Aschenbrenner’s Thesis vs His Fund’s Q3 Book: Power Is 25%

In June 2024 Leopold Aschenbrenner published Situational Awareness: The Decade Ahead, an essay series that wrote: “Probably the single biggest constraint on the supply-side will be power.” The September 2026 13F of his hedge fund, Situational Awareness LP, reports $1.58 billion.

We grouped every row of that filing under the essay’s themes. By our classification, power is 25.4% of the book, GPU compute 49.0%, memory 20.5% and data-centre site operators 5.1%. The grouping is ours, not the fund’s, and the filing does not say why any position was held.

Business Pill · GPUs

An evergreen Business Engineer explainer on why AI runs on GPUs. It teaches the general idea only and does not discuss Situational Awareness or any fund.

The key insight: As we read the essay next to the filing, the theme the essay ranks first, power, is 25.4% of the September book by our classification, behind GPU compute at 49.0%; in the fund’s first 13F, power producers were 42.7%.

What the Essay Says

The series is about the path to AGI; its third chapter, Racing to the Trillion-Dollar Cluster, is about money and hardware. Its summary says “many trillions of dollars will go into GPU, datacenter, and power buildout before the end of the decade”.

It puts the scale in electricity: “By the end of the decade, we are headed to $1T+ individual training clusters, requiring power equivalent to >20% of US electricity production.” It ranks the constraints too. Chips, it says, are “likely a smaller constraint than power”, while “HBM memory (for which demand is enormous) are already key bottlenecks for the current AI GPU scaleup”.

It also says “What any compute guy is thinking about is securing power, land, permitting, and datacenter construction.” The author adds that the essay draws on “publicly-available information, my own ideas, general field-knowledge, or SF-gossip”.

Situational Awareness LP's $1.58 billion 30 September 2026 13F, grouped by the themes of Leopold Aschenbrenner
Situational Awareness LP’s $1.58 billion 30 September 2026 13F, grouped by the themes of Leopold Aschenbrenner’s June 2024 essay series: GPU compute 49.0%, power 25.4%, memory 20.5%, data-centre sites 5.1%. The grouping is our classification; calls are valued at the underlying shares. Sources: SEC 13F-HR; situational-awareness.ai.

Power: 25.4% of the Book

Two holdings fit the power theme by our classification. Bloom Energy calls, six rows valued at $389.2 million on the underlying shares, and Fermi shares at $11.2 million together come to $400.4 million. Fermi’s 10-Q says it is “developing a utility-scale and utility-grade private-grid powered campus for AI-centric customers”.

Power was a bigger share at the start. The fund’s first 13F, for 31 December 2024, held Vistra, Talen Energy and Constellation Energy worth $108.7 million, 42.7% of $254.8 million, by our arithmetic. The rest was Marvell, at 34.1%, and Vertiv and Modine, at 23.3%.

Compute: 49.0%

The largest group is GPU cloud capacity: SharonAI shares at $440.7 million and Nebius calls at $332.6 million, $773.3 million in all. SharonAI’s 10-Q describes it as “an Australian neocloud operator”. The fund has filed Schedule 13D or 13G ownership reports on both companies, as our read of its 13F history shows. Matching them to the essay’s cluster theme is our classification.

Memory: 20.5%, All Calls

Sandisk calls at $182.7 million and calls on a Roundhill memory ETF at $140.3 million make $323.0 million. In June, memory was the largest part of the book: Sandisk and Micron shares were $11.25 billion, 55.6% of $20.24 billion. By September the Micron row was gone, as our first read of the filing found, and the memory exposure that remains is in call rows.

The essay’s memory passage names HBM. Whether these Sandisk and ETF positions track that part of the market is not shown by the filing.

Calls Change the Picture

Calls are $1.04 billion of the $1.58 billion. All the Bloom, Nebius, Sandisk and Roundhill exposure is in call rows, valued at the shares behind them, as our read of the call rows sets out. The 13F shows no strike, expiry or premium, so the bucket shares measure exposure, not money invested.

The Structural Read

As we read the rows, most of the essay-linked exposure sits in calls: all of the Bloom Energy, Nebius, Sandisk and memory ETF positions are call rows, valued at the shares behind them, so the bucket shares measure exposure, not money invested.

As we read the classification, the essay is a set of arguments about AI scaling, not a list of stocks; the filing shows positions on 30 September, and any link between the two is ours to draw and the fund’s to explain.

Leopold Aschenbrenner, Situational Awareness (June 2024)

“Probably the single biggest constraint on the supply-side will be power.”

Three Implications

FOR READING THE BOOK BY THEME By our classification: GPU compute $773.3 million (49.0%), power $400.4 million (25.4%), memory $323.0 million (20.5%), data-centre sites $80.8 million (5.1%).

FOR THE POWER THEME OVER TIME Power producers were 42.7% of the first 13F in December 2024 and power is 25.4% of September 2026, by our arithmetic.

FOR THE MEMORY THEME Memory went from $11.25 billion of Sandisk and Micron shares in June, 55.6% of the book, to $323.0 million of calls in September.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Memory Chokepoint.

The framework argues that the binding constraint on AI scaling has shifted to High Bandwidth Memory, and says: “Memory is becoming the new compute.”

As we read it alongside the filing, the essay and the framework both name HBM as a bottleneck, while the fund’s memory exposure fell from 55.6% of its June book to 20.5% in September, all of it in calls. The framework does not discuss Situational Awareness or its filings; the parallel is ours.

What Is Not Established

The filing does not show why the fund holds any position, whether the essay guides its choices, or what it holds today. Reuters reported on 30 July, citing two sources, that the fund sold most of its stock portfolio to Citadel; the fund did not comment in that report. The buckets and percentages are our classification and arithmetic, and a 13F leaves out shorts, private holdings and non-US listings, as our 13F guide explains.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Memory Chokepoint

The Business Engineer framework on why memory became a binding constraint on AI scaling.

Read the Map of AI →

The Bottom Line

Mapped onto the essay’s themes by our classification, Situational Awareness LP’s $1.58 billion September 13F is 49.0% GPU compute, 25.4% power, 20.5% memory and 5.1% data-centre sites. Power was 42.7% of its first 13F. The essay calls power the biggest constraint; the filing shows what was held on 30 September, not why.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. We read the introduction and chapter IIIa of Leopold Aschenbrenner’s essay series Situational Awareness: The Decade Ahead (June 2024) in full and quote it verbatim, and parsed every row of Situational Awareness LP’s 13F-HR filings on SEC EDGAR; company descriptions come from Fermi’s and SharonAI’s own 10-Q filings. The theme buckets are our classification, not the fund’s, and the percentages are our arithmetic. We make no claim about why the fund holds any position. Not investment advice; for information only.

Sources: Leopold Aschenbrenner: Situational Awareness, The Decade Ahead (June 2024) · Situational Awareness essay, IIIa. Racing to the Trillion-Dollar Cluster · SEC: Situational Awareness LP 13F-HR, quarter ended 30 September 2026 · SEC: Situational Awareness LP first 13F-HR, quarter ended 31 December 2024 · SEC: Fermi Inc. 10-Q, quarter ended 30 September 2026 · SEC: SharonAI Holdings 10-Q, filed 16 November 2026 · Reuters via The Star: Citadel buys most of Situational’s stock holdings (30 July 2026)

Scroll to Top
FourWeekMBA