Leopold Aschenbrenner is a former researcher on OpenAI's Superalignment team and the author of Situational Awareness: The Decade Ahead, a six-part essay series published in June 2024 that argues artificial general intelligence is achievable by around 2027 and frames AI development as a strategic contest with China. After leaving OpenAI in early 2024, he founded an AI-focused investment firm built around the theses in the essay.
Role: Founder, AI-focused investment firm (post-2024) Previous: Researcher, OpenAI Superalignment team (until early 2024) Known for: Situational Awareness: The Decade Ahead (June 2024)
Background and departure from OpenAI
Aschenbrenner worked on OpenAI's Superalignment team until early 2024, leaving before the team was dissolved later in 2024. He has publicly discussed concerns about OpenAI's security posture and its commitment to its superalignment program, citing those concerns as part of the motivation for Situational Awareness. Jan Leike was a colleague on the Superalignment team; OpenAI was his former employer.
Situational Awareness and its arguments
Situational Awareness is a six-part essay series dedicated to Ilya Sutskever. Its central claims are that AGI is achievable by around 2027 via compounding orders-of-magnitude gains in compute, algorithmic efficiency, and "unhobbling" (tool use, chain-of-thought, agency); that an intelligence explosion driven by automated AI research follows shortly after; that trillion-dollar compute clusters requiring roughly 100 GW of power will be built by the late 2020s; that current lab security is, in his characterization, catastrophically inadequate and trivially penetrable by state-level adversaries, especially the Chinese Communist Party; that the US government will nationalize or tightly integrate with frontier AI labs (which he terms "The Project") by 2027–28 as a Manhattan-Project analog; and that the "free world must prevail," on the argument that whichever coalition first controls superintelligence holds enduring military dominance.
The essay positioned AGI timelines and AI development as a hard-power geopolitical contest with China. Aschenbrenner is credited with shifting the Overton window on AGI timelines in Washington, venture-capital, and national-security circles. The China-contest framing he advanced is subsequently visible in America's AI Action Plan and OpenAI's Industrial Policy. The essay is widely cited both approvingly, by advocates of an AGI race, and critically, by skeptics who view the timeline and nationalization theses as speculative.
Subsequent work
Aschenbrenner founded an AI-focused investment firm premised on the theses in Situational Awareness.
The firm, Situational Awareness LP, sold its entire book of public equity positions on July 30, 2026 after losses on AI-infrastructure holdings and a short position in software left it seeking cash to meet margin requirements from prime brokers, with Citadel reaching a deal to buy the assets (Source: cnbc.com). The fund stood at $45 billion at the start of July 2026. Its four largest first-quarter holdings — Nebius Group, Sandisk, Micron and CoreWeave — were each down more than 35% over the month, while shorts including Adobe moved against it; Bank of America, Goldman Sachs and JPMorgan Chase worked with the fund on the unwind. A spokesman said reports that the firm was marketing its Anthropic stake were not accurate. See AI Bubble Debate.
Writing on August 1, 2026, Azeem Azhar of Exponential View said the fund peaked at $45 billion and ran the AGI capital-expenditure trade at roughly four times leverage, and that Aschenbrenner founded it in late 2024 on the thesis that the path to superintelligence would move trillions of dollars into compute, chips and power. Azhar attributed the reversal to the Philadelphia Semiconductor Index falling 28.6% from its June peak while software gained, and argued that the unwind is not itself evidence the underlying thesis was wrong (Source: exponentialview.co).
Reuters reported the scale and sequence on July 31, 2026: the portfolio lost 67% of its value over the course of July 2026, and the firm was forced to unwind most of a $16 billion public-equities book, with Citadel buying a portion of it by July 30. Ken Griffin spoke directly with Aschenbrenner on the morning of July 29, 2026, after which Citadel executives analysed the fund's positions overnight. Aschenbrenner wrote to investors: "We let you down." Reuters could not establish how much Citadel made on the deal (Source: reuters.com).
