- The Aschenbrenner Fund's Q1 2026 13F just dropped — and the AGI bull is net short Nvidia and the SMH semiconductor ETF, with $8.46B in notional put exposure against the AI chip stack.
- The disclosed longs (Bloom Energy $879M, Sandisk $724M, CoreWeave $556M) all express the same idea: the AGI bottleneck is shifting from silicon to power, memory, and neoclouds.
- The $13.7B headline is notional gross exposure, not managed capital. Actual AUM is materially smaller — the fund is leveraged via puts, so retail readers should treat the filing as a thesis map, not a copy-trade.
The man who wrote the 165-page essay arguing AGI arrives by 2027 just bet $8.5 billion (notional) against the chip stocks people are buying because they believe him. That’s the bombshell inside Situational Awareness LP’s Q1 2026 13F filing (filed May 18, 2026) — the Leopold Aschenbrenner fund’s first SEC disclosure that the AGI true-believer has flipped from semi-long to net semi-short across the entire AI chip stack: $1.57B notional NVDA puts, $2.04B against the VanEck SMH semiconductor ETF, $1.07B Oracle, $1.01B Broadcom, $969M AMD, plus smaller put exposures on Micron, TSMC, and ASML. At the same time, the disclosed longs read like a different thesis entirely: Bloom Energy at $878.7M, Sandisk at $724.4M, CoreWeave at $556.1M. Power, memory, neocloud. Not chips.
Bloom Energy — ticker BE — last printed $307.88 going into publication on May 23, 2026, and the trailing-twelve-month return tells you why Aschenbrenner sized it as his largest single long: the stock has gone from a back-of-the-pack hydrogen story to one of the cleanest data-center-power plays in public markets. The rest of this piece unpacks the rest of the book, the pivot away from chips, and what retail investors can actually do with a 13F that’s leveraged 7–35× gross-to-net.
Who is Leopold Aschenbrenner
Aschenbrenner was born in Germany around 2001 and entered Columbia University at fifteen. He graduated at nineteen as class valedictorian with a double major in economics and mathematics-statistics. While at Columbia he co-founded the campus effective-altruism chapter — a detail worth knowing because the fund’s investor base traces back to that community.
He joined OpenAI’s superalignment team in 2023 and was fired in April 2024 over what OpenAI described as an information leak, a characterisation he disputes. By his own account, the firing followed a memo to the board warning that OpenAI’s security posture left the company exposed to industrial espionage from foreign actors. The dispute became public in June 2024 when he released Situational Awareness: The Decade Ahead, a 165-page, 50,000-word essay arguing that AGI is roughly four years away and superintelligence about six. The essay went viral in tech and finance.
Situational Awareness LP launched in September 2024 with roughly $225 million in seed capital, co-managed by Aschenbrenner and Carl Shulman — the long-time AGI-safety researcher previously at Oxford’s Future of Humanity Institute. The disclosed equity book grew to about $5.52 billion by the end of 2025, an extraordinary expansion driven by both new capital and the fund’s concentrated positioning in AI-adjacent names. Seed backers include Patrick and John Collison of Stripe, Daniel Gross, Nat Friedman, and Graham Duncan of East Rock Capital.
What makes the fund unusual is the public legibility of the thesis. Most hedge fund managers won’t tell you what they think; Aschenbrenner published the whole argument as a free PDF before raising a dollar. That makes the 13F readable as a forecast — the longs and shorts are the bet on his own essay.
The 2027 thesis in plain English
The essay’s core argument is simpler than its length suggests. Compute keeps scaling. Models keep getting smarter as compute scales. Therefore “drop-in remote worker” AGI — an AI agent that can do most knowledge work end-to-end — should arrive around 2027, with superintelligence following inside the decade. The bottleneck in this story is not chips. It’s the physical infrastructure required to power them.
Aschenbrenner estimates that scaling current model architectures to AGI requires datacenter clusters drawing tens of gigawatts — an order of magnitude beyond what the largest existing hyperscaler facilities consume. That means new natural-gas turbines, restarted nuclear plants, behind-the-meter fuel cells, and a grid build-out the United States hasn’t seriously attempted since the 1960s. He frames it as industrial mobilisation comparable to the Manhattan Project, with US national security riding on whether American capital can move faster than Chinese state planning.
If you believe the thesis, the investable corollary is that chip demand stays voracious through 2027 — but the bottleneck pricing power migrates upstream to whoever supplies the electrons, the high-bandwidth memory, and the GPU rental capacity that’s already paid for. The 13F is a direct expression of that view.
