positions at a discount is the market's version of vultures circling. When a fund is forced to liquidate, it doesn't get to pick its timing. It doesn't get to negotiate gently. Castle got the assets it wanted at prices it dictated, and the Situational Awareness fund shrank to approximately $100 billion — a loss of roughly $35 billion in market value, or about 78% from peak.

Think about that for a second. In a matter of months, a fund that controlled nearly $50 billion in assets was reduced to a fifth of its size. The merge wasn't a slow bleed — it was a margin call cascade that forced public liquidation at the worst possible moment. That's not investing. That's gambling with borrowed time.

The Comeback: Same Thesis, Same Leverage, New Capital Structure

Here's where it gets interesting — and by interesting, I mean concerning. The same fund that just experienced one of the most dramatic drawdowns in recent memory is back. According to people familiar with the fund's recent activities, it has been purchasing hundreds of millions of dollars in AI infrastructure and storage-related options over the past several weeks. The thesis hasn't changed. If anything, it's become more concentrated.

The new positions include:

• SK Hynix: The South Korean memory giant that produces High Bandwidth Memory (HBM) chips critical for AI accelerators. SK Hynix is essentially the supplier that NVIDIA and other AI chip manufacturers depend on for the memory that makes their compute possible.

• SanDisk: Western Digital's consumer and enterprise storage brand, representing exposure to NAND flash memory that stores the massive datasets AI models are trained on.

• AMD: The Santa Clara chip designer positioned as the primary alternative to NVIDIA in the AI accelerator space. AMD's MI series chips are the credible second choice for organizations that can't get enough NVIDIA allocation.

• Bloom Energy: A company that makes fuel cell systems for on-site power generation. This is the wildcard — it represents the thesis that power infrastructure is the next bottleneck for AI data centers.

• CoreWeave: The GPU cloud rental specialist that has become the poster child for the "neocloud" phenomenon. CoreWeave rents access to NVIDIA GPUs to AI companies that don't want to buy their own hardware.

• Roundhill Memory ETF (DRAM): A diversified play on the memory industry through an exchange-traded fund that tracks DRAM and memory-related stocks.

The combination tells a clear story. This isn't a scattershot bet on AI. This is a vertical thesis about AI infrastructure bottlenecks — specifically, the idea that after GPUs, the next constraint on AI expansion will be memory (HBM, DRAM, NAND), power generation, and data center capacity. The fund is essentially betting that AI's growth will run into physical limitations that the market hasn't fully priced.

But here's what the headlines aren't telling you: the fund's leverage hasn't been disclosed. Nobody knows the actual leverage ratio. Nobody knows the option structures — are these simple calls? Vertical spreads? Iron condors? Calendar spreads? Each structure implies a completely different risk profile and market view. Are they betting on direction? Volatility crush? Time decay acceleration? The opacity is a red flag that the mainstream analysis is conveniently ignoring.

The Contrarian Angle Nobody Is Talking About

Here's what I keep coming back to: this fund retained its Anthropic private equity position throughout the liquidation. While nearly all public equities were sold at steep discounts to Castle Securities, the fund held onto its stake in the AI safety and development company. That decision reveals something critical about the fund's liquidity mismatch.

Private equity positions are beautiful when they're winning. They're nightmares when you need cash. You can't margin call your way out of an illiquid position when your broker is demanding more collateral on your options. The fund created a situation where its most liquid positions (public equities) were the first to get liquidated, while its least liquid positions (private equity) remain locked up. That's not portfolio management — that's triage.

And let's talk about the timing. The fund collapsed in July and is now back buying hundreds of millions in options. That means within a span of just a few months, the fund either raised new capital from investors or redeployed its remaining capital into a concentrated options strategy. Neither scenario is comforting. New capital from where? High-risk-preference family offices? Rehypothecated collateral? The sources say it's "unclear" where the money is coming from. In my experience covering crypto and traditional finance, "unclear" is a polite way of saying "we don't want to know."

The DRAM ETF position is particularly revealing. If you're confident about individual names like SK Hynix, why buy the ETF? The answer is either diversification (unlikely given the fund's concentrated history) or the need for liquidity. ETFs trade instantly. Individual stocks can have wider bid-ask spreads and less reliable execution during volatile periods. The fund might be using the ETF as a more liquid way to maintain exposure while keeping powder dry for the individual positions it really wants.

Or, more concerningly, the fund might not have enough capital to move individual stocks meaningfully and is using the ETF as a proxy. At $100 billion in AUM, hundreds of millions in options is still a significant position, but it's not the kind of overwhelming force that $45 billion represented. The fund is fighting with smaller ammunition now.

