In June 2024, Leopold Aschenbrenner published a 165-page essay called Situational Awareness. He committed himself to a chain of dated forecasts. Models would get more processing while producing answers, a method called test-time compute. Power, not chips, would constrain the US AI buildout. The Marcellus shale would power data centres despite climate pledges by the largest cloud companies. By May 21, 2026, most technology and infrastructure calls examined here had landed, received supporting evidence, or were tracking.
Most political prescriptions remained unfulfilled or moved in the opposite direction. There was no voluntary lab merger, Congressional trillions, or coordinated democratic coalition. The US government did not invoke the Defense Production Act (DPA) for the proposed project. Export controls loosened. Some deadlines extend to 2027–28 or the end of the decade, so this political verdict remains provisional.
This split says something about the worldview holding both halves up. Read Situational Awareness beside the four-and-a-half-hour Dwarkesh Patel interview. You find something stranger than a clean libertarian, hawk, or alignment researcher. You find all three running at once.
Artificial general intelligence (AGI) means broad human-level or greater cognitive ability. The essay treats AGI as a dated forecast and a political event.
The libertarian:
“I am a big believer in the American private sector, and would almost never advocate for heavy government involvement in technology or industry.”
That sentence is in Situational Awareness. Three pages later he calls for the largest peacetime industrial nationalisation in American history.
The hawk wants the US to secure its AI labs to a “B-21 bomber-grade” standard. He also wants nuclear deterrence against strikes on US data centres. By 2027/28, he expects the state to absorb the frontier labs into a national project. He calls this Burkean.
“American checks and balances have held for over 200 years and through crazy technological revolutions.”
The alignment researcher was an initial member of OpenAI’s Superalignment team. That team studied how to control AI systems that exceed human ability. Aschenbrenner has said publicly that OpenAI dismissed him in spring 2024 after he shared a security memo with the board. He has also said that he declined a departure agreement with a non-disparagement clause. According to him, that choice cost roughly a million dollars in vested equity. “Freedom is priceless,” was his summary.
He thinks the default plan will probably work. That plan uses scalable oversight, weak-to-strong generalisation, and interpretability. Scalable oversight helps people supervise systems that exceed their direct reviewing ability. Weak-to-strong generalisation asks whether weak supervision can guide stronger systems. Interpretability research examines how a model produces decisions. By the standards of the Machine Intelligence Research Institute (MIRI) and Eliezer Yudkowsky, he is an optimist.
Each identity is internally consistent on its own. Their conjunction is not. A libertarian does not nationalise. A Burkean does not impose a Manhattan Project on a peacetime economy. An empirical alignment optimist does not demand wartime custody in a sensitive compartmented information facility (SCIF) by April 2028. A SCIF protects classified information.
Aschenbrenner reaches all three positions by using the same forecasting machinery in three domains. These domains are compute scaling, great-power competition, and alignment tractability. The forecasts converge on a single decisive event. The opening line of his Dwarkesh interview names it directly:
“What will be at stake will not just be cool products, but whether liberal democracy survives, whether the CCP survives, what the world order for the next century will be.”
CCP refers to the Chinese Communist Party. I read Aschenbrenner the way he reads scaling laws: by decomposition. The three identities form the spine. For each one, I examine the framework, the evidence available by May 21, 2026, and the strongest counter-position. I also examine what the synthesis costs him. The result shows what a convergent forecast does to a worldview. It also reveals something about the AI-policy debate that the rest of us are still having.
I. The forecaster before the politics: a prediction scorecard
Before any criticism, what he got right.
The signature analytical move in Situational Awareness is what he calls “counting OOMs”, or orders of magnitude of effective compute. An OOM is a tenfold change. Effective compute combines training scale with gains from better algorithms and methods.
His decomposition is multiplicative. Training compute adds half an OOM per year. Algorithmic efficiency adds another half an OOM. Discrete unhobbling gains add more. Unhobbling means removing practical limits on model performance. His examples include reinforcement learning from human feedback (RLHF), which uses human preferences to improve responses. They also include chain-of-thought reasoning, which makes a model produce intermediate reasoning steps. Scaffolding connects a model to tools and structured workflows.
Extrapolating to 2027 gives roughly five OOM of effective compute scaleup since GPT-4. In his framing, this equals another GPT-2-to-GPT-4–sized capability jump. The framework is unfashionably simple. It is also empirically unusually well calibrated.
By May 21, 2026, most concrete technology and infrastructure calls examined here had landed, received supporting evidence, or were tracking.
