Aschenbrenner's Q1'26 13F — what it tells you about where the AI money is moving

I spent a few days going through Leopold Aschenbrenner's Q1 2026 13F filing alongside his four-and-a-half-hour Dwarkesh interview. Most coverage I've seen focuses on the headline (he's "shorting Nvidia") and misses what the book is actually doing. Posting my read here because a few of you have been asking what I make of it.

This is not investment advice. It's one operator's framework, laid out plainly.

Who he is

Worth grounding before reading the trade.

• German, born 2001/2002. John F. Kennedy School Berlin. Moved to the US at fifteen.
• Columbia, 2021 — valedictorian, BA in economics + math/stats, finished at nineteen.
• Worked at FTX's Future Fund (the philanthropy arm) before its collapse.
• Joined OpenAI's Superalignment team in 2023, under Ilya Sutskever and Jan Leike. Co-authored Weak-to-Strong Generalization (ICML 2024).
• Fired April 2024 over a disputed information-leak allegation. He walked away from roughly a million dollars in vested equity rather than sign a non-disparagement agreement.
• Published the Situational Awareness essay in June 2024. It went viral inside the AI and policy worlds.
• Founded Situational Awareness LP, a hedge fund, backed by Patrick and John Collison (Stripe), Daniel Gross, and Nat Friedman. Fund AUM around $1.5B.

His Q1 2026 13F shows $13.7B in reported value. That number includes option notional, so it isn't the same as AUM. It's the gross exposure his book casts onto the market. The reason that distinction matters is that most of the size came from puts overlaid on top of the long book this quarter, not from new capital.

What he's LONG (the picks-and-shovels stack)

The conviction longs cluster into three layers, all sitting underneath or downstream of the chip layer.

Energy / power infrastructure. The thesis that the actual bottleneck for a trillion-dollar AI cluster is gigawatts, not GPUs.

• Bloom Energy (BE) — $879M. Fuel cells for data-centre power. The biggest single non-derivative long in the book.
• Power Solutions International (PSIX) — ~$26M. Natural-gas engines for behind-the-meter gen-sets, the kind data centres are deploying to get power online faster than grid permitting allows.
• Babcock & Wilcox (BW), Solaris (SEI), T1 Energy (TE) — combined ~$150M. Industrial power, oilfield power, US solar + storage manufacturing.

AI compute hosting + miner pivots. The thesis that crypto miners with cheap power contracts and built grid interconnects get repriced as AI hosting facilities.

• CoreWeave (CRWV) — $556M. The template. AI cloud and GPU lessor.
• IREN (IREN) — $401M, and Core Scientific (CORZ) — $389M. Both mid-pivot from Bitcoin mining to AI hosting.
• Applied Digital (APLD) — $320M. AI data centre build.
• Riot Platforms (RIOT), Cleanspark (CLSK), Bitfarms (now Keel / KEEL), Bitdeer (BTDR), Hive (HIVE) — combined ~$320M. The option-value cluster. Same miner-to-AI-host pivot, smaller bets, broader coverage.

Storage / memory. The thesis that as inference scales out, the marginal dollar of value moves to whoever can feed the chip.

• SanDisk (SNDK) — $724M.
• Micron (MU) — $584M.
• Plus a separate $952M call book layered on top. MU, SNDK, TSM, CRWV, BE. Same names, levered upside.

The shape is consistent. Not "long AI" in the lazy sense. Long the substrate AI runs on, and concentrated below the layer where everyone else is already crowded.

What he's SHORT — broadly across the chip layer

The other side of the book is a roughly $7.74B notional put position spread across almost every major AI chip and fab name.

Important caveat that most coverage glosses over. 13F filings disclose put and call notional but not direction. They don't tell you whether the fund is long the puts (a hedge or outright short bet) or short the puts (premium harvesting, which is effectively synthetic long). The two look identical on the form.

The shape of the book, broad coverage of the entire semis complex paired with concentrated longs in the consumers of those chips, strongly suggests these are long-puts, i.e. an outright hedge against or short bet on the chip layer. But it isn't legally confirmable from the 13F alone. Worth holding that line carefully.

Worth flagging: seven tickers (NVDA, AMD, MU, TSM, ASML, INTC, GLW) appear on both sides. The small cash longs are token positions. NVDA's cash long is $498K against $1.57B of NVDA puts, roughly 3,000× smaller. The puts are the real exposure.

The macro view this trade comes out of

This is the part worth slowing down on. The book only makes sense once you see the frame underneath it.

