Leopold Aschenbrenner — 2027 AGI, China/US super-intelligence race, & the return of history

ELI5/TLDR

A former OpenAI researcher argues that today’s straight-line trends in computing power point to AI matching the smartest humans by around 2027, and that this will trigger a chain reaction: smart AIs that improve AI, leading to something far beyond human intelligence within a year or two of that. His main claim is that this is not just a story about better products. It is a national security story. Whoever gets there first, and whoever can keep their secrets from being stolen, may win a military and economic advantage as lopsided as the one America had in the first Gulf War. His warning: history, with its dictatorships and great-power struggles, is about to come back, and almost nobody in power is paying attention yet.

The Full Story

Leopold Aschenbrenner is the guest, and the timing matters: he recorded this with Dwarkesh Patel the same day he published Situational Awareness, a long essay series laying out his thesis. He grew up in Germany, graduated as Columbia’s valedictorian at 19, worked on OpenAI’s now-disbanded “superalignment” team, and is starting an investment firm with backing from the Collison brothers, Daniel Gross, and Nat Friedman. The conversation is the case for taking AGI seriously not as a tech demo but as the central geopolitical fact of the coming decade.

The straight line on the graph

The spine of the argument is a single trend. For roughly a decade, the computing power used to train the biggest AI systems has grown by about half an order of magnitude per year. (An “order of magnitude,” or OOM, means a 10x jump. Half an OOM a year means roughly tripling annually.) Aschenbrenner just plays that line forward.

GPT-4 was reported to have finished pre-training in 2022… a cluster size of about 25,000 A100s. That’s roughly a $500 million cluster. Very roughly, it’s 10 megawatts.

From there the numbers escalate fast. By 2026, a gigawatt cluster — the power of the Hoover Dam, costing tens of billions. By 2028, ten gigawatts, more power than most US states. By 2030, the “trillion-dollar cluster” drawing 100 gigawatts, over 20% of US electricity production. A “cluster” here is just the enormous warehouse of linked chips that trains a model. His point is that AI has quietly become an industrial process. The next model is not code; it is power plants, chip fabs, and steel.

He waves off the objection that energy will be the bottleneck (an argument he attributes to Mark Zuckerberg). Six months ago, he says, ten-gigawatt data centers sounded insane; now they are “the talk of the town.” The money is following: Nvidia’s data-center revenue went from a few billion a quarter to $25 billion in a year, and AMD has forecast a $400 billion AI-accelerator market by 2027. Will it pay off? He does the back-of-envelope math: sell a $100/month AI add-on to a third of Microsoft Office’s 300 million subscribers and that alone is $100 billion in revenue. For a knowledge worker, he argues, $100 buys back only a few hours of productivity a month — a low bar.

Unhobbling: why smart models aren’t yet useful

Here is the conceptual heart of the episode, and the part worth slowing down for. Today’s models are, in his words, “smart but limited.” They can answer a question but cannot go off and do a project. Aschenbrenner calls the fix “unhobbling” — removing the constraints that stop a smart model from acting like a competent worker.

The key bottleneck is what he calls the “test-time compute overhang.” Think of it this way. GPT-4 “thinks” for a few hundred words before answering — roughly equivalent to a human thinking for three minutes. A human, given a hard problem, can think for months: make a plan, write a draft, critique it, try, fail, fix the error, try again. Today’s models cannot sustain that. They write some code, get stuck, and can’t correct their own mistakes.

If you tried to answer a math question by saying the first thing that comes to mind, you wouldn’t be very good.

He borrows a useful frame from psychology: System 1 (fast, automatic, like driving on autopilot) versus System 2 (slow, deliberate, like working through a construction zone). Scaling improves the autopilot. The breakthrough that’s needed is teaching models the System 2 process — planning, error-correction, sustained reasoning over millions of words. If you crack that, you unlock months of equivalent thinking time on a single problem, and the model stops being a chatbot and becomes “a drop-in remote worker” you interact with over Slack and Zoom.