Inside the Aschenbrenner Fund’s Q1 2026 book
The Q1 2026 13F (filed May 18, 2026) covers positions held as of March 31, 2026. The fund’s total notional disclosed exposure expanded from $5.52 billion at the end of 2025 to $13.7 billion at March 31 — a 148% jump in one quarter, almost entirely driven by the new put leg layered on top of an already-concentrated long book. The largest disclosed long exposures break down as follows.
- Bloom Energy (BE) — $878.7M (6.49M shares) plus $55M call-option notional. Solid-oxide fuel cells; on-site power generation that bypasses the utility-grid bottleneck entirely. The largest single long, with extra bullish leverage layered on via calls.
- Sandisk (SNDK) — $724.4M. NAND flash memory, spun off from Western Digital in early 2025. Underloved in the broader semiconductor rally; Aschenbrenner’s bet is that AI-training storage hits the same shortage curve high-bandwidth memory hit in 2024.
- CoreWeave (CRWV) — $556.1M. “Neocloud” GPU rental specifically built for AI workloads. The disclosed long says the fund believes hyperscaler-owned GPUs alone will not be enough to meet training and inference demand.
- Bitcoin miner / HPC-pivot cluster — aggressively expanded. Aschenbrenner more than doubled or tripled positions across the cohort of miners pivoting to AI hosting on their permitted gigawatts of grid capacity. CleanSpark (CLSK) went from 1.64M to 12.28M shares; Riot Platforms (RIOT) from 6.17M to 11.50M; Bitfarms (BITF) from 6.90M to 19.88M; Bit Digital (BTDR) from 1.79M to 3.44M; HIVE Digital is a new 3.39M-share entry. The pattern says the fund views these names as power-and-grid plays sold to the market as crypto miners.
The shorts — almost all expressed as put options — are larger in notional dollars than the longs.
- SMH (VanEck Semiconductor ETF) — $2.04B notional puts (5.33M shares). The fund’s largest single bearish position. Taking the ETF rather than a single name is how you express a thesis on chip-multiples broadly without picking which name re-rates first.
- Nvidia (NVDA) — $1.57B notional puts (8.99M shares). The single-name bombshell. The company every long-AI fund owns is the company the fund’s second-largest single put position is against.
- Oracle (ORCL) — $1.07B notional puts (7.29M shares). Heaviest AI-infrastructure-narrative beneficiary outside the GPU designers. The put says Aschenbrenner thinks the multiple has run ahead of the actual hyperscaler-capture story.
- Broadcom (AVGO) — $1.01B notional puts (3.25M shares). Custom silicon for Google, Meta, and others; the bear case is that hyperscaler in-housing accelerates and AVGO’s content-per-system tops out earlier than the market expects.
- AMD — $969M notional puts (4.76M shares). The MI300/MI325 story; the puts implicitly say MI-series gross margins disappoint or the share-gain narrative slips.
- Micron (MU) — $584M notional puts. The HBM cohort — striking, because the fund is simultaneously long SNDK on the NAND side. Reads as a paired memory trade: long unloved storage, short the consensus HBM beneficiary.
- TSMC (TSM) — $535M notional puts. The foundry.
- ASML — $494M notional puts. EUV lithography — the upstream chokepoint of the whole stack.
The single most dramatic position change quarter-on-quarter was Intel (INTC). In Q4 2025 the fund held 20.24M shares of INTC long plus $747M of call-option notional — a heavily bullish wager. In Q1 2026 that flipped: equity holding cut to 202k shares (a 99% reduction) and a fresh $159M put-notional position layered on top. That’s a complete reversal on a single name in one quarter, and worth flagging because it shows the fund will pivot hard when its read of the catalyst window changes.
Total notional put exposure against the chip stack: $8.46B. The full 13F-HR is publicly filed on SEC EDGAR.
Critical context the aggregator headlines miss: the $13.68B “AUM” figure circulating in the financial press is notional gross exposure, not managed capital. Actual managed AUM is materially smaller — reported variously across sources at around $383M cash AUM (per MLQ Research’s February 2026 note) up to roughly $1.5B per Fortune’s October 2025 reporting, with a disclosed long-equity book of about $5.5B by end-2025. The differences are metric mismatches, but the takeaway is the same: the fund is leveraged roughly 7–35× gross-to-net via deep put exposure and concentrated longs. Reading the 13F as if Aschenbrenner controls $13.7B of cash is a misreading.
Why the AGI bull just shorted Nvidia
The obvious contradiction — long AI thesis, short the AI chip leader — admits three plausible readings.
