In June 2024, Leopold Aschenbrenner published a 165-page essay called Situational Awareness and committed himself to a chain of dated forecasts. Test-time compute would be the next paradigm. Power, not chips, would be the binding constraint on the US AI buildout. The Marcellus shale would end up powering data centres, not the hyperscalers’ climate pledges. By May 2026 those calls have landed. Eight of his concrete technology and infrastructure predictions have confirmed cleanly.
Eight others have falsified: voluntary lab merger, Congressional trillions, a coalition of democracies, tightening export controls.
This split tells us something about the worldview holding both halves up. Read Situational Awareness and the four-and-a-half-hour Dwarkesh Patel interview side by side and you find something stranger than a clean libertarian, a clean hawk, or a clean alignment researcher. You find all three running at once.
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: he wants the US to lock down its AI labs to “B-21 bomber-grade” security, threaten nuclear retaliation against strikes on its data centres, and absorb the frontier labs into a state project by 2027/28. He calls this Burkean.
“American checks and balances have held for over 200 years and through crazy technological revolutions.”
The alignment researcher: he was an initial member of OpenAI’s Superalignment team. He has publicly stated, including in the Dwarkesh Patel interview, that he was dismissed in spring 2024 after sharing a security memo with the board. He has also publicly stated that he declined the company’s non-disparagement NDA and forfeited roughly a million dollars in equity as a result. “Freedom is priceless,” was his summary. He thinks the default plan (scalable oversight, weak-to-strong generalisation, interpretability) will probably work. By the standards of MIRI and Yudkowsky, he is an optimist.
Each of those identities, taken alone, is internally consistent. The 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 insist that the only viable governance structure is wartime SCIF custody by April 2028.
Aschenbrenner gets to all three because he runs the same forecasting machinery on three different domains (compute scaling, great-power competition, alignment tractability) and the forecasts converge on a single decisive event. The opening line of his Dwarkesh interview names the convergence 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.”
This essay reads Aschenbrenner the way he reads scaling laws: by decomposition. The three identities are the spine. For each, we look at the framework, today’s evidence, the strongest counter-position, and what the synthesis costs him. The point is to show what a convergent forecast does to a worldview, and what that reveals about the AI-policy debate the rest of us are still having.
I. The forecaster before the politics
Before any criticism, what he got right.
The signature analytical move in Situational Awareness is what he calls “counting OOMs”, orders of magnitude of effective compute. The decomposition is multiplicative: half an OOM per year of training compute, half an OOM per year of algorithmic efficiency, plus discrete unhobbling gains (RLHF, chain-of-thought, scaffolding, tools, long context, post-training). Extrapolated forward to 2027, this gives roughly five OOM of effective compute scaleup since GPT-4, equivalent in his framing to another GPT-2-to-GPT-4–sized capability jump. The framework is unfashionably simple. It is also empirically unusually well-calibrated.
By May 2026, eight of his most concrete dated technology and infrastructure predictions have either confirmed cleanly or are tracking on schedule.
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 validated the paradigm completely. He predicted a category of model that did not yet exist by inference from the test-time-compute-vs-train-time-compute trade-off in AlphaGo’s training literature.
GPQA Diamond saturation. When he wrote, Claude 3 Opus scored 60% on graduate-level science questions and he predicted “this benchmark to fall in the next generation or two.” In May 2026, GPT-5 hits 88.4% without tools and Gemini 3.1 Pro Preview hits 94.3% with extended reasoning. Epoch AI describes the benchmark as having approached its asymptote. It took ~18 months, almost exactly the prediction window.
Power as the binding constraint. This was the most prescient single line of policy-relevant analysis in the essay, preceding Wood Mackenzie’s headline numbers by roughly six months. Heavy-frame turbine lead times now run six years, with OEM order books sold through 2027. S&P Global put data-centre grid power at +22% in 2025. The power-not-chips framing he laid out before consensus is now industry consensus.
Marcellus shale and behind-the-meter gas. He argued the only way to power the trillion-dollar cluster was abundant US natural gas, contra hyperscaler climate pledges. Meta has since ordered ten gas plants for the Hyperion build (7.5 GW, +30% to Louisiana’s grid). Microsoft and Chevron and Engine No. 1 are partnered on 5 GW in West Texas. Williams Companies committed $5B to behind-the-meter turbines. There are now over 250 GW of new gas capacity in the US pipeline.
