A 25-year-old researcher fired by OpenAI invested $225 million into $5.5 billion in one year

Former OpenAI researcher Leopold Aschenbrenner was fired in 2024 for “leaking secrets,” then published a 165-page AGI paper and founded Situational Awareness LP (SALP), whose holdings ballooned from $225 million to $5.52 billion within a year, growing nearly 24 times at an eye-catching speed.
(Background summary: Clearing out NVIDIA, buying fuel cells: The logic of a 24-year-old genius managing a $5.5 billion AI fund)
(Additional background: Anthropic spent 200 billion dollars over five years on Google Cloud, with two AI startups taking half of the cloud giants’ orders)

Table of Contents

Toggle

• Key Highlights
• FTX collapsed, he emerged from the ruins
• 165 pages, $5.5 billion
• He buys the electricity bill for AI
• SanDisk up 17%, CoreWeave down 11%
• FAQs

Key Highlights

• SALP’s holdings grew from $225 million to $5.52 billion in one year, nearly 24 times, with founder Leopold Aschenbrenner only 25 years old
• Holdings exclude NVIDIA, focusing on AI infrastructure: power, storage, fiber optics, and transforming Bitcoin mining farms into HPC
• Today’s portfolio divergence: SanDisk +16.6%, Intel +14% lead gains vs. CoreWeave -11.4%, as the market reprices AI bottlenecks

An eviction notice was sent in spring 2024 from OpenAI’s HR department. The recipient was a German-American researcher, Leopold Aschenbrenner, citing leaking company secrets, which he refused to cooperate with in an internal investigation.

In Silicon Valley, being fired from a top AI lab usually means the end of your career. You’ve signed NDAs thicker than your resume, and the circle is so small that the next interview’s interviewer is probably a former colleague who just fired you.

Two months later, Aschenbrenner released a 165-page paper titled “Situational Awareness: The Decade Ahead.” Its content relates to everything he saw inside OpenAI, claiming AGI’s timeline is faster than everyone thought, and the world isn’t ready.

Eighteen months later, his self-founded fund SALP’s public holdings exploded from $225 million to $5.52 billion.

24 times, in one year.

FTX collapsed, he emerged from the ruins

Leopold Aschenbrenner’s resume reads like a guide on how to escape from well-known companies.

Before OpenAI, he worked at FTX Future Fund—yes, the fund used by Sam Bankman-Fried to package the “effective altruism” persona—investing in a bunch of now-very-valuable AI startups, the faucet of that era.

FTX blew up in November 2022 at one of the fastest speeds in financial history, evaporating $8 billion of customer funds. SBF is now in federal prison, serving 25 years.

That year, Aschenbrenner was about 21.

He didn’t stay in the crypto rubble; he turned and walked into OpenAI’s doors. Then OpenAI pushed him out again. Most careers can’t survive either of those events. He survived twice, then started a fund, but now it’s no longer a small seed-stage fund—it’s the fastest-growing AI thematic fund on Wall Street in the past year.

The mission of FTX Future Fund was to invest in the future with a long-term mindset, but it didn’t last a quarter. What Aschenbrenner picked up from that wreckage wasn’t money, but methodology.

165 pages, $5.5 billion

That paper, known industry-wide as “SA,” isn’t academic junk or PR fluff.

In it, Aschenbrenner drew a forecast no one dared to make: he suggests AGI won’t be in twenty years, but could arrive in the second half of this decade. And when it arrives, the bottleneck won’t be algorithms, data sets, or talent.

It will be power, chips, fiber optics, cooling systems, and physics.

That paper caused a stir in Silicon Valley and Wall Street. Supporters called it “the clearest analysis of this decade,” while opponents called it “the vengeful fantasy of a fired employee.” A year later, his fund returned 24 times the investment, providing the answer.

Everyone is rushing to invest in “AI companies,” including OpenAI, Anthropic, xAI, with valuations exceeding $2 trillion combined.

What Aschenbrenner sees is another layer: regardless of which company ultimately develops AGI, they all need the same things. Electricity, more electricity. Data that runs on fiber. Training datasets filling SSDs. Cooling systems for a million GPUs.

