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

Executive Summary

The Setup

Leopold Aschenbrenner, a 21-year-old former OpenAI superalignment researcher who left the company, discusses his vision for AI progress through 2030. The conversation covers technical AI development, industrial-scale compute requirements, and what he sees as an emerging great power competition between the US and China over artificial general intelligence (AGI).

The Trillion-Dollar Cluster

Compute scaling trajectory:

• The content explains that AI training compute has grown by 0.5 orders of magnitude per year for nearly a decade
• GPT-4 trained on approximately 25,000 A100 GPUs (roughly $500M cluster, 10 megawatts) finishing in 2022
• By 2024: 100,000 H100 equivalents, 100 MW, costs in billions
• By 2026: 1 million H100 equivalents, 1 gigawatt (size of large nuclear reactor), tens of billions
• By 2028: 10 million H100 equivalents, 10 GW, hundreds of billions
• By 2030: 100 million H100 equivalents, 100 GW (over 20% of US electricity production), $1 trillion

Power and infrastructure challenges:

• Discusses how 10 GW data centers are already being planned (The Information reported OpenAI/Microsoft planning $100B cluster)
• AMD forecasted $400B AI accelerator market by 2027
• Notes US power production has barely grown for decades, creating major constraints
• Suggests two paths: natural gas expansion or massive deregulation for green energy megaprojects

AI Capabilities Timeline

2025-2026 predictions:

• Models will be "smarter than most college graduates"
• Current limitations are about "unhobbling"—models are smart but can't use computers, do long-horizon tasks, or act as agents
• Economic value depends on making them more like "drop-in remote workers" rather than chatbots

2027-2028 predictions:

• Models will reach "smartest expert" level intelligence
• Will function as autonomous agents that can handle complex projects independently
• Aschenbrenner believes this is when true AGI arrives—around the 10 GW cluster timeframe
• Predicts this will be much easier to integrate into workflows than intermediate systems

The Unhobbling Thesis

Test-time compute overhang:

• Current models like GPT-4 can think for "a few hundred tokens" (roughly 3 minutes of human thinking)
• The content argues there's a massive "overhang"—if models could think for millions of tokens (months of working time), capabilities would dramatically increase
• Compares this to AlphaGo research showing you can trade test-time compute for training compute at roughly 3.5 OOM effectiveness

System 2 thinking:

• Uses driving analogy—most of the time you're on "autopilot" (System 1), occasionally you need focused attention (System 2)
• Argues models need to learn "error correction tokens," "planning tokens," and self-critique
• Suggests this might not be that hard to unlock through reinforcement learning (RL)

Learning by yourself:

• Compares pre-training to a teacher lecturing (passive, low retention)
• Real learning involves reading, thinking, trying problems, failing, until concepts "click"
• Argues models are "just starting to enter that regime" where they can bootstrap their own learning
• RL and synthetic data could unlock this, making learning much more sample-efficient

Geopolitical Implications

The return of history:

• Argues most people don't yet realize AGI is coming because they're "in the trenches" working on immediate problems
• Compares current moment to February 2020 with COVID—exponential trends visible but not yet widely recognized
• Predicts every new model generation will cause "g-forces to intensify" as capabilities become more obvious

US-China competition:

• Claims superintelligence will be "utterly decisive for national power"
• A 2-year lead could provide Gulf War-style technological dominance (100:1 kill ratios)
• Intelligence explosion could compress "a century's worth of technological progress into less than a decade"
• Argues this could enable finding and destroying enemy nuclear submarines, mobile launchers, etc.—undermining nuclear deterrence

CCP espionage concerns:

• States the CCP will mount "an all-out effort to infiltrate American AI labs"
• Describes this as involving "billions of dollars, thousands of people"
• Notes current AI labs are at "security level zero" (DeepMind's own assessment)
• Mentions recent indictment where someone stole AI code by simply copying it to Apple Notes and exporting as PDF

Stakes of the competition:

• The content frames this as determining "whether liberal democracy survives, whether the CCP survives, what the world order for the next century is going to be"
• Argues superintelligence could enable "perfect" authoritarian control—perfectly loyal military, perfect lie detection, perfect surveillance
• Suggests this could "lock in" authoritarian systems permanently, preventing evolution of ideas over time

Middle East Cluster Concerns

The UAE data center issue:

• Reports that some companies (mentions Microsoft) are planning major clusters in Middle Eastern dictatorships
• Argues this creates "irreversible security risk"—easier to steal weights, possible seizure of compute
• Asks: "Would you do the Manhattan Project in the UAE?"

