# Leopold Aschenbrenner Has No Nvidia and a $5.5 Billion AGI Portfolio
> We rebuilt five quarters of 13F filings from the Situational Awareness fund: $255M to $5.5B in a year, zero Nvidia longs, and a bet that AGI runs on power.
- Author: Barebone Research, Barebone AI
- Published: 2026-03-09
- Canonical: https://barebone.ai/resources/leopold-aschenbrenner-agi-portfolio-power-not-chips
- Publisher: Barebone AI (https://barebone.ai)
---
## The Manifesto That Became a Hedge Fund
In April 2024, OpenAI fired a young researcher from its superalignment team — the group tasked with keeping future superhuman AI systems under control — over an alleged information leak. He disputes that account. He was two years out of Columbia, where he had graduated as the 2021 valedictorian at age 19.
Two months later he published a 165-page PDF titled *Situational Awareness: The Decade Ahead*. It argued that AGI — AI that can match or beat humans at most cognitive work — is "strikingly plausible" by 2027, and that almost everyone was underestimating what building it would actually require. Not better algorithms. Power plants.
Silicon Valley read the essay as a manifesto. Leopold Aschenbrenner treated it as a prospectus. Within months he had launched a hedge fund named after the PDF, seeded by Stripe founders Patrick and John Collison, Nat Friedman, and Daniel Gross. He is now 24.
We used Barebone AI to pull every quarterly 13F the fund has filed with the SEC — five filings, from its first in February 2025 through the latest on February 11, 2026 — rebuilt the book position by position, and repriced the names at the March 9 close.
Three headline numbers. The disclosed US book grew from **$254.8 million to $5.52 billion** in twelve months. Across all five filings, the fund has reported **zero shares of Nvidia** — the only Nvidia position it ever disclosed was a bet *against* the stock. And **25% of the book** now sits in current and former bitcoin miners.
## The Thesis: AGI Is an Industrial Project
The essay's core argument is a ladder. Each new frontier AI model has consumed roughly an order of magnitude — ten times — more compute than the last, and Aschenbrenner argued the ladder would keep extending through the decade. Follow it out a few rungs and the numbers stop being a software story:
| Year | Training cluster | Power draw | Cost | Power equivalent |
|------|-----------------|------------|------|------------------|
| 2022 (GPT-4) | ~10k H100-equivalents | **~10 MW** | ~$500M | ~10,000 homes |
| ~2024 | ~100k | ~100 MW | $ billions | ~100,000 homes |
| ~2026 | ~1M | **~1 GW** | $10s of billions | the Hoover Dam |
| ~2028 | ~10M | ~10 GW | $100s of billions | a small US state |
| ~2030 | ~100M | **~100 GW** | $1T+ | >20% of US electricity |
That last row is the whole thesis. A 100 GW cluster running continuously for a year consumes 876 terawatt-hours; total annual US electricity production is about 4,250 TWh. The essay put it bluntly:
> "The talk of the town has shifted from $10 billion compute clusters to $100 billion clusters to trillion-dollar clusters."
If even half of that ladder gets built, the scarce input stops being GPUs — which the supply chain eventually learns to print in volume — and becomes electrons: generation capacity, grid connections, and buildings already wired for megawatt-scale loads. Chips are a manufacturing problem. Power is a permitting problem, and permitting problems take a decade.
That repricing of the bottleneck is the entire fund.
## Five Filings, One Direction
The 13F record — the quarterly snapshot of US long positions every large manager must file — reads like the essay being converted into positions in real time.
| Filing | Filed | Positions | Disclosed value |
|--------|-------|-----------|-----------------|
| Q4 2024 | Feb 11, 2025 | 6 | $254.8M |
| Q1 2025 | May 14, 2025 | 12 | $1.01B |
| Q2 2025 | Aug 14, 2025 | 9 | $2.12B |
| Q3 2025 | Nov 14, 2025 | 28 | $4.14B |
| Q4 2025 | Feb 11, 2026 | 29 | **$5.52B** |
The first book, just six names, was already two-thirds power and cooling: Marvell at 34.1%, then Vistra (23.2%), Vertiv (20.3%), Talen (11.0%), Constellation (8.5%), and Modine (3.0%). Note what those power names are — Vistra, Talen, Constellation. The consensus "AI power trade" of 2025. He owned it in 2024, and by the end of 2025 he had exited all three.