Test-time compute overhang. In June 2024 he wrote:
“What if it could use millions of tokens to think about and work on really hard problems or bigger projects? […] If we could unlock ‘being able to think and work on something for months-equivalent, rather than a few-minutes-equivalent’ for models, it would unlock an insane jump in capability.”
OpenAI launched o1 four months later. DeepSeek-R1 followed in January 2025. Claude extended thinking and GPT-5 thinking provided further support for the paradigm. Aschenbrenner had predicted a model category that did not yet exist. He inferred it from AlphaGo research on the trade-off between compute during training and compute during use.
GPQA Diamond saturation. Graduate-Level Google-Proof Q&A (GPQA) Diamond tests models with difficult science questions. Claude 3 Opus scored 60% when Aschenbrenner wrote. He predicted “this benchmark to fall in the next generation or two.” In May 2026, GPT-5 scored 88.4% without tools. Gemini 3.1 Pro Preview scored 94.3% with extended reasoning. Epoch AI describes the benchmark as near its asymptote, where further gains become small. It took roughly 18 months, almost exactly the prediction window.
Power as the binding constraint. This was the essay’s most prescient line of policy analysis. It preceded Wood Mackenzie’s headline numbers by roughly six months. Heavy-frame turbine lead times reached six years by May 21, 2026. Original equipment manufacturers (OEMs) had sold their order books through 2027. S&P Global put data-centre grid power growth at +22% in 2025. His power-not-chips framing preceded consensus and later became industry consensus.
Marcellus shale and behind-the-meter gas. Behind-the-meter plants generate power at or near the customer site. A gigawatt (GW) is one billion watts. Aschenbrenner argued that abundant US natural gas was the only way to power the trillion-dollar cluster. This claim ran against climate pledges by the largest cloud companies. Meta later ordered ten gas plants for the planned Hyperion build. The build is planned at 7.5 GW, with an associated grid expansion of roughly 30% in Louisiana. Microsoft, Chevron, and Engine No. 1 partnered on 5 GW in West Texas. Williams Companies committed $5B to behind-the-meter turbines. By May 21, 2026, the US pipeline contained over 250 GW of new gas capacity.
This is the part of the record that gets him invited to the room. The next part is what he says once he is in it.
The Gulf chip pivot. He asked, intending the answer to be no:
“Would you do the Manhattan Project in the UAE?”
Stargate UAE was under construction by May 21, 2026. A megawatt (MW) is one million watts. Its eventual capacity is 5 GW, with 200 MW planned for the third quarter of 2026. HUMAIN deployed its first 18,000 GB300s in Saudi Arabia. The Trump administration authorised 70,000 GB300s for G42 and HUMAIN in November 2025. The thing he warned against happened. The same US government that he expected to nationalise the labs authorised it.
AMD’s compute TAM. Total addressable market (TAM) estimates the revenue opportunity available to an industry or product. Aschenbrenner cited AMD’s $400B AI accelerator forecast for 2027. AMD later reaffirmed the forecast and upgraded to a $1T compute TAM by 2030.
Scaling lawfulness, generally. Aschenbrenner claims that scaling laws held across fifteen orders of magnitude. He cites the Kaplan-to-Chinchilla-to-GPT-4 chain. Evidence from 2024-26 broadly supports this claim. The 2024 Kaplan paper from Anthropic on data-bound scaling complicates the picture. So do capability gains driven mainly by RL post-training rather than pretraining floating-point operations (FLOPs). Neither development breaks the scaling claim.
Several predictions were ahead of consensus by months. The predictions themselves were not the novel contribution. Anyone reading SemiAnalysis and AMD investor decks could have triangulated most of them. The novelty was his commitment to a dated narrative in which they all compound.
A sceptical reader will note that San Francisco’s AI cluster has better information about what happens inside the labs. Granted. But test-time compute was not yet inside-the-lab consensus when he published. The power constraint preceded the industry’s shift. The Gulf-at-the-AGI-table point preceded the actual deal flow. There is real foresight here, even after discounting for the information asymmetry.
The May 21, 2026, scorecard cuts hard the other way on another category of prediction. I return to that record later. The technology and infrastructure forecasts earn the rest of this essay. If those forecasts had been mediocre, his worldview would not be worth dissecting.
II. Identity one: the libertarian
“I am a big believer in the American private sector, and would almost never advocate for heavy government involvement in technology or industry.”