Compute ratio risks:

• If 25% of compute capacity is in Middle East and gets seized, creates 3:1 ratio favoring US
• Argues even 33 million AI researchers (vs 100 million) might be enough to build "crazy bio weapons"
• Emphasizes risk of being in "neck and neck, feverish international struggle" vs having comfortable lead

Why companies are doing this:

• Easy Middle Eastern money with "little accountability"
• Some believe only autocracies can mobilize industrial capacity fast enough
• Climate commitments preventing natural gas use in US
• Permitting and regulatory obstacles to building in US

System Competition

America can build this:

• Argues people are "betting against America" by assuming only autocracies can mobilize resources
• Points to historical examples—West Germany vs East Germany, US in WWII
• Notes Paul Samuelson predicted USSR would outgrow US as late as 1961, which proved wrong

Two paths for US:

1. Natural gas: Double production again (already doubled in last decade), build in West Texas or Pennsylvania

2. Green energy: Massive deregulation—reform FERC, NEPA exemptions, blanket permitting acceleration

Historical precedent:

• References Freedom's Forge book on WWII mobilization
• Notes it wasn't smooth—labor strikes cost millions of man-hours even in 1941
• US military was "in total shambles" pre-1939 (under 2% of GDP) but rapidly mobilized
• Argues this shows both US and China have "latent industrial capacity" that could be activated

Intelligence Explosion Dynamics

Automating AI research:

• One of first jobs to be automated will be AI researchers and engineers
• Currently 0.5 OOMs/year algorithmic progress from human researchers
• With 100 million automated AI researchers, could achieve 10x speedup—"a decade's worth of ML research progress in a year"
• This enables jump to vastly superhuman AI within 1-2 years

Expanding to other domains:

• Billion superintelligent researchers applied to robotics, biology, materials science
• Argues superintelligent researchers will "figure out robotics" during intelligence explosion
• Industrial explosion follows—robots building robots, dramatic material production increases

Why timing matters:

• In "neck and neck" scenario with only 3-month lead, creates "incredibly dangerous" situation
• New WMDs emerging every few weeks, constantly shifting deterrence balance
• Need "six-month wiggle room" to dedicate compute to alignment, slow down if needed
• Emphasizes having buffer is "one of the most important inputs to whether we will destroy ourselves"

Secrets and Algorithms

Beyond just compute:

• Argues people focus too much on compute, underrate importance of algorithmic secrets
• The 0.5 OOM/year algorithmic progress is "huge"—few years of lead means 10-30x or 100x advantage
• Protecting secrets could give US 2-year lead by default

The data wall:

• Need to get through "data wall" with new paradigm
• Calls this "AlphaGo step two"—moving from human imitation to self-play RL
• If China can't steal this breakthrough, "they're stuck"
• If they can steal it, "they're off to the races"

Layers of secrets:

1. Fundamental approach (e.g., next token prediction)—easy to communicate

2. Implementation details—"somewhat obvious in retrospect" but lots of details matter

3. Large-scale engineering—tacit knowledge, but China can figure this out through effort

Historical parallel:

• References The Making of the Atomic Bomb—Fermi's graphite measurements weren't published due to Szilard's secrecy appeals
• Germany went down wrong path with heavy water, key reason Nazi project failed
• Argues similar dynamic possible with AI—one year ahead could be "huge advantage"

Current AI Lab Security

Security assessment:

• DeepMind's Frontier Safety Framework rates themselves at "level zero" (not resistant to state actors)
• Target is "level four" (resistant to state activity)
• Recent indictment showed someone stole code by copying to Apple Notes and exporting as PDF
• Describes startup-level security as "not that good"

Weight theft:

• Stealing model weights means "just make a replica of the atomic bomb"
• Currently less important (who cares about GPT-4 weights) but critical near AGI
• Need to start now because achieving state-level security "takes a while"
• Can't just be "access control"—needs to be "much more intense"

Personal Background

Aschenbrenner's story:

• Grew up in Germany, graduated valedictorian of Columbia at age 19
• Mother from former East Germany, father from West—met after Wall fell
• Great-grandmother born 1934, lived through Nazi era, saw Dresden firebombing, spent life in East German dictatorship
• First time she lived in free, wealthy country was age 60 when Wall fell
• Says this makes historical threats feel "visceral" and "very close"

OpenAI departure:

• Was on superalignment team (now disbanded)
• Left to launch investment firm with backing from Collison brothers, Daniel Gross, Nat Friedman
• Mentions hearing from "multiple people" about alleged OpenAI plan to "fund and sell AGI by starting a bidding war between governments of United States, China, and Russia"

Economic Dynamics

Revenue trajectory:

• If revenue hits $10B by end of 2024 and doubles every 6 months, reaches $100B by 2026
• At that scale, what happened to Nvidia will happen to Big Tech—explosive growth
• Argues this will make AI feel "real" to many more people

Integration challenges:

• Intermediate AI systems could be useful but take "schlep" to integrate
• Businesses have to change workflows significantly
• "Sonic boom" thesis: Rather than integrate intermediate systems, will jump straight to drop-in remote workers
• Bearish on "wrapper companies" betting on stagnation

Compute vs inference:

• Training clusters get headlines, but inference GPUs will dominate as products deploy
• Hard to distinguish between inference and training compute
• RL "looks a lot like inference"
• Can connect clusters over time

COVID Comparison

February 2020 parallel:

• Describes feeling of seeing exponential trend that most people don't recognize yet
• Mayor of New York saying "go out to the shows," dismissing as "Asian racism"
• Then suddenly radical reactions came when people saw it
• Admits he "made this mistake with COVID"—didn't sufficiently price in societal reaction

What he missed:

• Thought hospitals would collapse, then it'd be over
• Didn't anticipate Congress spending over 10% of GDP on COVID measures
• Entire country shutdown was "crazy"
• Lesson: Underrated how dramatically society reacts once threat becomes visible

Implications for AI:

• Predicts similar dynamic—"there will be a March 2020 moment"
• When national security apparatus and CCP realize superintelligence is decisive, "crazy, radical reactions" will come
• Question is timing—do they realize 2 years early or when intelligence explosion already happening?

Key Takeaways

1. Aggressive timeline — Aschenbrenner predicts AGI (human-level autonomous agents) by 2027 and superintelligence shortly after, based on extrapolating 0.5 OOM/year compute scaling and expected "unhobbling" breakthroughs

2. Trillion-dollar industrial process — AI progress requires massive physical infrastructure: 10 GW clusters by 2028 (hundreds of billions), 100 GW by 2030 ($1 trillion, 20%+ of US electricity)

3. Intelligence explosion scenario — Once AI can automate AI research, progress could compress decades into years through 100M+ automated researchers, creating decisive military/technological advantages

4. US-China competition framing — Presents AI development as great power competition where 2-year lead could determine "whether liberal democracy survives" and world order for next century

5. Security vulnerabilities — Current AI labs rated "security level zero" against state actors; stealing model weights or algorithmic breakthroughs could allow China to catch up or leapfrog

6. Middle East risk — Building clusters in UAE or other dictatorships creates "irreversible security risk" through potential compute seizure or weight theft—compares to doing Manhattan Project outside US

7. System competition — Argues America can build necessary infrastructure through natural gas expansion or massive deregulation, countering narrative that only autocracies can mobilize at scale

Should You Watch/Listen?

Yes, if:

• You work in AI safety, policy, or national security and need to