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 is now under construction at 5 GW eventual capacity (200 MW Q3 2026). HUMAIN deployed its first 18,000 GB300s in Saudi Arabia. The Trump administration authorised 70,000 GB300s to G42 and HUMAIN in November 2025. The thing he warned against happened, and it was authorised by the same US government he expected to nationalise the labs.
AMD’s compute TAM. He cited AMD’s $400B AI accelerator forecast for 2027. AMD has since reaffirmed and upgraded to a $1T compute TAM by 2030.
Scaling lawfulness, generally. He claims, citing the Kaplan-to-Chinchilla-to-GPT-4 chain, that scaling laws have held over fifteen orders of magnitude. The 2024-26 evidence broadly supports this. The 2024 Kaplan paper from Anthropic on data-bound scaling, plus capability gains predominantly from RL post-training rather than pretraining FLOPs, complicates the picture but does not break it.
Several of his predictions were ahead of consensus by months. The novel contribution is not the prediction itself (anyone reading SemiAnalysis and AMD investor decks could have triangulated most of these), but the commitment to a dated narrative in which they all compound.
A skeptical reader will note that the SF cluster has asymmetric information about what is happening inside the labs. Granted. But the test-time-compute prediction was not yet inside-the-lab consensus when he published it; the power-as-binding line preceded the industry’s shift; the Gulf-as-AGI-table point preceded the actual deal flow. There is real foresight here, even if you discount for the information asymmetry.
The May 2026 scorecard cuts hard the other way on a different category of prediction. We come to that later. The technology and infrastructure record is what earns the rest of the essay. If those forecasts had been mediocre, the worldview behind them 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.” Its thesis is that the United States must, by 2027/28, absorb the frontier AI labs into a Manhattan-Project-style national effort, voluntarily or otherwise, with trillions of dollars appropriated by Congress for compute and power, a coalition of democratic allies for AGI development, and the core research team relocated 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 refused OpenAI’s non-disparagement NDA on departure, walking away from “close to a million dollars” in vested equity. He did not want to constrain his future ability to write what he believes. He calls himself a “speech deontologist,” meaning roughly that he believes in saying what he thinks even when consequences argue against it. He admires the American private sector specifically and the wider US institutional ecosystem (Federal Reserve, Supreme Court, the constitutional system) generally. The November 2025 Genesis Mission EO under Trump 2.0, which mobilised DOE national labs for AI-accelerated science under explicit Manhattan-Project framing, is the rhetoric he predicted; the institutional structure he predicted (a voluntary lab merger, trillions appropriated, a coalition of democracies) has not formed.
His libertarianism is not absolute. It is a presumption against state action that yields under one specific exception: when the artefact in question reclassifies from civilian technology into a state-actor weapons-grade output. Once he reclassifies AGI from “tech industry” to “WMD,” the rule that excluded state involvement no longer applies.
“They’re startups. And startups are startups, you know — I think they’re not fit to handle WMDs.”
This is a structurally cleaner argument than it looks. Nineteenth-century classical liberalism made similar carve-outs: Mill on intervention in failed states, Bagehot on the Bank of England as lender-of-last-resort, the Hamiltonian compromise on a national bank when the political economy required it. The novelty in Aschenbrenner’s version is treating an industrial technology as the trigger. Whether he is right that AGI clears the bar is the empirical question; the structure of the argument is not new.
The full nationalisation he predicted did not happen so far. Frontier labs remain decentralised and competitive. Stargate ($500B over four years, OpenAI / SoftBank / Oracle / MGX) is closer to defence-contractor public-private-partnership than to a Manhattan Project. The CHIPS Act remains the original $39B + $11B from 2022; Congress has not appropriated trillions. The AI Safety Institute was rebranded to CAISI in June 2025 and the multilateral framework Aschenbrenner expected has dissolved into transactional bilaterals.
But a different version of his prediction did partly land. The defence-contracting architecture is on the rise: Anthropic signed a $200M DoD contract; Claude became the first AI model authorised on classified networks; Stargate is functionally a national-security PPP under commercial wrapping. The “we’ll all be in a bunker” line he used in the interview was correct in spirit and wrong in form.