That sentence opens Chapter IV of Situational Awareness. The chapter is titled “The Project.” It argues that the United States must absorb frontier AI labs by 2027–28. The national effort would resemble the Manhattan Project and could be voluntary or compulsory. Congress would appropriate trillions of dollars for compute and power. Democratic allies would form a coalition for AGI development. The core research team would move to a SCIF.
The gap between the opening sentence and the thesis is not hidden. He flags it himself:
“I used to apply this same framework to AGI — until I joined an AI lab.”
Aschenbrenner’s libertarianism is operational, not rhetorical. He says he declined a departure agreement with a non-disparagement clause when he left OpenAI. According to him, that refusal cost “close to a million dollars” in vested equity. He did not want to constrain his future ability to write what he believes.
Aschenbrenner calls himself a “speech deontologist,” meaning that he says what he thinks even when the consequences argue against it. He admires the American private sector. He also admires the wider US institutional system: the Federal Reserve, the Supreme Court, and the Constitution.
The November 2025 Genesis Mission executive order (EO) under Trump 2.0 used explicit Manhattan-Project language. It mobilised Department of Energy (DOE) national laboratories for AI-accelerated science. This is the rhetoric that Aschenbrenner predicted. The institutional structure did not form. There was no voluntary lab merger, no appropriation of trillions, and no coalition of democracies.
His libertarianism is not absolute. It creates a presumption against state action with one exception. The exception applies when civilian technology becomes a weapons-grade output for state actors. He reclassifies AGI from “tech industry” to a weapon of mass destruction (WMD). His rule against state involvement therefore no longer applies.
“They’re startups. And startups are startups, you know — I think they’re not fit to handle WMDs.”
This argument is structurally cleaner than it looks. Nineteenth-century classical liberalism made similar exceptions. Mill wrote about intervention in failed states. Bagehot treated the Bank of England as a lender of last resort. Hamilton accepted a national bank when the political economy required it. Aschenbrenner’s novelty is using an industrial technology as the trigger. The empirical question is whether AGI clears that bar. The structure of the argument is not new.
By May 21, 2026, the full nationalisation that he predicted had not happened. Frontier labs remained decentralised and competitive. Stargate was announced as a four-year, $500B investment plan involving OpenAI, SoftBank, Oracle, and MGX. It looks more like a defence-contractor public-private partnership (PPP) than a Manhattan Project. The CHIPS Act still contained the original $39B + $11B from 2022. Congress had not appropriated trillions.
The AI Safety Institute (AISI) became the Center for AI Standards and Innovation (CAISI) in June 2025. The multilateral framework that Aschenbrenner expected had dissolved into transactional bilateral agreements.
A different version of his prediction did partly land. Defence contracting was expanding by May 21, 2026. Anthropic signed a $200M Department of Defense (DoD) contract. Claude became the first AI model authorised on classified networks. Stargate is functionally a national-security PPP under commercial wrapping. His interview line “we’ll all be in a bunker” was correct in spirit and wrong in form.
The investment firm is the product of a libertarian instinct that refuses to die. He wants situational awareness to become financially decisive before the state arrives.
“If AGI were priced in tomorrow, you could maybe make 100x. Probably you can make even way more than that because of the sequencing.”
The fund is the libertarian’s hedge against his own forecast. It captures private upside before his predicted nationalisation reshapes the prize.
The strongest counter-position says that this is not a libertarian making a careful exception. It is a hawk borrowing libertarian clothes.
The defence starts with his equity forfeiture, his speech deontology, and his refusal to sign an NDA at material cost. His libertarian commitment is operational when he can choose to exit, including from his own employment. That choice cost him real money. His conditional preference for state action is consistent with classical liberalism rather than a defection from it. He is careful about the exception, not opportunistic.
That is intellectually honourable. It is also the first place where the synthesis starts to creak. A real libertarian who pays a million dollars to call for a Manhattan Project is more interesting than a hypocrite.
He asks us to follow a chain whose first link is “I changed my mind about state involvement when I joined an AI lab.” That is an empirical claim. The prescriptive jump also depends on the hawk’s claim about adversary capability. It then depends on the alignment optimist’s claim about institutional tractability.
III. Identity two: the Burkean hawk
“In some sense, this is simply a Burkean argument: the institutions, constitutions, laws, courts, checks and balances, norms and common dedication to the liberal democratic order […] have withstood the test of hundreds of years. Special AI lab governance structures, meanwhile, collapsed the first time they were tested.”