Then the book started doing stranger things. In Q1 2025, calls on Intel were **45.7%** of the entire filing, and CoreWeave appeared in the same quarter it went public. In Q2 2025, the largest line was not a stock at all: **puts on SMH**, the VanEck Semiconductor ETF, at 26.9% of the book — long power, short chips, in the same portfolio. By Q3 the put list had grown to include Broadcom, TSMC, Micron, and Nvidia itself.
The one public performance number we have: the fund returned **47% after fees in the first half of 2025**, as the Wall Street Journal reported in August — against roughly 6% for the S&P 500.
The book did not just grow. Each quarter, it moved further from wherever the crowd had just arrived.
## The Nvidia Position That Never Existed
The viral version of this story says Aschenbrenner "dumped every share" of his Nvidia. The filings say something cleaner: there was never a share to dump.
In five quarterly filings covering December 2024 through December 2025, Situational Awareness LP has never reported a single Nvidia share held long. The only Nvidia position in the record is in the Q3 2025 filing: **put options covering 1.6 million NVDA shares**, roughly $300 million of underlying value — 7.2% of the book, sitting alongside those puts on Broadcom, TSMC, Micron, and the semiconductor ETF.
By December 31, every chip put was gone. The lone put remaining in the Q4 book is on Infosys — an $8.9 million wager that AI eats outsourced IT services, which may be the most quietly brutal position in the portfolio.
Two honest caveats before reading too much into this. A 13F shows only US long positions and options; short sales and swaps are invisible, so we see half the portfolio at best. And the first filing covers December 2024 — anything the fund did in its earliest months is dark. What we can say precisely: in the entire public record, the only disclosed Nvidia bet was a bet that it would go down.
This is not an anti-AI position. It is more specific than that: chips are where the AI trade is crowded, and power is where it is constrained.
## Three Layers of the Bottleneck Trade
The Q4 2025 book sorts almost perfectly into three layers — who makes the power, what runs the compute, and where the machines live.
| Layer | Names | Value | % of book |
|-------|-------|-------|-----------|
| Compute & connectivity | CRWV, INTC, LITE, SNDK, COHR, TSEM | $2.86B | 51.8% |
| Converted bitcoin miners | CORZ, IREN, APLD, CIFR, RIOT, HUT, WYFI, BTDR, CLSK, BITF | $1.38B | 25.0% |
| Power & fuel | BE, EQT, SOI, PSIX, BW, LBRT, PUMP | $1.22B | 22.1% |
| Everything else | KRC, INFY puts | $0.06B | 1.1% |
**The power layer** is led by Bloom Energy — the largest single line in the book at $875.5 million of stock, **16.5%** combined with its calls. Bloom builds solid-oxide fuel cells that generate power on site, which matters because of one number: Bloom says it can deliver power to a data center in as little as 90 days, and signed a 2025 deal with Oracle built around exactly that promise. A new grid interconnection — permission to plug a gigawatt load into the utility network — routinely takes two to five years, sometimes seven. By early 2026, Brookfield had committed $5 billion to deploy Bloom's fuel cells across AI data centers, and American Electric Power had placed a $2.65 billion order for roughly 900 megawatts. Around Bloom sit EQT (natural gas), Solaris (mobile gas turbines), and a tail of generator and gas-services names.
**The compute layer** is half the book. CoreWeave is the largest combined exposure at **22.0%** — 6.1 million shares plus calls on another 10.8 million. The Intel position is pure conviction expressed in options: calls on 20.2 million shares, **$746.8 million** of underlying value, a position held through four consecutive filings while the stock more than doubled. And Lumentum, at 8.7%, makes the lasers that move data between AI chips — as clusters scale, the optical links between GPUs become their own bottleneck.