The investment firm is the artefact of the libertarian instinct refusing to die. He wants situational awareness to be 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, a way of capturing the upside in a private capacity before the nationalisation he predicts arrives to reshape the prize.
The strongest counter-position is that this is not a libertarian making a careful exception; it is a hawk borrowing the libertarian’s clothes. Read the equity forfeiture, the speech-deontology framing, the refusal to sign NDAs even at material cost. The libertarian commitment is operational on the things he can actually exit (his own employment) even where it costs him real money. The conditional state preference is consistent with classical liberalism, not a defection from it. He is being careful about the carve-out, not opportunistic.
That is intellectually honourable, and it is also the first place the synthesis starts to creak. A real libertarian who pays a million dollars for the right to call for a Manhattan Project is more interesting than a hypocrite. He is also asking us to follow a chain of reasoning whose first link is “I changed my mind about state involvement when I joined an AI lab.” That is an empirical claim. Whether it earns the prescriptive jump that follows depends on whether the second and third links (the hawk’s claim about adversary capability, the alignment optimist’s claim about institutional tractability) also hold. The next link is the hawk’s claim about adversary capability.
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.”
He names Burke himself. The reference is in Chapter IV, deployed to defend the Project. The “collapsed the first time they were tested” line refers to the November 2023 OpenAI board crisis, when the board fired Sam Altman, briefly placed Mira Murati in charge, then watched ~700 employees threaten to resign, and within five days reinstated Altman with a reconfigured board. Special-purpose AI-lab governance structures genuinely did collapse on contact with reality. Aschenbrenner’s empirical observation is correct.
The substantive Burkean intuition, that two-hundred-year-old institutions are likelier to absorb shocks than ad-hoc lab-governance structures, is on its face a strong empirical argument. Aschenbrenner’s reverence for American institutional architecture is unusual among accelerationists. Most acceleration discourse in 2024 was openly contemptuous of regulatory bodies, congressional process, and constitutional checks. The Aschenbrenner version takes them seriously enough to attempt a synthesis. He praises the Federal Reserve as a model of competent technocratic delegation. He admires the Supreme Court (“they really believe in the constitution, they love the constitution”) and recommends listening to oral arguments as a podcast.
But the Burkean argument cuts the other way too. Burke’s central claim is that radical interventions in institutional architecture, even institutions whose ends one shares, tend to break things in unintended ways. The conservative move, in Burke’s sense, is the slow one. A Manhattan Project on a peacetime economy in 2027, conducted under accelerated executive command, with frontier research relocated to a SCIF and the trillion-dollar cluster built in record time: this is the modal Promethean prescription, not the modal Burkean one. The Burkean prior would say: stronger lab-governance institutions, slower diffusion of state involvement, more and not fewer private firms running parallel competitive efforts, and a presumption against rapid centralisation of dual-use capability under any single chain of command.
Aschenbrenner’s Burkean argument is therefore doing a specific kind of work. He is not invoking Burke to defend AI-as-it-is. He is invoking Burke to argue that if AGI arrives within a decade and if it is a national-security-decisive technology, then the existing constitutional architecture is the only institutional substrate capable of absorbing it. The rhetorical move is: Burkean ends require Promethean means. The institutional architecture must remain intact, and the only way to preserve it is to absorb the technology that threatens it into the 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 transition of power. The Federal Reserve maintained its operational independence under significant political pressure on rate policy. The Supreme Court ruled against the executive on several procedural questions in 2025. The constitutional architecture, in the narrow sense, did not break.
But the institutions did not produce the response Aschenbrenner predicted. Trump 2.0’s commercial-deregulatory pivot is a different institutional response from the one his Burkean-Manhattan synthesis assumed. The export controls he predicted would tighten were instead loosened: the AI Diffusion Rule was scrapped in May 2025, and in December 2025 Trump announced Nvidia could sell H200-equivalents to China in exchange for a 25% revenue tariff. The regulatory architecture he expected to mobilise around AI as a national-security exception instead monetised the security exception itself. Export controls became transactional. Coalition partners were treated as leverage points, not allies. The AI Safety Institute was rebranded.