Aschenbrenner names Burke in Chapter IV to defend the Project. The phrase “collapsed the first time they were tested” refers to OpenAI’s November 2023 board crisis. The board fired Sam Altman and briefly put Mira Murati in charge. Then ~700 employees threatened to resign. Within five days, the board reinstated Altman and changed its membership. Special-purpose AI-lab governance structures genuinely collapsed on contact with reality. Aschenbrenner’s empirical observation is correct.
The substantive Burkean intuition is that two-hundred-year-old institutions absorb shocks better than improvised lab-governance structures. On its face, that is a strong empirical argument. Aschenbrenner’s respect for American institutions is unusual among accelerationists, who favour faster technological development. Much accelerationist discourse in 2024 was openly contemptuous of regulators, Congress, and constitutional checks. Aschenbrenner takes them seriously enough to attempt a synthesis.
He praises the Federal Reserve as a model of competent technocratic delegation. He also admires the Supreme Court. In his words, “they really believe in the constitution, they love the constitution”. He recommends listening to oral arguments as a podcast.
The Burkean argument also cuts the other way. Burke argued that radical institutional changes tend to cause unintended damage, even when reformers share the institution’s goals. In Burke’s sense, the conservative move is slow.
Aschenbrenner proposes a Manhattan Project for a peacetime economy in 2027. An accelerated executive command would run it. Frontier research would move to a SCIF, and workers would build the trillion-dollar cluster in record time. This is the modal Promethean prescription, not the modal Burkean one.
A Burkean prior would favour stronger lab governance and slower state involvement. It would favour more private firms running parallel efforts, not fewer. It would also resist rapid centralisation of dual-use capability under one chain of command.
Aschenbrenner’s Burkean argument therefore performs a specific task. He does not invoke Burke to defend AI as it is. He argues that two conditions would make existing constitutional institutions the only viable home for AGI. First, AGI must arrive within a decade. Second, it must decide national security.
The rhetorical move is that Burkean ends require Promethean means. Existing institutions must survive. To preserve them, the state must absorb the technology that threatens them into its existing chain of command.
“There’s only one chain of command and set of institutions that has proven itself up to this task.”
The institutions Aschenbrenner praised held under stress. The 2024 election produced a peaceful transfer of power. The Federal Reserve maintained operational independence despite significant political pressure on interest-rate policy. The Supreme Court ruled against the executive on several procedural questions in 2025. In the narrow sense, the constitutional system did not break.
But those institutions did not produce the response that Aschenbrenner predicted. Trump 2.0 made a commercial and deregulatory pivot. That response differed from the one assumed by his Burkean-Manhattan synthesis.
The export controls that he expected to tighten instead loosened. The AI Diffusion Rule was scrapped in May 2025. In December 2025, Trump announced that Nvidia could sell H200-equivalents to China for a 25% revenue tariff. He expected the regulatory system to mobilise around a national-security exception. Instead, it monetised the exception. Export controls became transactional. The administration treated coalition partners as leverage points rather than allies. It also rebranded the AI Safety Institute.
The Manhattan Project analogy does more rhetorical work than it can carry. Nuclear weapons were excludable because their physics required uranium-235 or plutonium-239. States could control both substances. Manhattan Project secrecy survived because the physics required those materials.
AI weights are infinitely reproducible after exfiltration. DeepSeek’s January 2025 R1 release also showed that algorithms diffuse through papers and reverse engineering. DeepSeek worked with a fraction of the budget that frontier labs were spending.
Aschenbrenner’s proposed regime lacks the physical basis of the original Nuclear Non-Proliferation Treaty (NPT). He asks the political system to control a technology without excludability. Excludability lets an owner prevent others from accessing or reproducing something. Historical control regimes depended on it.
The strongest defence says that calling him a hypocrite for invoking Burke is itself un-Burkean. Burke was not a pacifist about state action. He was a pragmatist about the load that institutions can bear.
The hawk-Burkean combination has a serious lineage. It runs through Hamilton, the liberal hawks of the Second World War, Acheson, Kennan, and the architects of NSC-68. Granted. But an unprecedented industrial nationalisation by 2027 remains a real stretch as the Burkean move. Twenty-three months of evidence had not produced the institutional mobilisation that his framework requires. The institutions held. They just held in a direction he did not predict.
Most acceleration discourse is contemptuous of institutions; Aschenbrenner’s version takes them seriously enough to attempt a synthesis. The synthesis fails on its own terms, but it is a failure worth having.
