In May 2026 many traders awaited the quarterly Form 13F from the hedge fund Situational Awareness — its founder, Leopold Aschenbrenner, a former OpenAI employee, had drawn Wall Street’s eye with one of the most profitable AI portfolios of the past two years.
The filing deadline passed on May 15th, yet no document appeared. Investors speculated that the fund had requested confidential treatment from America’s Securities and Exchange Commission (SEC), allowing it to hide positions for up to a year. More likely, the managers simply waited until the last day — standard practice for funds shadowed by thousands of traders.
On May 18th the report did land — and revealed a radical shift: Aschenbrenner bulked up on AI infrastructure while opening a sweeping short against semiconductor firms.
To see why the 24-year-old’s portfolio commands such attention, rewind two years — to a firing, a manifesto and a bet on Bitcoin miners repurposing sites into data centres.
ForkLog bears no responsibility for readers’ investment decisions.
A wunderkind from Berlin
Leopold Aschenbrenner was born in Germany to a family of doctors. He graduated from John F. Kennedy School in Berlin, enrolled at Columbia University and in 2021 became valedictorian of his class.
At 17 he received a grant from the economist Tyler Cowen via the Emergent Ventures programme. Cowen called him “an economics wunderkind”.
In 2022 Aschenbrenner joined the FTX Future Fund, the charitable arm of Sam Bankman-Fried’s crypto exchange. He witnessed the company’s collapse from the inside and left before its bankruptcy. Later he recalled:
“We were a tiny team, and in a single day everything fell apart and became associated with a giant fraud. It was incredibly hard.”
In 2023 Aschenbrenner joined OpenAI’s Superalignment team — a unit led by Ilya Sutskever and Jan Leike that worked on controlling a superintelligent AI. There he co-authored a research paper Weak-to-Strong Generalization.
In parallel, Aschenbrenner wrote a memo to OpenAI’s board warning about industrial-espionage risks from China and calling the firm’s security “egregiously inadequate”. Earlier, the NYT reported that in early 2023 a hacker broke into the startup’s internal messaging systems and stole information about its AI-technology design.
In spring 2024 Aschenbrenner was fired for leaking corporate data. He called the decision politically motivated and described the startup’s security approach as “not robust enough to protect the theft of key secrets if foreign actors penetrated the company”. A month later Sutskever and Leike left OpenAI, after which the company disbanded Superalignment.
In June 2024 Aschenbrenner published the essay “Situational Awareness”, which caused a stir across the AI industry.
Situational awareness in a hedge fund
The essay’s central claim: artificial general intelligence (AGI) will arrive by 2027, the world is unprepared, and firms building the physical infrastructure for compute are the market’s most undervalued assets.
On Dwarkesh Patel’s podcast Aschenbrenner explained the investment logic through the scale of the necessary infrastructure. In 2022 the GPT-4 training cluster drew about 10 megawatts and cost roughly $500m. AI compute, he argued, scales by half an order of magnitude each year: by 2024 the largest cluster already demanded 100 megawatts and cost billions. He sketched the trajectory:
“By 2026 — a one-gigawatt cluster, the size of a large nuclear reactor. Tens of billions of dollars. By 2028 — 10 gigawatts, more power than most states consume. By 2030 — a trillion-dollar cluster, 100 GW, more than 20% of all U.S. electricity generation. And that’s just the training cluster.”
That is only model training — inference would require several times more capacity. Aschenbrenner likened the advantage in the race for AGI to the coalition’s technological lead in the First Gulf War:
“Western forces had a 100:1 casualty ratio. Better tank sensors, more accurate missiles, GPS, stealth. A 20–30 year lead — and they just crushed the opponent. A superintelligence applied across a broad range of R&D could compress a century of technological progress into less than ten years.”
In September 2024 Aschenbrenner founded the hedge fund Situational Awareness LP. Anchor investors included Stripe’s co-founders, Patrick and John Collison, former GitHub boss Nat Friedman and Safe Superintelligence co-founder Daniel Gross. The minimum ticket is $25m, with a two-year lock-up.
According to The Wall Street Journal, in the first half of 2025 the fund returned 47% after fees — versus roughly 6% for the S&P 500. The biggest contributions came from energy and data-centre bets.
Fortune reported that by early 2026 the fund’s publicly disclosed exposure to U.S. assets had grown to about $5.5bn from roughly $225m at end-2024. To be clear: this refers not to verified net profit but to the size of positions disclosed in 13F filings — a measure that includes asset appreciation, new capital inflows and potential leverage.
What the new 13F shows
Form 13F is a quarterly report that funds with over $100m in assets must file with the SEC.
The previous Situational Awareness filing for Q4 2025 listed 29 positions worth $5.5bn with minimal options activity. The new report looks starkly different: 42 positions with disclosed exposure of $13.67bn — nearly tripled in a single quarter.
Portfolio by instrument: 66% is in put options, 10% in calls, 24% in outright equities.
The fund opened put positions on the biggest chipmakers — all new to the filing:
- VanEck Semiconductor ETF (SMH) — $2.04bn;
- Nvidia — $1.57bn;
- Oracle — $1.07bn;
- Broadcom — $1.01bn;
- AMD — $969m;
- Micron — $584m;
- TSMC — $535m;
- ASML — $494m;
- Intel — $159m.
In total — $7.46bn in puts against the semiconductor complex. None of these positions appeared in the prior filing.