He turned the paper into an investment thesis. In 2024, SALP was established, with a 13F filing in Q4 showing about $225 million in holdings. By Q4 2025, it had grown to $5.52 billion.

Wall Street fund managers take thirty years to build that scale; he did it in twelve months. And he’s only 25, roughly the age most people start their first full-time job.

He buys the electricity bill for AI

Opening SALP’s 13F holdings, you won’t find NVIDIA. You won’t find Microsoft. You won’t find Google, Meta, Amazon.

What you do find are:

• Bloom Energy (BE)—fuel cells, powering AI data centers
• Core Scientific (CORZ)—former Bitcoin mining farms, now running AI data centers
• IREN—former Bitcoin farms, transitioning to high-performance computing
• Bitdeer (BTDR)—former Bitcoin farms, transitioning to AI compute
• Cipher Mining (CIFR)—former Bitcoin farms, transitioning to HPC
• SanDisk (SNDK)—storage devices, the kind needed for PB-scale AI training
• Intel (INTC)—yes, the Intel that’s been mocked by Wall Street for five years
• Coherent (COHR), Lumentum (LITE)—optical components, fiber connections between data centers
• Applied Digital (APLD)—AI data center developer
• CoreWeave (CRWV)—GPU cloud service, this year’s most watched IPO

See the logic now?

He’s not buying “companies making AI.” He’s buying “the stuff that makes AI run,” including power, cooling, storage, fiber, and data centers.

Five Bitcoin mining farms form the core of his holdings because large-scale compute facilities already built, powered, and cooling are largely Bitcoin farms—marking the start of private enterprise building large-scale compute infrastructure.

CoinShares reports that mining companies have collectively raised over $70 billion to transition into AI. What’s truly notable is that a 25-year-old former crypto industry player had already turned this trend into a portfolio allocation before most institutional investors.

If AI is a gold rush, Aschenbrenner isn’t buying shovels; he’s buying water and electricity.

SanDisk up 17%, CoreWeave down 11%

At yesterday’s close, SALP’s holdings shifted again, with the portfolio diverging.

Gainers: SanDisk +16.6%, Intel +13.96%, IREN +7.65%, Coherent +5.03%, Core Scientific +2.5%, Lumentum +1.26%, Bloom Energy +0.92%.

Losers: CoreWeave -11.4%, Bitdeer -2.67%, Applied Digital -0.67%, Cipher Mining -0.65%.

The biggest winner is the “most boring” company: SanDisk, which started with hard drives, and Intel, the CPU in your laptop as a kid. The biggest loser is the “sexiest”—CoreWeave, built on NVIDIA GPUs, just completed Wall Street’s most watched IPO of the year, and dropped 11.4% in a single day.

The market is doing what Aschenbrenner’s 165-page paper predicted long ago: re-pricing the bottlenecks of AI. Not the GPUs themselves—that’s CoreWeave’s business—but everything that keeps GPUs running. Power. Cooling. Storage. Fiber.

Buying 100k H100s isn’t hard. The hard part is finding a place that can supply 100MW of power, with no outages, cooling, and enough fiber bandwidth to connect the compute.

A paper written at 25, turning into a portfolio at 26.

The day OpenAI kicked Aschenbrenner out, they probably thought his story was over.

But the person they kicked out is now laying cables for the company that kicked him out, and collecting tolls.

FAQs

What is the Situational Awareness LP (SALP) fund?

SALP is an AI infrastructure-focused fund founded by former OpenAI researcher Leopold Aschenbrenner in 2024. It doesn’t buy AI model companies but focuses on power supply, data centers, semiconductors, optics, and Bitcoin farms transitioning into HPC. Its public holdings reached $5.52 billion in Q4 2025.

Why doesn’t SALP hold NVIDIA?

Aschenbrenner’s 165-page paper argues that the bottleneck for AGI is in physical infrastructure, not GPUs themselves. Therefore, the fund sold off NVIDIA and shifted into power (Bloom Energy), storage (SanDisk), fiber optics (Coherent), and other foundational infrastructure enabling GPU operation.