The Storage Cycle Thesis: Real or Overhyped?

Let's give the thesis its due. The argument that memory will be the next AI bottleneck has genuine merit. HBM (High Bandwidth Memory) is a specialized product that requires significant manufacturing expertise and capital expenditure. SK Hynix is the current leader, having secured a substantial share of NVIDIA's HBM supply. Samsung and Micron are racing to catch up, but capacity takes time to build.

The AI server market is consuming memory at unprecedented rates. A single AI training server can require multiple terabytes of DRAM — orders of magnitude more than traditional enterprise servers. The training of large language models requires not just compute but memory bandwidth, and the inference stage requires rapid access to stored weights and context. Memory isn't a footnote in the AI stack — it's a critical component that the market may have underestimated.

From my experience watching DeFi protocols and their infrastructure requirements, the lesson is clear: bottlenecks create the most profitable positions. When everyone is focused on compute (GPUs), the smart money looks for what comes next in the supply chain. Memory, networking, storage, power — these are the unsexy parts of AI infrastructure that nobody was talking about six months ago. Now they're the thesis.

But here's the counterargument: memory is cyclical. DRAM and NAND have experienced boom-bust cycles for decades. The "AI growth stock" re-rating of memory companies assumes that AI demand will be structural and sustained enough to prevent the traditional downcycle. SK Hynix might be the memory leader today, but Samsung's memory division is massive and well-funded. Micron has U.S. government support through the CHIPS Act. The competitive dynamics aren't static.

Hackers don't just look at the shiny new narrative — they look at what breaks the narrative. A memory oversupply in 2025 would crush this thesis regardless of AI's growth trajectory. The fund is betting on timing as much as direction.

The Power Play: Bloom Energy's Wildcard Position

Bloom Energy is the most unusual position in this portfolio, and it's the one that reveals the deepest conviction in the infrastructure thesis. Bloom makes solid oxide fuel cells that can generate electricity on-site, bypassing the traditional power grid. For data centers that need reliable, continuous power and can't wait for utility companies to expand grid capacity, Bloom represents a solution.

The logic is compelling: AI data centers are power-hungry. A single large-scale AI training facility might consume 100 megawatts or more — equivalent to powering a small city. The grid wasn't designed for this. Utility interconnection queues in many regions are years long. The only way to get power quickly is to generate it on-site.

Bloom Energy has been positioning for this moment. The company's servers provide continuous power with lower emissions than grid electricity in many regions. For hyperscale data center operators under pressure to meet sustainability commitments while rapidly scaling capacity, Bloom's value proposition is increasingly relevant.

But here's the catch: Bloom Energy's fuel cells run on natural gas or biogas. In an era of increasing climate scrutiny, this creates a reputational and regulatory risk that the market might not be pricing. The fund is betting that power availability trumps power source in the near term, but longer-term ESG pressure could limit Bloom's addressable market among the most sustainability-conscious buyers.

CoreWeave and the Neocloud Question

CoreWeave represents the fund's bet on GPU rental economics. The neocloud model — companies that buy NVIDIA GPUs and rent them to AI developers — has been extraordinarily profitable during the GPU shortage. CoreWeave went from a relatively obscure crypto mining operation to one of the fastest-growing AI cloud providers in the industry.

The thesis is straightforward: organizations want NVIDIA GPU access but don't want to buy $30,000+ servers and wait months for delivery. CoreWeave provides immediate access to GPU compute on a rental basis. The company has secured massive GPU allocations from NVIDIA through its early-mover relationships and has been signing long-term contracts with AI companies hungry for compute.

The risk is concentration. CoreWeave's primary value proposition depends on continued NVIDIA GPU supply and sustained demand for rental compute. If NVIDIA's supply catches up with demand, GPU rental rates would fall. If AI development slows — due to regulatory pressure, economic headwinds, or technology limitations — the demand side of the equation weakens. CoreWeave's customers are primarily AI startups and research organizations that may not have the capital buffers to weather a funding winter.

The fund's CoreWeave position suggests it's betting on continued GPU scarcity and sustained demand. That's a reasonable view given current supply-demand dynamics, but it's not a certainty. The company's private market valuation has been rising rapidly, and at some point, the premium for access to compute has to be justified by actual economics rather than scarcity premium.

The Systemic Risk Nobody Is Measuring

Here's the part that keeps me up at night: we don't know the leverage. The fund's $45 billion peak AUM almost certainly included significant borrowed money. When a fund of that size uses leverage, its positions can move markets. The forced liquidation in July likely contributed to the broader weakness in AI-adjacent names during that period. A fund with $45 billion in gross positions dumping everything at once creates price impact that cascades through the market.