• The real bottleneck is energy, not chips. A trillion-dollar cluster requires 100+ gigawatts. That's not a software problem or a startup problem. It's century-scale industrial mobilisation. Natural gas, small modular reactors, solar megaprojects, permitting reform. The US has not attempted physical building at that timescale since the mid-twentieth century.
• The chip layer is consensus-long with diminishing marginal returns. Everyone owns NVDA. The trade is priced. The marginal dollar of value in the AI stack is no longer flowing to the chip layer, it's flowing to whoever can deliver the power, the hosting, and the memory.
• AI beta is migrating downstream. Chips → power → hosting → storage. The fund is positioned for that migration to continue.
• The doubling of book size this quarter is the put overlay. The book went from $5.52B at year-end 2025 to $13.68B by 2026-03-31. The longs didn't double in value. The structural change is the addition of the large put overlay across the chip layer. Either as a hedge against the long book, or as an outright short call on chips. Either way, it's a position on where AI beta is moving, not a bet against AI itself.

The broader views worth knowing

These don't sit inside the 13F, but the trade only makes sense once you've read the worldview that produced it. Compressed:

• AGI is coming on a schedule set by physics and engineering. Compute scales ~0.5 OOM/year. Pair that with algorithmic unhobbling (test-time compute, agents, self-play) and AGI lands by 2027–2028. No new fundamental discovery required. He treats this as engineering, not mystery.
• The intelligence explosion is the decisive move. Once AGI exists, AI does AI research, and capability compounds on itself. A six-to-twelve-month lead in the compounding phase translates into something like decades of post-explosion advantage. His analogy is the Gulf War kill ratio.
• The race is US-China binary. He believes a CCP-led superintelligence, paired with perfect surveillance + lie detection + engineered loyalty, locks in a form of authoritarianism with no historical precedent. He doesn't treat that outcome as morally equivalent to a US-led one.
• The national security state will inevitably take over. Private labs cannot defend weights against state-level espionage. Current lab security is "startup-grade against state-level adversaries." A Manhattan-shaped government project is the least-bad endgame.
• Atoms-for-Peace with China, but only from a lead. The endgame is a deal. US gets there first, locks down the secrets, then offers Beijing a stable settlement. Parity is the most dangerous configuration; daylight is what makes negotiation possible.
• Secrets beat compute. A genuine algorithmic lead is worth 10–100× more than another data centre.
• Open-source frontier collapses once algorithms stop being published, which they will, because national security requires it. "Open source is betting on stagnation."
• The American constitutional order is load-bearing. He treats it as a five-hundred-year inheritance that survived nuclear weapons and can survive this too, if it isn't traded away for whichever CEO happens to be holding the model. This is the softer civic argument underneath everything else.

His epistemic posture is unusual. He's genuinely humble on exact dates. Six months, two years, who knows. But he's fully committed to the trajectory. The curve is the curve.

How I'd hold this

A few things worth holding in tension when reading the trade.

• The framework is unusually coherent. Every claim does work the others depend on. That's rare in macro takes, and it's the main thing that makes him worth reading.
• The China view rests on historical analogy, not firsthand exposure. He hasn't spent meaningful time in China. His model of authoritarian state behaviour comes from German family history (great-grandmother lived through the Nazi era, then East German communism, the Stasi) and Cold War Soviet material. Whether the Soviet analogy maps cleanly onto contemporary China is a fair question.
• The trade is his thesis with money on it. If AGI happens this decade, human capital, including his own, depreciates toward zero, and the only thing that retains value is financial capital positioned in front of the wave. If it doesn't, the fund underperforms but he's still young. Both branches of his life point at the same conviction.
• The 13F is a backward-looking snapshot. It's the book as of 2026-03-31. By the time you read it, positions have moved. Use it as a frame, not a copy trade.
• The put direction is not legally confirmable. I keep flagging this because most secondary coverage doesn't. Read carefully.

The reason I find this one worth thinking about isn't that I'm certain he's right. It's that he's arranged his own life around the framework, which is a more demanding kind of seriousness than most commentators offer. Whether the framework holds is a different question. Either way it's a sharper instrument than the consensus take, and worth running your own positions past.

If you want to read the source material: the 4.5 hour Dwarkesh interview is on YouTube, the 13F is on SEC EDGAR under CIK 0002045724 (accession 0002045724-26-000008), and the Situational Awareness essay from June 2024 is the public version of the worldview.

Happy to talk about any specific line of this in person.