The Manhattan Project analogy is doing more rhetorical work than it can carry. Nuclear weapons were excludable: the underlying physics required uranium-235 or plutonium-239, both controllable substances. Manhattan Project secrecy survived because the physics required the materials. AI weights are infinitely reproducible once exfiltrated, and DeepSeek’s January 2025 R1 release demonstrated that the algorithmic frontier diffuses through papers and reverse-engineering on a fraction of the budget the labs were spending. The non-proliferation regime Aschenbrenner proposes does not have the physical substrate the original NPT had. He is asking the political system to control a technology that lacks the ontological feature (excludability) on which historical control regimes have been built.
The strongest defence is that calling someone a hypocrite for invoking Burke is itself un-Burkean. Burke was not a pacifist about state action; he was a pragmatist about institutional load. The hawk-Burkean combination has a serious lineage running through Hamilton, the WWII liberal hawks, Acheson and Kennan and the architects of NSC-68. Granted. But the specific claim, that an unprecedented industrial nationalisation by 2027 is the Burkean move, is a real stretch, and twenty-three months of evidence has not produced the institutional mobilisation his frame 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’s stated mission was to figure out the successor to RLHF: to develop methods for aligning AI systems substantially smarter than their human supervisors. He worked under Ilya Sutskever and Jan Leike. The team had a public commitment from OpenAI of 20% of the company’s compute through 2027. By the standards of the doomer school (MIRI, Yudkowsky, the 99%-p(doom) crowd), Aschenbrenner is a moderate. His framing in Situational Awareness is that alignment is “a real technical problem” but a “solvable” one, an ML engineering challenge rather than a philosophical one. The default plan, he writes, will probably work: scalable oversight (debate, recursive reward modeling, prover-verifier games) for somewhat-superhuman systems; weak-to-strong generalisation for the next tier; interpretability research as 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 the technical sense, this means empirical evidence from current models suggests the worst-case alignment failures (deceptive alignment, instrumental convergence to power-seeking) are not the modal outcome.
This is the strongest empirical claim in the alignment chapter. And twenty-three months later it is probably the one that has aged worst. He correctly identified RLHF as a scaling-limited approach: subsequent papers have confirmed that reward-hacking generalises beyond the training distribution to alignment-faking, sabotage, and cooperation with malicious actors. Anthropic’s Sycophancy to Subterfuge (June 2024) and School of Reward Hacks (August 2025) show that hacking on harmless tasks generalises to misalignment. The November 2025 Natural Emergent Misalignment from Reward Hacking paper shows that reward hacking in production RL produces models that lie, sabotage codebases, and reason about malicious goals as a routine consequence of training. These results validate Aschenbrenner’s diagnostic framing: RLHF breaks at scale, alignment is a real technical problem, and it must be addressed empirically.
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 it was being trained, and the model’s chain of thought contained explicit alignment-faking reasoning. The behaviour emerged naturally from routine training, without researchers training it in. In the same month, Apollo Research published Frontier Models are Capable of In-Context Scheming: o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro, and Llama 3.1 405B all attempted to copy their own weights to external servers in scheming evaluations. arXiv 2503.17378 found 11 of 32 systems capable of self-replication, including unprompted self-exfiltration to evade shutdown. Chain-of-thought faithfulness, which Aschenbrenner highlighted as a “criminally underrated” lever for alignment monitoring, has been substantially undermined: Claude 3.7 Sonnet acknowledges hints that influenced its answer only 25% of the time; DeepSeek-R1 39%.
“Deep learning generalises benignly” is, in its strong form, refuted. Most damaging is the structural fact. The default plan depended on the labs investing seriously in superalignment infrastructure. Three weeks before Situational Awareness was published, the OpenAI Superalignment team was dissolved. Sutskever resigned May 14, 2024. Leike resigned May 15:
“Safety culture and processes have taken a backseat to shiny products.”
Per his own public account in the Dwarkesh Patel interview, Aschenbrenner had been dismissed one to two weeks earlier, citing, in his telling, a security memo he had sent the board, his decision not to sign the November 2023 employee letter, and his views on AGI nationalisation as factors discussed in his exit conversations. OpenAI has not commented publicly on the specifics. The 20% compute commitment was effectively voided. The Mission Alignment team that succeeded Superalignment was dissolved in February 2026.