Now the fund is back with a reported $100 billion in AUM and hundreds of millions in new options positions. Even if the fund is using less leverage now, its positions in relatively liquid names like SK Hynix ADRs, AMD, and the DRAM ETF could still create meaningful market impact if the fund needs to exit quickly.

The options market is particularly concerning. Options positions can create gamma squeezes when large players are on the wrong side of a move. If the fund sold put options and the market moves lower, delta hedging by market makers can force rapid selling that accelerates the decline. If the fund bought call options and the underlying stocks rally, market makers buying to hedge can create a feedback loop of upward pressure. Either way, the fund's positions are part of a complex derivatives ecosystem that could amplify volatility rather than dampen it.

For retail investors and smaller institutions watching this situation from the outside, the key lesson is this: when a fund with this much opacity and recent trauma starts making concentrated bets, you're watching a live experiment. The outcomes affect not just the fund's investors but everyone who holds similar positions.

What This Means for the AI Infrastructure Trade

Despite all the risks, the fund's thesis about AI infrastructure bottlenecks deserves serious consideration. The idea that the next constraint on AI growth will be memory, power, and data center capacity — rather than compute — is a coherent view that aligns with what I'm seeing in the crypto infrastructure space. Blockchain networks face similar bottlenecks: not just computational capacity but storage, bandwidth, and energy costs that determine what's actually possible.

The memory trade in particular has structural support. HBM production is capacity-constrained, and the lead times for new fab construction are measured in years, not quarters. SK Hynix and Samsung are investing billions in new HBM capacity, but demand is growing at least as fast. For traders looking at this sector, the question isn't whether memory demand will grow — it's whether current prices already reflect that growth.

The power trade is earlier in its development curve. Bloom Energy and competitors like Tesla's energy division are scaling, but the grid bottleneck is real and getting worse in many regions. Data center power demand is expected to grow 100% or more by 2030 according to some estimates. The companies that solve power delivery — whether through on-site generation, renewable integration, or grid expansion — will capture significant value.

The Bottom Line: Watch the Margin Calls

So where does this leave us? The Situational Awareness fund has made its bet. High-leverage, concentrated, opaque — everything that makes risk managers nervous and market watchers fascinated. The fund's collapse in July proved that even the most conviction-driven theses can unravel when leverage meets liquidity crisis. The fund's comeback suggests either exceptional conviction or exceptional stubbornness.

The difference between those two interpretations will be determined by what happens next. If AI infrastructure demand continues to accelerate and the bottleneck thesis plays out, the fund's options positions could generate returns that dwarf the losses from July. If the AI cycle shows signs of topping, if memory supply catches up faster than expected, if power solutions scale more quickly than anticipated — the fund could face another margin call cascade that makes July look like a preview.

For everyone watching this situation, the key metric isn't the fund's stated positions. It's the shadow data: options open interest changes, unusual activity in related names, correlation between the fund's portfolio and broader market moves. The merge wasn't the end of the story — it was the first chapter of a sequel that's just getting started.

The real question isn't whether Leopold is right about AI infrastructure. It's whether the market has already priced in the most optimistic version of that thesis, leaving the fund chasing returns that have already been made. And whether a fund with this much recent trauma and this much remaining leverage can survive the inevitable volatility that comes with betting against the consensus.

Forward Watch: Three Scenarios for the Next 90 Days

Bull Case: AI infrastructure demand accelerates through Q4 2024, HBM supply remains tight, memory prices rise, and the fund's concentrated positions generate returns that restore confidence. AUM climbs back above $150 billion. New investors pile in. The July collapse becomes a footnote in a larger success story.

Base Case: AI infrastructure demand grows but at a slower pace than the fund's leveraged positions require. Memory prices stabilize rather than soaring. The fund's options positions theta-decay into losses. AUM drifts lower. The fund limps along as a cautionary tale about leverage and concentration.

Bear Case: A broader risk-off event — rate spike, recession fears, AI regulation, or a major AI company's earnings disappointment — triggers another margin call. The fund is forced to liquidate not just public equities but potentially its private positions. Castle Securities or another distressed buyer acquires remaining assets at steep discounts. The AI infrastructure thesis survives; the fund does not.

The Situational Awareness fund's comeback is the trade that reveals whether the AI infrastructure cycle has legs or whether it's a crowded trade waiting for an exit. Watch the positions. Watch the leverage. Watch the margin calls. And remember: in markets, the loudest thesis isn't always the correct one. Sometimes the most conviction is just the most leverage.