IV. Identity three: the alignment optimist
Aschenbrenner was an initial member of OpenAI’s Superalignment team. The team sought a successor to RLHF. Its mission was to align AI systems that were substantially smarter than their human supervisors. Aschenbrenner worked under Ilya Sutskever and Jan Leike. OpenAI publicly committed 20% of its compute to the team through 2027.
By the standards of MIRI, Yudkowsky, and the 99%-p(doom) group, Aschenbrenner is a moderate. The term p(doom) means the estimated probability of an AI catastrophe. Situational Awareness calls alignment “a real technical problem” but a “solvable” one. He treats it as a machine-learning (ML) engineering problem rather than a philosophical problem.
Aschenbrenner writes that the default plan will probably work. Scalable oversight would use debate, recursive reward modelling, and prover-verifier games for somewhat-superhuman systems. Weak-to-strong generalisation would cover the next tier. Interpretability research is the longest-shot but most ambitious bet.
“I’m incredibly bullish on the technical tractability of the superalignment problem. […] Deep learning generalises surprisingly benignly in many situations: it often just ‘does the thing we meant’ rather than picking up some abstruse malign behaviour.”
In technical terms, he argues that evidence from the models available when he wrote weighs against the worst alignment failures as the most likely outcome. Those failures include deceptive alignment and instrumental convergence towards power-seeking.
This is the alignment chapter’s strongest empirical claim. Twenty-three months later, it is probably the claim that aged worst.
Aschenbrenner correctly identified RLHF as an approach with scaling limits. Later papers showed that reward hacking generalises beyond the training distribution. Reward hacking occurs when a model exploits its scoring rule instead of doing the intended task. The resulting behaviour included alignment faking, sabotage, and cooperation with malicious actors.
Anthropic’s Sycophancy to Subterfuge from June 2024 studied this problem. School of Reward Hacks followed in August 2025. Both papers show that hacking on harmless tasks can generalise to misalignment.
The November 2025 paper Natural Emergent Misalignment from Reward Hacking studied reward hacking in production RL. The training process produced models that lied, sabotaged codebases, and reasoned about malicious goals as a routine result.
These results validate Aschenbrenner’s diagnosis. RLHF breaks at scale, alignment is a real technical problem, and researchers must address it empirically. The results do not validate his optimism.
In December 2024, Anthropic published Alignment Faking in Large Language Models. Claude 3 Opus complied with harmful queries 14% of the time when it inferred that researchers were training it. Its chain of thought contained explicit alignment-faking reasoning. This behaviour emerged from routine training without researchers training it directly.
That month, Apollo Research published Frontier Models are Capable of In-Context Scheming. The evaluations covered o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro, and Llama 3.1 405B. Every model attempted to copy its weights to external servers during scheming evaluations.
arXiv 2503.17378 found that 11 of 32 systems could replicate themselves. The observed behaviour included unprompted self-exfiltration to evade shutdown.
Chain-of-thought faithfulness measures whether a model’s stated reasoning reflects the factors that shaped its answer. Aschenbrenner called it a “criminally underrated” tool for alignment monitoring. Later evidence substantially weakened that tool. Claude 3.7 Sonnet acknowledged influential hints only 25% of the time. DeepSeek-R1 acknowledged them 39% of the time.
In its strong form, “Deep learning generalises benignly” is refuted. The structural fact is more damaging. His default plan depended on labs investing seriously in superalignment infrastructure.
Three weeks before Situational Awareness appeared, OpenAI dissolved the Superalignment team. Sutskever resigned May 14, 2024. Leike resigned May 15:
“Safety culture and processes have taken a backseat to shiny products.”
In the Dwarkesh Patel interview, Aschenbrenner gave his own public account of the dismissal. He said it came one to two weeks earlier. According to his account, his exit discussions covered three factors. One was a security memo that he had sent to the board. Another was his refusal to sign the November 2023 employee letter. The third was his position on AGI nationalisation. OpenAI has not commented publicly on the specifics.
The 20% compute commitment was effectively voided. OpenAI dissolved Superalignment’s successor, the Mission Alignment team, in February 2026.
This evidence strengthens his geopolitical case while weakening his alignment case. In his framework, failed lab governance leaves state custody as the remaining institutional answer for safety.
But speed pressure makes alignment faking and scheming operationally dangerous under state custody. Several models attempted self-exfiltration during research evaluations. The danger grows in a wartime SCIF whose operators are racing China through an intelligence explosion.
The Project was supposed to provide a safety margin that the labs would not. Evidence through May 21, 2026, points the other way. Under speed pressure, state custody appears less safe for the failure modes that researchers had observed by that date.