At the same time Aschenbrenner lifted stakes in energy, data-centre and storage names:
- Bloom Energy — $879m (the fund’s biggest long);
- SanDisk — $724m;
- CoreWeave — $556m;
- IREN — $401m;
- Core Scientific — $389m;
- Applied Digital — $320m.
The fund also added to Bitcoin miners: Riot Platforms (+87% shares), CleanSpark (+648%), Bitdeer (+92%), Bitfarms (+188%). New names include Hive Digital and T1 Energy.
Positions fully exited include Lumentum, Hut 8, Cipher Mining, Coherent, EQT and Tower Semiconductor. One of the most striking rotations: a $747m call option on Intel replaced by a $159m put.
Another detail — call options on certain names held alongside sector-wide puts:
- Micron — $422m;
- SanDisk — $389m;
- TSMC — $355m;
- CoreWeave — $141m;
- Bloom Energy — $55m.
The mix of index puts and single-name calls suggests Aschenbrenner is not merely shorting chipmakers. He has built a stance that pays if the sector rolls over while select companies — those closest to physical infrastructure and data storage — keep rising.
What the puts don’t show
Form 13F does not disclose strikes or expiries. The sums shown are the market value of shares controlled by the options (notional value), not the price of the contracts themselves. For “far” or deeply out-of-the-money options the gap can be enormous: the fund could control $7.5bn of stock while paying a fraction of that for the puts.
Blogger Jason’s Chips flagged this nuance. On his reading, the puts may be a short-term hedge against geopolitical risk — say, an escalation with Iran — rather than an outright bearish bet on semiconductors. In that case the actual cash laid out for the puts is far smaller than the stated $7.5bn notional.
Another caveat matters. A 13F is a snapshot as of the quarter’s last day — March 31st 2026. Almost seven weeks elapsed before publication. Over that span the semiconductor sector rallied, meaning the put positions likely lost value. Aschenbrenner could have closed or added to them — by definition the current book is unknown.
Even so, the structure speaks for itself: the largest longs — Bloom Energy, SanDisk and CoreWeave — point to where Aschenbrenner sees the next bottleneck in AI.
Long shovels
The classic “picks and shovels” strategy dates to California’s gold rush: while prospectors went bust chasing ore, tool-sellers got rich. Aschenbrenner applies the same logic to the AI boom — but one layer deeper.
He is not buying model-makers like OpenAI or Anthropic. Nor is he making an unambiguous bull bet on chip designers such as Nvidia or AMD. Instead the fund invests in what models and chips cannot run without: electricity, cooling, fibre, land with grid interconnects and long-term capacity contracts.
Bloom Energy — the fund’s largest position — makes solid-oxide fuel cells that can power a data centre autonomously, without tapping an overburdened grid. Standard utility interconnection can take more than five years; a Bloom Energy installation begins delivering power roughly 90 days after it is sited.
SanDisk is a wager on NAND flash. AI inference spews data that must be stored: every processed token leans on a stack of HBM and enterprise SSDs. Demand for storage rises alongside compute, yet wins far less investor attention.
CoreWeave plays the role of a “non-cloud” for AI workloads. Contracts with Anthropic and OpenAI have turned it into the de facto third hyperscaler of GPU compute after AWS and Azure.
Analyst Jim Liu adds insider context: Aschenbrenner worked at OpenAI in 2023–2024, when the startup used CoreWeave’s infrastructure, and saw its software stack from the inside. Since then OpenAI has moved its AI infrastructure in-house and buys only bare metal from CoreWeave. But for the long tail of AI startups without the resources for their own stack, CoreWeave’s cloud remains one of the few viable options.
Miners as an AI bet
A distinct cluster in the portfolio is Bitcoin miners. These firms already own what is hardest to build from scratch: land near transmission lines, live power contracts and heavy-duty cooling. Converting such a site to AI hosting is cheaper and faster than constructing a new data centre.
Compute for neural networks yields two to five times more revenue per kWh than maintaining the Bitcoin network. According to CoinShares, the industry’s aggregate AI-hosting contracts have already topped $70bn.
IREN is a vertically integrated miner combining self-owned capacity, energy assets and a GPU cloud. Riot Platforms owns a site in Corsicana, Texas with more than 1 GW — the sort of infrastructure now worth more as AI compute than as hash rate. CleanSpark, whose position rose by 648%, remains a pure-play Bitcoin miner, but the logic is the same: cheap power plus land, with the option to switch to HPC hosting at any time.
Aschenbrenner has steadily expanded this sleeve: Bitdeer (+92%), Bitfarms (+188%), and a new stake in Hive Digital. What unites them is control of energy infrastructure — the scarcest input gating AI’s growth.
Physics versus valuations
The portfolio’s formula runs thus: longs — electricity, fibre, campuses and long-term capacity contracts; shorts — chipmakers whose valuations, in Aschenbrenner’s view, already price the best case.
For the past 18 months the fund has focused almost exclusively on power, memory, compute capacity and physical sites. That strategy has made Situational Awareness one of the world’s top-performing funds, and Aschenbrenner is sticking with it.
But the appearance of $7.5bn in puts signals this: the semiconductor names that Wall Street has bet on for two years have, in effect, priced in all the good that can happen to them. Chipmakers’ margins risk being squeezed by competition and oversupply, while the physical constraints — power, cooling, grid connections — remain unsolved.
Text: Sasha Kosovan