Leopold Aschenbrenner: 2027 AGI, the China/US Super-Intelligence Race, and the Return of History

Podcast: Dwarkesh Podcast
Host: Dwarkesh Patel
Guest: Leopold Aschenbrenner — former OpenAI superalignment team, now launching an investment firm backed by the Collison brothers, Daniel Gross, and Nat Friedman
Duration: ~4 hours 30 minutes
Listen: Apple Podcasts | YouTube

Leopold Aschenbrenner graduated as valedictorian from Columbia at 19, worked on OpenAI’s superalignment team, and wrote the “Situational Awareness” essay series. This conversation covers his full thesis: the path from GPT-4 to AGI, the trillion-dollar clusters it will require, and why superintelligence will reshape global power.

The Trillion-Dollar Cluster

Aschenbrenner’s central projection: AI training compute is growing at 0.5 orders of magnitude per year. By 2026, clusters will consume 1 gigawatt (tens of billions of dollars). By 2028: 10 gigawatts, hundreds of billions. By 2030: 100 gigawatts, over $1 trillion, representing 20% of US electricity production.

This is no longer a software story. AI development has become an industrial process requiring massive physical infrastructure, and that makes it inherently geopolitical.

The Unhobbling

Current models like GPT-4 have latent capabilities that are locked because they cannot engage in extended reasoning. Aschenbrenner calls this the “test time compute overhang.” The next big jump will come when models can think for millions of tokens (equivalent to months of human reasoning) instead of hundreds.

He draws an analogy to AlphaGo: the same model, given more time to think, became dramatically more capable. The same dynamic will play out in language models.

The Intelligence Explosion

Once AI systems can do AI research themselves, the dynamic changes entirely. Running 100 million human equivalents of AI researchers could compress decades of progress into years. This acceleration cascades beyond software to robotics, materials science, and military technology.

The Geopolitical Stakes

Aschenbrenner warns that the Chinese Communist Party will eventually recognize superintelligence as decisive for national power. He expects “an all-out effort to infiltrate American AI labs” involving billions of dollars and the full force of the Ministry of State Security. Current AI lab security is startup-level, not state-resistant.

He is sharply critical of building AI clusters in the UAE and other Middle Eastern countries, calling it shortsighted. “Would you do the Manhattan Project in the UAE?”

The Historical Lens

Aschenbrenner draws from his German family background — experiences with Nazi Germany, East German communism, and the Cold War. He emphasizes how quickly liberal democratic norms can collapse when superintelligence-enabled surveillance becomes possible.

His prediction: as AI capabilities become obvious through successive model releases, the response will resemble COVID-19’s March 2020 moment — sudden, total recognition that changes everything.

Crepi il lupo! 🐺