**The miner layer** is the strange one, and the most contrarian idea in the whole portfolio. Ten current and former bitcoin miners — Core Scientific, IREN, Applied Digital, Cipher Mining, Riot, Hut 8, WhiteFiber, Bitdeer, CleanSpark, Bitfarms — together make up a quarter of the book. The logic: miners spent a decade acquiring the one asset that now takes years to permit — buildings with hundreds of megawatts of live grid connection. If the queue is the moat, miners own pre-built moats. Repurposing one into an AI data center is a retrofit; building one from scratch is a half-decade of paperwork.
The lone head-scratcher is a 0.9% position in Kilroy Realty, a San Francisco office landlord. Make of that what you will.
## Is the Market Paying the Trade?
As of tonight's close, the answer is: partially — and narrowly.
Year to date, Bloom Energy is up **+74.2%** (including a near-12% jump on March 9 alone) and Lumentum **+73.8%**, while Nvidia is *down* 2.1% and the S&P 500 is off 0.7%. Intel, his single biggest options bet, is up **+23.5%**. The divergence between the bottleneck trade and the chip trade is no longer theoretical — it is on the tape.
But look at the miner layer: Core Scientific +4.1%, IREN +2.8%, Cipher -3.0%. CoreWeave is up just 3.9% and still trades 16% below its early-February close. The market is paying the electrons and the optics — and still auditing the real estate.
A rotation that only pays two tickers is not a rotation yet. It is a stock-picking contest.
## Where This Goes Wrong
Every number above is real, and several of them are doing less work than they appear to.
**The $5.5 billion is not "returns."** Going from $254.8 million to $5.52 billion in disclosed positions is mostly new investor money plus rising markets — and 13F rules report option positions at the value of the underlying shares, which inflates the apparent size of a calls-heavy book. The only public, dated performance figure is that 47% first half of 2025. Everything since is unaudited silence.
**The track record is six quarters in one regime.** Aschenbrenner is 24, has never managed money through a down cycle, and before OpenAI he worked at FTX's Future Fund — a reminder that proximity to brilliant people and abundant capital is not the same thing as risk control. His access is also his selling point: his fiancée is chief of staff to Anthropic's CEO, and his seeders run Stripe. That vantage cuts both ways — it is exactly why everything in this article comes from public filings, not from anyone's word.
**The thesis can be right and the trade still fail — on timing.** The 1999–2001 telecom buildout laid the fiber the internet genuinely needed; traffic grew exactly as promised, and the companies that built it still went bankrupt because capacity arrived a decade before the demand curve. If AGI lands in 2031 instead of 2027, a 100 GW power buildout becomes a glut, and fuel cells trading at 2027 expectations reprice to 2031 reality.
**Copying the book gets the shape, not the timing.** Filings arrive 45 days stale, the top three issuers are **52%** of the book, and this manager trades fast — the Nvidia and semiconductor puts appeared and vanished within two quarters. Anyone mirroring the longs also carries none of the shorts that, per the fund's own described strategy, sit on the other side.
## What the Filings Are Actually Telling You
The value of this portfolio is not any single ticker. It is the migration pattern.
Trace it: chips were the consensus by 2023. Grid-scale power — Vistra, Constellation, Talen — became the consensus in 2025, and he had already owned them in 2024 and left. The 2026 book is positioned one step further upstream: behind-the-meter generation that skips the queue entirely, and miners whose grid connections are the queue, already won. Each filing shows the same investor asking the same question — *where does the bottleneck move next?* — and refusing to stand where the answer used to be.
The signals worth watching from here, none of which require believing a word of the manifesto: whether behind-the-meter power orders keep landing at Brookfield and AEP scale; whether miner-to-AI conversions produce signed hyperscaler leases or just press releases; whether interconnection queues actually shorten, which would thin the very moat this book is built on; and the fund's Q1 filing, due by mid-May, which will show whether the miner layer survived the spring.
If Aschenbrenner is wrong, this is the most articulate way anyone has ever been long a utility buildout. If he is right, the money keeps moving upstream of the chip — and the filings are where he shows you, 45 days late, which way the water is flowing.
---
*Data: Barebone | Sources: Situational Awareness LP Form 13F filings (SEC EDGAR), Situational Awareness: The Decade Ahead (June 2024), Fortune (October 8, 2025 and March 5, 2026), Wall Street Journal (August 2025) | Prices as of March 9, 2026 close*