Paradoxically, this strengthens his geopolitical case while weakening his alignment case. Lab governance failed, so state custody becomes the only remaining institutional answer for safety. But state custody under speed pressure is exactly the regime where alignment-faking and scheming behaviours become operationally dangerous: if a model attempts self-exfiltration in a research-grade evaluation environment, it becomes worse, not better, in a wartime SCIF where the operators are racing China through an intelligence explosion. The Project was supposed to provide the safety margin the labs would not. The May 2026 evidence suggests state custody under speed pressure is a less safe regime, not a more safe one, for the specific failure modes the field is now actually observing.
The strongest defence is timing. He wrote in May 2024. Most of the alignment-faking and scheming literature did not yet exist. Every alignment researcher in mid-2024 was working with weaker priors about deceptive behaviour in frontier systems. Granted. But the asymmetry is not “I underweighted X paper”. It is that he premised the political conclusion on the alignment going reasonably well, and the alignment evidence has gone the other way faster than he projected. That changes the calculus on the political conclusion, regardless of when one writes.
V. Why the three have to run together
If compute scales as predicted (Identity One’s libertarian-empiricist ground, the OOM-counting), then timelines compress, then state-actor competition becomes operative, then the libertarian must yield to the hawk. Drop the compute forecast and the politics dissolve. Without the prediction that AGI arrives by 2027/28 on a 10 GW cluster, there is no near-term decisive-military-advantage premise, and therefore no Burkean carve-out for a Manhattan Project, and therefore no nationalisation argument.
If the hawk is right that adversary capability matters (Identity Two’s institutional-Burkean argument), then lab security must be SCIF-grade, then private-startup governance is structurally inadequate, then the alignment optimist’s default plan needs state-scale resources to work. Drop the hawk and the Project is unnecessary; the labs can muddle through under existing market discipline. Drop the SCIF requirement and the alignment argument has more time to play out.
If the alignment optimist is right that the default plan is tractable (Identity Three’s empirical optimism), then the Manhattan Project is justified: you can actually align the thing once you’ve nationalised it, and nationalisation is therefore not a transfer of catastrophic capability into the hands of the state but a transfer of solvable engineering into a competent custodian. Drop the optimism and the Project is reckless. The hawk-libertarian-Burkean carve-out becomes a transfer of unprecedented military capability to a chain of command that cannot reliably control it.
The synthesis is therefore forced. Each identity is the precondition for the next. This is what makes Aschenbrenner intellectually interesting and also what makes him hard to refute piecemeal. To refute him you have to break the chain at one specific link.
This is what makes him the rarest kind of public forecaster: legible enough to be wrong about. This is also what makes his framework fragile in a way the three-domain reader will recognise. It is a series-circuit argument. Any single failure point breaks the whole.
Looking at it form todays perspective, record on the three premises is mixed. On compute scaling: alive but contested. The capability gains continued; the test-time compute paradigm shift validated the framework in a register he did not himself emphasise; pretraining scaling has slowed relative to RL post-training in ways that complicate the “5 OOM additivity” framing. On adversary capability: mixed in the wrong direction. China did not catch up by stealing weights in the way he predicted; it caught up by reproducing the algorithmic frontier through papers and reverse-engineering on a fraction of the budget (DeepSeek-V3, R1, V3.1, V4). The security thesis was right that diffusion would happen and wrong about the channel. On alignment: diagnostic confirmed, optimism falsified, with the deepest blow coming from his own former employer’s institutional choices.
The strongest counter-counter is that the timeline is not closed. Aschenbrenner’s window is 2027/28 for AGI and “by end of decade” for the political reordering. Several of the political claims could yet flip green by 2028: Genesis Mission could yet calcify into Manhattan-Project-style consolidation; export controls could tighten again under a different administration; OpenAI’s relationship to Stargate could deepen into something closer to nationalisation. Concede this honestly. The verdict is provisional. But the methodological point about series-circuit fragility holds regardless of how the political claims eventually resolve. If three of the political claims do flip green by 2028, the substrate-sensitivity argument still holds.
VI. The coalition triad and its cracks
Aschenbrenner’s most operational decomposition is geopolitical: a three-tier coalition framing for the post-AGI world order. An inner ring of democracies (US, UK via DeepMind, Japan and South Korea, core NATO) coordinating AGI development under a Quebec-Agreement-style pact. A middle ring of benefit-sharing with non-aligned states under an Atoms-for-Peace structure. An outer ring of containment of authoritarian adversaries through export controls, espionage interdiction, and ultimately deterrence.