Reading one: valuation is the trade. NVDA trades above twenty times trailing sales going into mid-2026, and the SMH ETF is up substantially over the prior eighteen months. The fund can believe AGI is real and believe the chip stocks are pricing in five more years of monopoly margin. The expected demand is already in the price; the unexpected demand is in the power and memory cohort that hasn’t re-rated yet. This is a valuation trade, not a thesis reversal.
Reading two: the bottleneck shifted. The essay itself argues power is the binding constraint by 2027, not silicon. If that is right, the marginal dollar of AI capex flows to fuel-cell installers, gas-turbine builders, nuclear-uprate developers, and neoclouds that already own GPU capacity — not to the chip designer whose product is fully sold out for the next two years. BE, CRWV, SNDK are all consistent with that read. The puts on chip stocks are an explicit statement that the next leg of return is somewhere else.
Reading three: it’s a hedge. A leveraged hedge fund running a concentrated long-AI book carries a real tail risk if the AI capex narrative cracks. $8.5B of notional puts against the chip basket buys explicit downside protection. Many funds that look bullish actually run net-flat exposure through put structures — the disclosed longs grab the upside and the puts hedge the disaster scenario.
The most useful interpretation combines all three. The essay’s thesis is consistent with shorting overvalued chips, going long power and memory and rental capacity, and running the chip puts as portfolio insurance against the disaster outcome where the AI trade unwinds before AGI actually shows up. None of those are mutually exclusive. The 13F is just what an internally-consistent expression of the Situational Awareness essay looks like on a brokerage statement.
Where the money flows now
The fund’s disclosed positions map to investable buckets retail can actually access. The mapping is the useful part; the actual position sizing is harder to copy because Aschenbrenner is using leverage and concentration that most retail accounts cannot replicate.
- Power and grid: the fund’s BE long is the headline expression, but the broader theme — on-site power for AI loads — reaches Vistra (VST), Constellation Energy (CEG), Talen Energy (TLN), GE Vernova (GEV), and increasingly the small-cap fuel-cell cohort including FuelCell Energy. The nuclear-power AI thesis is part of the same setup — SMR developers and uprate plays sit in the same investable bucket.
- Memory: SNDK is the disclosed long, but the broader cohort includes Micron (MU) on the HBM side, and non-US names SK Hynix and Samsung. The trade thesis is that AI-training storage and inference-side memory hit the same shortage that high-bandwidth memory hit in late 2023.
- Neoclouds: CRWV is the disclosed long; this is the smallest of the three clusters by public-company count because most of the comparable companies (Lambda, Voltage Park) are private. The bet is that GPU rental capacity already paid for by neoclouds is a more efficient way to ride training-cycle demand than buying the chip designer.
- The shorts retail can express: the SMH ETF and NVDA can both be shorted by retail via inverse ETFs (SOXS, NVDS) or via puts on a brokerage that supports options. Reality check on this: theta decay and premium cost make short positions expensive to maintain — running an active short for twelve months can cost 4–7% of notional even when the stock goes nowhere. Aschenbrenner is paying that cost too; he just sized for it. Worth knowing before mimicking. The SMH ETF Aschenbrenner just shorted with $2B notional puts is the same basket retail readers have been buying for AI exposure — the irony is part of the story.
The single biggest caveat: 13F filings have a 45-day reporting lag. The positions disclosed are point-in-time as of March 31, 2026. By the time anyone reads the filing in May, the fund may have flipped again. Treat the disclosure as a snapshot of the thesis, not a real-time map of the positions.
What kind of stocks does Aschenbrenner buy
Backing out a financial fingerprint from the holdings cohort gives a clean picture of the archetype. Across the disclosed longs the median capex-to-revenue ratio is roughly 0.38 — about ten times the S&P 500 baseline. The median price-to-sales sits around 10.5×. Revenue growth runs roughly +25% year-over-year. Return on equity is low at about 3.8%, and debt-to-equity is elevated at 0.85.
Translation: he buys capital-intensive growth at premium prices. These are companies pouring most of their operating cash flow back into physical capacity — data centers, fuel cells, memory fabs, GPU clusters — with the expectation that revenue scales faster than capex over the next several years. The low ROE is a feature, not a bug; cash isn’t being returned because every dollar of free cash has higher-return capacity to fund. The high D/E is a feature too; the buildout is debt-financed because the asset base supports it.
This is the structural opposite of Buffett-style value investing. Berkshire’s archetype is high-ROE, low-capex, cash-generative businesses bought at modest multiples. Aschenbrenner’s archetype is low-ROE, high-capex, growth-trajectory businesses bought when the market hasn’t yet priced the buildout. The two strategies are not in opposition — they’re answering different questions about where return comes from. Buffett bets on compounding distributable cash. Aschenbrenner bets on the unit-economics of future distributable cash, after the buildout phase ends.