Why does he think this is solvable, and soon (his guess: six months to three years)? Because of an analogy to how humans learn. Pre-training — feeding a model the entire internet — is like a teacher lecturing while the words fly by; you absorb a little. Real learning is active: you read a dense page, argue with a study buddy, try a practice problem, fail, and at some point it clicks. That moment of working-it-out-yourself is the richest possible training signal, and reinforcement learning (RL) is the attempt to give models that same loop. He thinks models are just now entering the stage where, like a college student, they can start teaching themselves.

If both forces compound — raw scaling plus unhobbling — he expects another “preschooler to high schooler” leap on top of GPT-4 by 2027-2028. That, in his framing, is true AGI: drop-in remote workers smart enough to replace cognitive labor.

The intelligence explosion

The next move is the one that turns a productivity story into a geopolitical one. The first job to be automated, he argues, is the AI researcher’s own.

If you can automate AI research, things can start going very fast.

Run tens of millions of GPUs, populate them with 100 million automated researcher-equivalents, and you might compress a decade of machine-learning progress into a year. That lets you jump from human-level AI to “vastly smarter than humans” inside a year or two. Then the explosion broadens: a billion superhuman researchers turn their attention to robotics (which he reframes as fundamentally a software problem), then biology, then every field of R&D, with robots eventually building the physical world to match.

The picture he paints of the transition is gradual at the edges but compressed at the center. First, factory workers wearing AR glasses and earpieces, coached in real time by an AI that can turn anyone into a skilled technician (Meta’s Ray-Bans, he notes, are a complement to its Llama models). Then the robots arrive.

History is back

This is where Aschenbrenner gets out over his skis deliberately, and it is the most distinctive thread of the conversation. His claim is that superintelligence applied to military R&D could compress a century of progress into under a decade, producing a decisive advantage like the one coalition forces had in the 1991 Gulf War — a 100-to-1 kill ratio off the back of 20-30 years of technological lead. Picture, he says, billions of mosquito-sized drones that locate and neutralize nuclear submarines and mobile missile launchers — an advantage that could “preempt nukes.”

He insists his generation has been lulled by an unusual 80 years of peace and American hegemony into thinking nothing of historical consequence ever happens. The actual historical norm is brutal: in the Thirty Years’ War, up to half the population of swaths of Germany died; the Seven Years’ War killed 20-30% of Prussia; both world wars saw countries borrow over 100% of GDP. His bet is that when the national security establishments of the US and China finally grasp what’s coming, forces “we haven’t seen in a long time” will activate. The stakes, as he frames them, are whether liberal democracy survives and whether the Chinese Communist Party survives.

He makes this personal. His great-grandmother, born in 1934, watched the firebombing of Dresden as a child, then lived most of her life under the East German dictatorship — a son jailed by the Stasi for trying to ride a motorcycle across the Iron Curtain — and only first lived in a free country at nearly 60. Reading The Gulag Archipelago, he dwells on how superintelligence could make dictatorship permanent: perfectly loyal security forces, perfect lie detection, total surveillance, no Gorbachev ever rising because every doubter is identified and removed. A regime where “truth is what the party says,” locked in forever.

The race, the theft, and the secrets

If you accept the timeline, the operational question becomes: who gets there first, and can they keep it? Aschenbrenner’s answer is that American AI labs are wildly insecure. DeepMind’s own safety framework rates security from level zero (none) to four (resistant to state actors); the labs, he says, are at level zero. He cites a real indictment: a man stole critical AI code and took it to China simply by pasting it into Apple Notes and exporting a PDF — slipping past all monitoring. “I would think of the security of a startup,” he says. “It’s not that good.”

He separates the threat into three layers:

The weights — the trained model itself. Steal these near the moment of AGI and you have “a direct copy of the atomic bomb.”
The algorithms / secrets — the half-OOM-a-year of algorithmic progress. He argues people fixate on compute (sexy) and underrate secrets. Much of the recent progress (transformers, Chinchilla scaling laws, mixture-of-experts) was openly published, which is exactly why Chinese models are decent. If the US started actually keeping the next paradigm secret, a few years of protected lead could mean a 10-100x effective advantage.
Tacit engineering knowledge — the schlep of making giant training runs work. This, he concedes, China will figure out.