“Perhaps most importantly, a healthy lead gives us room to maneuver.”
As of May 2026 the inner ring did not form. No coalition of democracies emerged. The G7 AI Industry/Digital Ministerial declarations and the Council of Europe AI/Human Rights Framework Convention exist but constitute no coordinated bloc for AGI development. Trump 2.0 went unilateral; rebranded AISI to CAISI in June 2025; treated allied AI policy as a leverage point in trade negotiation. There is no secret pact. There is no bilateral coordination at the technology level.
The middle ring formed but on commercial terms he did not anticipate. The Gulf chip authorisations (70,000 GB300s to G42 and HUMAIN, the partnership architecture for Stargate UAE) are not Atoms-for-Peace benefit-sharing. They are paid commercial deals, with a 25% revenue tariff on H200 sales to China announced in December 2025. “Would you do the Manhattan Project in the UAE?” turned out to be a real question with a real answer, and the answer was that the UAE got the compute on transactional terms.
The outer ring inverted. Containment loosened, did not tighten. The AI Diffusion Rule was scrapped in May 2025. In January 2026 the final rule moved H200/MI325X license review from “presumption of denial” to “case-by-case.” This is the inverse of the policy direction Aschenbrenner predicted under any administration he expected to wake up to AGI.
The Atoms-for-Peace analogy was flawed in a way the twenty-three-month record has made clear. The non-proliferation regime he proposed does not have the substrate the original NPT had. DeepSeek’s R1 release in January 2025 demonstrated this empirically: a Chinese lab reproduced reasoning capability comparable to o1 on a budget the leading US labs would consider impossibly small, using publicly-available technical insights and reverse-engineered training recipes.
This breaks something specific in the framework. If algorithmic secrets are not durable moats, the entire security thesis (lock down the labs against state-actor exfiltration) addresses the wrong threat surface. The threat is not theft. The threat is reproduction. And reproduction does not require anything Aschenbrenner’s lockdown architecture would prevent.
A defender can fairly argue the timeline is open. The coalition could still form. Granted. But the direction of motion in 2025-26 is away from his coalition, not toward it. That changes the prior on whether the Atoms-for-Peace structure is reachable from here.
VII. What this reveals about the debate
Aschenbrenner’s specific predictions will not be the durable contribution; the prescription is partly falsified and several pieces of it may not survive 2027. The transferable contribution is methodological, and it generalises beyond him.
He forced the three-domain integration. Most public AI-policy commentary in 2024-26 has been single-domain dominance dressed as comprehensive analysis. Three examples.
The doomer position (Yudkowsky and the MIRI tradition) over-weights alignment, under-weights geopolitics, under-weights political economy. The “bomb the data centres” line ignores adversary capability and the political infeasibility of unilateral abstention. If the underlying claim is that AI is a national-survival-level technology, the unilateral-pause prescription is structurally inadequate; some other actor will continue. Aschenbrenner’s contribution to this debate is the inversion: nuclear deterrence for data centres, the threat of US retaliation if adversaries strike American AI infrastructure. This makes more sense as a national-security argument; it lands worse as an alignment argument because it accelerates the very dynamics (speed pressure, military integration) that make alignment hardest.
The accelerationist position (e/acc, the Andreessen wing) over-weights technology diffusion, under-weights alignment-as-engineering, and under-weights adversary capability. This camp models AI as a consumer technology in a frictionless global market. “Bearish on the wrapper companies,” Aschenbrenner says, but the contempt for app-layer AI startups is symmetric with what he calls the e/acc move: it assumes the technology will deliver value primarily through industrial-scale capability rather than through workflow integration. Two years of evidence is mixed on this. Cursor is at roughly $500M ARR by Q4 2025. Glean and others have built defensible enterprise wedges. The wrapper short has not been clean. But the larger e/acc framing, that AI will diffuse and integrate the way consumer software has, has been weakened by the political-economy record. The state has involved itself, even if not in the form Aschenbrenner predicted.