For retail readers, the practical implication is that the fund’s longs will look “expensive” by traditional value metrics — high multiples, low current ROE, leveraged balance sheets. That isn’t an accident. It’s the screening criterion.
What could prove this wrong
The AGI timeline slips. If a GPT-7-equivalent doesn’t show drop-in-worker capability by 2027, the whole investment thesis decompresses. The power names lose their premium bid because the gigawatt buildout case weakens. The chip-short payoff also shrinks — not because chip demand collapses, but because the multiple expansion the fund is shorting against doesn’t crack. Both sides of the book underperform.
NVDA earnings keep beating — but the multiple cracks anyway. Q1 FY2027 results dropped on May 20, 2026 and Nvidia delivered: revenue of $81.6 billion (+85% year-over-year), data-center revenue +92% YoY to about $75 billion, EPS of $1.87 against a $1.76 consensus, and a fresh $80 billion buyback authorisation. The stock fell roughly 1.75% on the print. That sequence — huge beat, stock down — is exactly the dynamic Aschenbrenner’s puts are positioned for: the issue isn’t demand, it’s that demand was already priced in. If hyperscaler capex stays above $700 billion annually through 2027 (see the hyperscaler capex thesis we mapped previously) and Nvidia keeps delivering, the short still bleeds on theta and any sustained re-rally. Theta plus spot drift can cost 4–7% of notional per year even when the underlying moves sideways.
Power capex hits permitting or grid bottleneck. Bloom Energy specifically depends on natural-gas-adjacent fuel cells; CEG and VST depend on nuclear restarts and uprates. Permitting delays, interconnection queue blocks, or grid-stability issues could mean the gigawatt buildout slips from 2027 to 2030. That timeline drift would compress the BE thesis more than the SNDK or CRWV thesis.
The leverage cuts both ways. $13.7B notional on managed capital reported between $383M and $1.5B is a gross-to-net ratio that magnifies returns in both directions. A drawdown of just a few percent on gross exposure can become a margin-call scenario at fund level. Aschenbrenner is young, brilliant, and unproven in a sustained bear market. The track record — +61.3% trailing-twelve-month, +47% versus S&P in the first six months — was earned in a market that mostly rewarded AI exposure. The next regime will test whether the framework or the regime was the source of returns.
What to watch next
- Q2 2026 13F in August. Whether the SMH and NVDA shorts stay on tells you whether the pivot is structural or tactical.
- Public commentary from Aschenbrenner. The fund moves when he speaks. Track interviews, essay updates, and any successor to the original Situational Awareness paper.
- BE, SNDK, CRWV earnings cadence. The long thesis either plays out in quarterly results or it doesn’t. Watch capacity utilisation at Bloom, AI-training NAND demand at Sandisk, and customer-concentration risk at CoreWeave.
- NVDA price action since the May 20 print. The stock fell on a huge beat — the most important data point for the chip-short thesis since the fund went live. Watch whether the multiple compression continues into the August quarter or whether the buyback floor catches it.
The bottom line
Situational Awareness LP exists in part because Aschenbrenner published his thesis in plain English before raising capital. The fund is the bet on that thesis, and the 13F is the cleanest public expression of an AGI-aware investment framework that retail has access to without buying private-market exposure. The shorts — chip puts at $8.5B notional — are the noise in the picture; reasonable people can read those three ways. The longs — Bloom Energy, Sandisk, CoreWeave, plus the bitcoin-miner cohort — are the signal. They map onto power, memory, and rental capacity. The names are accessible; the underlying thesis is documented in a free 165-page essay; the manager’s track record is real, if short.
The discipline retail needs is to treat the Aschenbrenner fund’s filings as a map of what is being argued, not a real-time map of what is owned. The leverage is significant. The positions can flip by next quarter. The thesis can be wrong. None of that changes the fact that this is the most legible long-AGI bet outside private markets — and the chip-short is the surprise that made the Q1 2026 13F worth reading in the first place.
Get early access to Orbit
Orbit is Luna3.ai’s AI-augmented research engine. 12 algorithmic signals + a gradient-boosted ML model + an agentic LLM that reads each top pick’s filings and writes a daily thesis with conviction score and catalyst proximity. Three regimes, three playbooks — growth in expansion, defensives in late-cycle, recovery plays at panic bottoms. The 3 in Luna3.ai.
No spam. Unsubscribe any time.
No comments yet. Be the first to share your thoughts!