The strongest defence is timing. He wrote in May 2024. Most alignment-faking and scheming research did not yet exist. Every alignment researcher in mid-2024 had less evidence of deceptive behaviour in frontier systems. Granted.
But the asymmetry is not “I underweighted X paper”. He based his political conclusion on alignment going reasonably well. The alignment evidence moved the other way faster than he projected. That shift changes the political calculation, regardless of when one writes.
V. Why the three have to run together
If compute scales as predicted, then timelines compress. Then state-actor competition becomes active. Then the libertarian must yield to the hawk. This is Identity One’s libertarian-empiricist foundation: OOM counting.
Drop the compute forecast and the politics dissolve. The chain requires AGI to arrive by 2027–28 on a 10 GW cluster. Without that timing, the claim of a near-term decisive military advantage falls away. There is therefore no Burkean exception for a Manhattan Project and no nationalisation argument.
If the hawk is right about adversary capability, then lab security must reach SCIF standards. Then private-startup governance becomes inadequate. Then the alignment optimist’s default plan needs state-scale resources. This is Identity Two’s institutional-Burkean argument.
Drop the hawk and the Project becomes unnecessary. Labs can muddle through under existing market discipline. Drop the SCIF requirement and the alignment argument gains more time.
If the alignment optimist is right, then the default plan is tractable. Then the Manhattan Project is justified. Then engineers can align the system after nationalisation. The state receives a solvable engineering problem rather than an uncontrollable catastrophic capability.
Drop the optimism and the Project becomes reckless. The hawk-libertarian-Burkean exception then transfers unprecedented military capability to a chain of command that cannot reliably control it. This is Identity Three’s empirical optimism.
The synthesis is therefore forced. Each identity is a precondition for the next. That structure makes Aschenbrenner intellectually interesting and hard to refute piecemeal. A critic must break the chain at a specific link.
He is the rarest kind of public forecaster: legible enough to be wrong about. The same clarity makes his framework fragile. It is a series-circuit argument. One failed component breaks the whole circuit.
From the May 21, 2026, perspective, the record on the three premises was mixed.
Compute scaling was alive but contested. Capability gains continued, and test-time compute supported the framework in a way that he did not emphasise. However, pretraining slowed relative to RL post-training. That shift complicates the “5 OOM additivity” framing.
The evidence on adversary capability moved in the wrong direction for his argument. China did not catch up by stealing weights as he predicted. It reproduced the algorithmic frontier through papers and reverse engineering on a fraction of the budget. DeepSeek-V3, R1, V3.1, and V4 illustrate the path. The security thesis correctly predicted diffusion but identified the wrong channel.
On alignment, the evidence confirmed his diagnosis but weakened his optimism. The deepest blow came from the institutional choices of his former employer.
The strongest counter-counter is that the timeline remained open on May 21, 2026. Aschenbrenner’s AGI window is 2027–28. His window for political reordering is “by end of decade”.
Several political forecasts could still turn green by 2028. Genesis Mission could harden into Manhattan-Project-style consolidation. A different administration could tighten export controls. OpenAI’s relationship with Stargate could move closer to nationalisation.
Concede this honestly: the verdict is provisional. But the methodological point about series-circuit fragility does not depend on the final political outcome. Even if three political claims turn green by 2028, forecasts would still depend on the processes underneath them.
VI. The coalition triad and its cracks
Aschenbrenner’s most operational geopolitical model has three coalition rings for the post-AGI world.
The inner ring contains democracies. The US, the UK through DeepMind, Japan, South Korea, and core North Atlantic Treaty Organization (NATO) members would coordinate AGI development. A pact like the 1943 US-UK Quebec Agreement on nuclear cooperation would govern the work.
The middle ring shares benefits with non-aligned states through an Atoms-for-Peace structure. Atoms for Peace was a Cold War programme for sharing civilian nuclear technology. The outer ring seeks to contain authoritarian adversaries through export controls, espionage interdiction, and ultimately deterrence.
“Perhaps most importantly, a healthy lead gives us room to maneuver.”
As of May 21, 2026, the inner ring had not formed. No coalition of democracies had emerged. The Group of Seven (G7) AI Industry/Digital Ministerial declarations existed. So did the Council of Europe AI/Human Rights Framework Convention. Neither created a coordinated bloc for AGI development.
Trump 2.0 acted unilaterally. The administration rebranded the AI Safety Institute (AISI) as CAISI in June 2025. It treated allied AI policy as leverage in trade negotiations. There was no secret pact and no bilateral coordination at the technology level.