His historical touchstone is the Manhattan Project. Leo Szilard fought bitterly for secrecy; he persuaded Enrico Fermi not to publish a key result about graphite as a reactor moderator. Because that stayed secret, Nazi Germany went down the wrong path (heavy water) and fell hopelessly behind. Aschenbrenner asks whether America will “instantly leak how to get past the data wall” or guard it.

Dwarkesh pushes back well here, repeatedly. If progress is an industrial process of compute plus published algorithms, doesn’t that make catch-up easier? If the key insight can be written “on the back of a napkin,” why can’t China just figure it out themselves? And he offers a sharp counter-anecdote from Richard Rhodes: the Soviets, ordered by the ruthless Lavrentiy Beria to copy the American bomb exactly, would have built a better bomb had they trusted their own design — and great inventions (the light bulb, the bomb) tend to arrive in parallel because the tech tree dictates what’s next. Aschenbrenner’s reply is that even a six-month-to-two-year lead is decisive when each year of the intelligence explosion might be the gap “between a system that’s human-level and a system that is vastly superhuman.”

Why it has to be built in America

A running argument concerns the Gulf states. There is “free-flowing Middle Eastern money” trying to build clusters, and reports of companies (he names Microsoft, and references Sam Altman’s reported multi-trillion-dollar chip ambitions) courting it.

Would you do the Manhattan Project in the UAE?

His objection is that putting AGI-capable clusters in authoritarian states creates an irreversible security risk: easier weight theft, easier compute seizure when things get tense, and implicit leverage — “a seat at the AGI table” — for dictators who otherwise have nothing but money. He proposes a two-tier structure: a narrow coalition of democracies builds AGI, while a broader coalition (including dictatorships) gets to use AI products and older models — an “Atoms for Peace” arrangement.

The counterargument — if the US won’t work with the UAE, they’ll just back China — he treats as partly fair (hence benefit-sharing) but partly a bluff: the UAE is export-controlled and money alone doesn’t translate into AI progress. He’s also openly suspicious of the labs’ motives, relaying a secondhand claim he hasn’t verified: that OpenAI leadership once sketched a plan to fund AGI by starting a bidding war between the US, China, and Russia. “That’s pretty fucked up,” Dwarkesh says.

He’s adamant America can build domestically. The obstacle isn’t capacity but will. Two paths: natural gas (West Texas, the Marcellus Shale — US gas production has nearly doubled in a decade), blocked mainly by corporate climate commitments; or a green megaproject of solar, batteries, small modular reactors, and geothermal, blocked by permitting, FERC, and NEPA. He reaches for Freedom’s Forge and WWII mobilization to puncture the myth that mid-century America had effortless state capacity: labor strikes crippled aircraft factories into 1941, and the pre-war military was a shambles at under 2% of GDP. The latent capacity was there; the country eventually got its act together. His worry is that some people “shitpost about loving America” while privately betting the autocracies will out-build it — a bet he calls simply wrong.