The unwind's cost reached at least one of the fund's investors. Jane Street told employees in a note dated August 14, 2026 that it took a $15 billion hit in July 2026 from its investment in Situational Awareness and from other technology positions, its first negative month of trading revenue since 2016, leaving revenue down roughly 25% from its end-of-June peak. The firm had generated more than $40 billion in trading revenue year to date, one source said, above the $39.6 billion it made in all of 2025. The note attributed the failure of its hedges to the shape of the drawdown: "We generally worry most about sharp drawdowns, and buy puts that would help in those scenarios. The losses in AI stocks were relatively spread out throughout the month, so those short-term hedges provided little help." Several of the largest memory and semiconductor stocks fell around 50% during July (Source: reuters.com).
Following the sale of the public book, the firm's activity moved to private positions. In an account published August 7, 2026, Situational Awareness invested a further $400 million in Source Foundry, a chip-manufacturing startup, bringing its total investment in the company to $500 million; it was the latest known private-company position taken by the fund after the stock portfolio was sold. Source Foundry was founded in 2025 by Stanford University researchers Abdulmalik Obaid and Joe Burg, has raised at least several hundred million dollars to date including a recent round at a $5 billion valuation, and was previously backed by Sequoia Capital (Source: wsj.com).
Retrospective assessment
Two independent one-year-later retrospectives grade Aschenbrenner's forecast: Delisle (June 2025, quantitative) and Harris (March 2026, eight-category). The fuller evaluation is collected in Situational Awareness Retrospectives.
The retrospectives find several elements on-pace or ahead of his forecast: compute, algorithmic-efficiency, capex, chip, and power trajectories; the inadequacy of lab security, which they treat as vindicated by subsequent state-actor exploitation evidence; the field's rebranding from safety toward security, which tracks his framing; and the US-China strategic-contest framing, which they treat as confirmed by espionage, an estimated $2.5B in chip smuggling, a 7nm Chinese fab, and the decision by both the US and China to opt out of the February 2026 responsible-AI-military declaration.
The retrospectives also identify elements they assess as wrong or partially wrong. On AI revenue, Aschenbrenner predicted a $100B run-rate by mid-2026; the actual figure was around $60B, which the retrospectives characterize as off by roughly 1.7× but directionally validated. On "The Project," no US-government nationalization or Manhattan-style integration has materialized; the retrospectives note that the direction of travel through export controls and AISI buildout is consistent, but the nationalization step has not occurred. On the structure of Chinese competition, they find that Aschenbrenner missed that Chinese labs (DeepSeek, Qwen, Kimi) would produce genuine architectural innovation rather than merely distill US models. On open-source at the frontier, they read the essay as implicitly assuming a collapse that did not occur, with open-weight models instead continuing to thrive. On the qualitative "shocking leap," they note that the visceral wake-up moment he anticipated has not landed, with the trajectory proceeding benchmark-by-benchmark rather than as a discontinuous jump. The Harris grading (EA Forum, March 2026) summarizes the eight categories as infrastructure ahead, capabilities on-target, revenue 1.7× low, Chinese-innovation structure missed, and security and safety vindicated (Source: Raw Sources/How Did Leopold Do.md).
Across both retrospectives, the assessment is that Aschenbrenner's physical and infrastructure forecasts have tracked more accurately than his political forecasts, and that his record on direction has held up better than his record on mechanism.
Relationships
- related: Situational Awareness — The Decade Ahead — his central work
- related: Situational Awareness Retrospectives — one-year-later evaluations
- related: (Source: Raw Sources/How Did Leopold Do.md) — Harris (EA Forum, March 2026) eight-category grading: infrastructure ahead, capabilities on-target, revenue 1.7× low, Chinese innovation structure missed, security/safety vindicated
- related: OpenAI — former employer
- related: Jan Leike — former Superalignment colleague
- supports: AGI Timelines (aggressive position), AI Race Dynamics, Recursive Self-Improvement
- contradicts: AI as Normal Technology