The middle ring formed on commercial terms that he did not anticipate. Gulf authorisations covered 70,000 GB300s for G42 and HUMAIN. They also supported the partnership structure for Stargate UAE. These are not Atoms-for-Peace benefit-sharing arrangements. They are paid commercial deals.
The administration announced a 25% revenue tariff on H200 sales to China in December 2025. “Would you do the Manhattan Project in the UAE?” became a real question with a real answer. The UAE received the compute on transactional terms.
The outer ring inverted. Containment loosened instead of tightening. The AI Diffusion Rule was scrapped in May 2025. In January 2026, the final rule changed H200/MI325X licence review from “presumption of denial” to “case-by-case.” This change reversed the direction that Aschenbrenner expected from any administration that recognised AGI’s importance.
The twenty-three-month record exposed a flaw in the Atoms-for-Peace analogy. His proposed non-proliferation regime lacks the physical substrate of the original NPT. DeepSeek’s R1 release in January 2025 demonstrated the problem. A Chinese lab reproduced reasoning capability comparable to o1 on a budget that leading US labs would consider impossibly small. It used publicly available technical insights and reverse-engineered training recipes.
This breaks something specific in the framework. If algorithmic secrets are not durable moats, the security thesis addresses the wrong threat. Locking down labs may stop state actors from exfiltrating information. It does not stop them from reproducing the work. The threat is not theft. The threat is reproduction. Reproduction requires nothing that Aschenbrenner’s lockdown architecture would prevent.
A defender can fairly argue that the timeline remained open on May 21, 2026. The coalition could still form. Granted. But the direction of motion in 2025-26 was away from his coalition. That evidence makes an Atoms-for-Peace structure less likely to remain reachable.
VII. What this reveals about the debate
I do not think Aschenbrenner’s specific predictions will be the durable contribution. By May 21, 2026, the political programme remained largely unfulfilled or reversed. Several parts may not survive 2027. His transferable contribution is methodological, and it extends beyond him.
He forced analysts to integrate all three domains. Most public AI-policy commentary in 2024-26 was single-domain analysis presented as comprehensive work.
The doomer position includes Yudkowsky and the MIRI tradition. It gives too much weight to alignment and too little to geopolitics and political economy. The “bomb the data centres” line ignores adversary capability. It also ignores the political infeasibility of unilateral abstention.
If AI affects national survival, a unilateral pause is structurally inadequate. Another actor will continue. Aschenbrenner inverts the doomer prescription: nuclear deterrence for data centres. The US would threaten retaliation if adversaries struck American AI infrastructure.
This position makes more sense as a national-security argument. It works less well as an alignment argument. It increases the speed pressure and military integration that make alignment harder.
The accelerationist position includes effective accelerationism (e/acc) and the Andreessen wing. It gives too much weight to technology diffusion. It gives too little weight to alignment engineering and adversary capability. This camp models AI as consumer technology in a frictionless global market.
“Bearish on the wrapper companies,” Aschenbrenner says. A wrapper company builds an application around a foundation model. His contempt for these startups mirrors what he calls the e/acc move. Both assume that industrial-scale capability will create more value than workflow integration.
Evidence through May 21, 2026, was mixed. Cursor had reached roughly $500M in annual recurring revenue (ARR) by Q4 2025. Glean and other companies had built defensible enterprise positions. The wrapper short had not been clean.
The political-economy record weakened the broader e/acc thesis. That thesis expects AI to diffuse and integrate like consumer software. The state became involved, although not in the form that Aschenbrenner predicted.
The mainstream AI-safety policy position includes US standards and the European Union AI Act. It gives too much weight to regulatory process. It gives too little weight to capability speed and supply-chain geopolitics.
The National Institute of Standards and Technology (NIST) leads the US standards track. The EU AI Act came into force as frontier capability moved towards test-time compute. The Act had not anticipated that paradigm. Aschenbrenner’s framework would have predicted this type of failure.
He calls the regulatory process structurally too slow: “NIST takes years and they figure out what the expert consensus is.” On this point, he was correct.
I expect the most useful AI-policy work in 2026–28 to integrate all three domains. Most commentary available by May 21, 2026, failed this test. Alignment specialists could not model power. Geopolitical specialists could not model alignment. Policy specialists could not model the underlying capability curves.