Key Takeaways

Training compute for frontier AI has grown ~0.5 orders of magnitude (roughly 3x) per year for a decade; extrapolating gives a 100-gigawatt, trillion-dollar training cluster (>20% of US electricity) by ~2030.
The GPT-4 cluster cost ~$500 million to build (not the oft-cited $100 million, which is only the rental price).
AGI as a “drop-in remote worker” — as smart as the smartest experts, able to run multi-step projects — is his best guess for 2027-2028, tied to ~10-gigawatt clusters.
The missing ingredient is “unhobbling,” especially unlocking the “test-time compute overhang”: teaching models sustained System 2 reasoning (planning, error-correction) so they can think for the equivalent of months on one problem.
He estimates the System 2 / RL breakthrough is 6 months to 3 years away — “not that hard.”
Automating AI research itself is the trigger for an “intelligence explosion”: ~100 million automated researchers could compress a decade of ML progress into a year, jumping from human-level to vastly-superhuman AI in 1-2 years.
A 6-month-to-2-year lead could translate into a Gulf-War-style decisive military advantage — potentially one that “preempts nukes.”
Superintelligence could make dictatorship permanent (perfect surveillance, lie detection, loyal forces), removing the pluralism that historically corrects bad regimes.
US AI labs have near-zero security against state actors (he cites the Apple-Notes-to-PDF code theft); weights, algorithmic secrets, and tacit knowledge are three distinct theft vectors.
He argues “secrets” (algorithmic progress) are underrated versus compute; keeping the next paradigm secret could compound into a 10-100x effective lead.
Building AGI clusters in Gulf dictatorships creates irreversible risks (weight exfiltration, compute seizure, leverage); he favors a two-tier democratic coalition for AGI plus benefit-sharing for others.
America’s constraint on building domestically is political will (climate commitments, permitting/FERC/NEPA), not capital or physical capacity.
His framing: the public, governments, and national security states will have a “March 2020 moment” — a sudden, COVID-like collective realization that AGI is the central thing happening.

Claude’s Take

A note on completeness first: the cached transcript runs to about line 1,473 and cuts off mid-sentence (“…or ‘data wall’”). The video is 4h32m; what’s captured here is roughly the first hour. This summary is faithful to that portion — the compute/scaling case, unhobbling, the intelligence explosion, the geopolitics, and the security argument — but the back half of the interview (likely the alignment deep-dive, the “superalignment” team’s demise, his investment thesis, and Dwarkesh’s harder object-level pushback) isn’t in the file and isn’t covered. Worth knowing before treating this as the whole conversation.

On the substance: this is one of the most influential AI essays-as-podcast of its era, and it rewards engagement even where it overreaches. The strongest part is the descriptive core — the compute trend, the cost mechanics, and especially the unhobbling argument. Those are grounded in real numbers and real research, and the System 1/System 2 framing is a genuinely clarifying way to think about why a model can ace a benchmark yet fail to act like a coworker. Two years on from recording, the “test-time compute” thesis has aged unusually well; reasoning models built on exactly this idea did arrive, which lends his framework retroactive credibility.

The argument gets progressively more speculative as it stacks. Each step — scaling continues smoothly, unhobbling is easy, AI researchers automate themselves, that produces a fast explosion, that yields a decisive military edge — is individually plausible and collectively a tower where the conclusions inherit every prior uncertainty multiplied together. The intelligence-explosion step in particular smuggles in a lot: that ML progress is bottlenecked mainly by researcher quantity rather than by compute, experiment latency, or ideas that don’t parallelize. Dwarkesh’s pushback is the most valuable thing in the transcript, and he’s right to press the parallel-invention point (the Soviet bomb anecdote is a real dent in the “secrets are decisive” thesis). Aschenbrenner’s “two years makes all the difference” is an assertion, not a demonstration.

The geopolitics is where confidence and evidence diverge most. The historical sweep — that peace is the anomaly and great-power brutality the norm — is bracing and probably healthy as a corrective to complacency. But “superintelligence equals a Gulf-War-style decisive advantage” is an analogy doing the work of an argument; the leap from “much smarter R&D” to “mosquito drones that neutralize the nuclear deterrent” skips the entire question of whether physics, manufacturing, and counter-measures cooperate on his timeline. Treat the timeline and the military determinism as the speculative load-bearing beams. Treat the compute trend, the cost math, and unhobbling as the grounded foundation.

Scoring it 8/10. It is clear, original, unusually concrete for futurism, and it genuinely shifted how a lot of serious people frame the AI-and-power question. It loses points for a confidence level that outruns its evidence on the geopolitical and intelligence-explosion claims, and — for this vault specifically — because the cached transcript only covers the opening stretch. Still essential listening if you want to understand the worldview that’s now baked into how Washington and the labs talk about each other.