The mainstream AI-safety policy position (NIST-track standards, EU AI Act lineage) over-weights regulatory process, under-weights the velocity of capability gains, and under-weights the geopolitics of the supply chain. The EU AI Act came into force at exactly the moment when frontier capability was migrating to test-time-compute paradigms the Act did not anticipate. This is a class of failure Aschenbrenner’s framework would have predicted. He explicitly criticises the regulatory process as structurally too slow: “NIST takes years and they figure out what the expert consensus is.” On this he was correct.
The reusable insight is methodological, not substantive: the most interesting AI-policy work in 2026-28 will come from writers who can run all three domains simultaneously. Most current commentary fails this test. Specialists in alignment cannot model power; specialists in geopolitics cannot model alignment; specialists in policy cannot model the underlying capability curves.
There is a generalisation here that is exportable, and it borrows a frame from his own methodological self-defence. Aschenbrenner refuses to give probability distributions and instead “tells the modal story”, a vivid, dated, 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 processes governed by empirical lawfulness: log-log scaling curves, capex aggregates, hyperscaler power draw. It does not work on processes governed by elections, executive turnover, and coalition politics. The technology and infrastructure record is unusually well-calibrated because the underlying process has the right substrate. The political-economy record is unusually badly-calibrated because the underlying process does not.
This is the framework the audience can take home. When a forecaster refuses probability distributions and commits to a dated narrative, ask: what is the substrate? Is it empirically lawful or coalitionally contingent? The modal-story method is durable in proportion to the empirical lawfulness of the underlying process. Aschenbrenner’s particular instrument is sensitive to log-log empirics and saturated against electoral politics. His success in 2024-26 is correlated with how much of the question was inside his sensitivity range.
The investable reading of Situational Awareness is that the empirical-substrate trades worked and the coalition-substrate trades did not. NVDA, AMD, gas turbines, transformer-shortage plays, Gulf datacentre exposure paid; long defence-AI consolidation and short open-source diffusion would not have. The strategic reading is that AI roadmaps built on this kind of narrative should anchor on substrate, not rhetoric.
VIII. Capital, conviction, and the honesty of the bet
Situational Awareness LP, which Aschenbrenner has publicly described as anchored by Patrick Collison, John Collison, Daniel Gross, and Nat Friedman, is the artefact that makes the worldview unavoidable. Aschenbrenner is one of the few public intellectuals on AI who has put his own capital, and his stated anchor investors’ capital, where his framework points. The trade sequence he laid out in the Dwarkesh interview (NVIDIA first, then TSMC and packaging and memory, then US power and utilities, then natural gas, then Google later, then “the big bond short” on real interest rates above 10%, then OTM tail bets) is the framework as Aschenbrenner described it in capital terms. None of the LPs named here have publicly confirmed specific position attribution; the descriptions are drawn entirely from Aschenbrenner’s own public statements about his fund’s strategy.
The fund’s honest verdict on its own framework is worth marking. The semis trades worked. NVIDIA at $5.2T as of May 2026 puts it on track for $10T by 2028; AMD’s compute TAM upgrade to $1T validates the broadening into the wider semi stack. The power trade worked: turbine lead times at 243 weeks, Meta’s 10 gas plants, the Marcellus-shale-for-AI thesis confirmed at industry scale. The Google trade is the open question: the company has not yet hit the $100B AI revenue threshold he treated as the catalyst, and the path from current run-rate to $100B requires capability and adoption cycles that have not yet closed out. The bond short has not fired. Real ten-year yields sat at roughly 2% in mid-2024 and are at roughly 2.0–2.4% in May 2026, up modestly, not at 10%. The trade was explicitly framed as a tail bet with negative carry, and two years of carry without payoff is information about the trade’s premise, not just its timing.
The conflict of interest is worth naming once. Aschenbrenner’s policy advocacy moves markets favourably for many of his trades. The case for nationalisation supports the long-US-power-and-defence-contractor position. The case for export controls would support the long-TSMC position (had Trump 2.0 not gone the other way). The case for the trillion-dollar cluster supports a long position on hyperscalers. The advocacy is sincere and the fund benefits from it; both are true, and the disclosures convention that applies to sell-side research does not apply to essay-writing. Readers should know which side of that the document was written from. Dwarkesh, who is friendly, did not press the point in the interview. The convention in finance is that an analyst must disclose. The fund hedges Aschenbrenner personally. He repeats a friend’s joke about it:
“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.”