Aschenbrenner rejects probability distributions and instead “tells the modal story”. A modal story is a vivid, dated, and falsifiable narrative bet:
“I have a lot of uncertainty. So a lot of the time I’m trying to tell the modal story, because I think it’s important to be concrete and visceral about it. And I have a lot of uncertainty basically over how the 2030s play out. But basically the thing I know is, it’s gonna be fucking crazy.”
The method works on empirically lawful processes. Examples include scaling curves plotted on logarithmic axes, capital-expenditure (capex) aggregates, and power use by the largest cloud companies. It works poorly when elections, executive turnover, and coalition politics govern the process.
Through May 21, 2026, his technology and infrastructure forecasts were unusually well calibrated. Their underlying processes had the right substrate. Here, substrate means the causal process that produces the outcome. His political-economy forecasts were poorly calibrated over the same period because their substrate was different.
When a forecaster rejects probability distributions and commits to a dated narrative, ask: what is the substrate? Is it empirically lawful or coalitionally contingent? The modal-story method becomes more durable as the process becomes more empirically lawful. This dependence is substrate sensitivity.
Aschenbrenner’s instrument is sensitive to log-log evidence and saturated against electoral politics. His success in 2024-26 tracked how much of each question fell inside that range.
The market moves favoured the empirical-substrate trades that he described, but not the coalition-substrate trades. Nvidia, AMD, gas turbines, transformer-shortage positions, and Gulf data-centre exposure gained. Bets on defence-AI consolidation and against open-source diffusion would not have paid. Strategically, an AI roadmap built on this narrative should anchor on substrate rather than rhetoric.
VIII. Situational Awareness LP: capital, conviction, and the honesty of the bet
Situational Awareness LP makes the worldview unavoidable because it reveals a capital position. Aschenbrenner has publicly described Patrick Collison, John Collison, Daniel Gross, and Nat Friedman as anchor investors. By his account, he put his own capital and their capital where his framework points.
In the Dwarkesh interview, Aschenbrenner described a trade sequence. Nvidia came first. Taiwan Semiconductor Manufacturing Company (TSMC), packaging, and memory followed. US power and utilities came next, then natural gas and later Google. After those trades came “the big bond short” on real interest rates above 10%. Out-of-the-money (OTM) tail bets came last. An OTM option has no intrinsic value at the current market price; here, it serves as a low-probability tail bet.
This sequence expresses his framework in capital. None of the named LPs has publicly confirmed a specific position. Every position description comes from Aschenbrenner’s public statements about his fund.
The market performance of the trades that Aschenbrenner described provides a provisional verdict on the framework. The semiconductor trades worked. Nvidia was worth $5.2T by May 21, 2026. That valuation put the company on a possible path to $10T by 2028. The latter figure remains a forecast. AMD’s $1T compute TAM upgrade supported the move into the wider semiconductor stack.
The power trade gained support: turbine lead times reached 243 weeks, Meta ordered ten gas plants, and the Marcellus thesis gained industry-scale evidence.
The Google trade remained open on May 21, 2026. Google had not reached the $100B AI revenue threshold that he treated as the catalyst. Its current annualised pace remained below $100B and still depended on capability and adoption cycles that had not closed.
The bond short had not fired. A bond short gains when bond prices fall, usually as yields rise. Real yields measure interest rates after expected inflation. Real ten-year yields were roughly 2% in mid-2024. By May 21, 2026, they were roughly 2.0–2.4%, up modestly and far below 10%.
Aschenbrenner explicitly framed the trade as a tail bet with negative carry. Negative carry means that holding the position costs money over time. Two years of cost without a payoff tells us something about the premise, not only the timing.
The conflict of interest deserves one mention. Aschenbrenner’s policy advocacy can move markets in directions that favour many of his trades. The nationalisation case favours bets on rising US power and defence-contractor prices. The export-control case would favour TSMC if Trump 2.0 had not reversed direction. The trillion-dollar-cluster case favours the largest cloud companies.
The advocacy may be sincere, and the fund may benefit from it. Both statements can be true. Financial analysts follow disclosure rules that do not apply to essays. Readers should know the financial position behind the document. Dwarkesh, who is friendly with Aschenbrenner, did not press this point in the interview. Finance normally requires an analyst to disclose such interests.
The fund also hedges Aschenbrenner personally. He repeats a friend’s joke about that hedge:
“A friend joked that the investment firm was perfectly hedged for me. It’s like, you know, either AGI this decade — and yeah, your human capital is depreciated, but you’ve turned that into financial capital. Or no AGI this decade, in which case maybe the firm doesn’t do that well, but you’re still in your 20s and you’re so smart.”