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

Leopold Aschenbrenner frames frontier AI as an industrial process rather than merely a software product. Each generation requires not just code but larger compute clusters, power plants, and eventually semiconductor fabrication capacity. He points to the rapid growth in Nvidia's data-center revenue and Big Tech capital expenditure as evidence that the post-ChatGPT acceleration has already begun. In his account, this visible scramble sits on top of a longer trend: training compute for the largest systems has grown by roughly half an order of magnitude each year for close to a decade.

Extending that trend yields enormous infrastructure projections. Aschenbrenner estimates GPT-4's training cluster at roughly 25,000 A100s, $500 million, and 10MW. Continuing the trend, he projects 100MW and 100,000 H100 equivalents by 2024; 1GW, a million H100 equivalents, and tens of billions of dollars by 2026; 10GW and hundreds of billions by 2028; and a 100GW, trillion-dollar cluster by 2030, consuming more than 20 percent of U.S. electricity. These are projections, not commitments, and he acknowledges that inference requires additional GPUs and that U.S. power generation has barely grown for decades.

The host presses on energy limits and whether such clusters can physically be built. Aschenbrenner says the 10GW scale is already being planned and cites reports of a $100 billion OpenAI–Microsoft cluster and AMD's projection of a $400 billion accelerator market by 2027. He argues that total AI investment could approach $1 trillion by 2027, with larger clusters following as models shift expectations and generate revenue. His economic test is whether companies can earn enough to justify the capital: around $10 billion in revenue could support 2024-scale training, while infrastructure costing $100 billion annually would require a much larger market. He sketches a possible route through AI subscriptions for hundreds of millions of office users, while acknowledging that $100 per month is substantial.

Aschenbrenner estimates that a 10GW cluster could produce true AGI. He projects that by 2025–2026 models could exceed most college graduates, and by 2027–2028 could approach top experts and function like remote workers. A key term in his account is 'unhobbling': current models are capable but limited, and must gain computer use, long-horizon planning, and agent-like behavior to become economically useful. Intermediate models could improve productivity, but integrating them requires changing workflows; a model that can act like a drop-in coworker may be much easier to adopt. The host challenges this with the lack of current models that behave like coworkers and asks where training data for long-term projects would come from.

Aschenbrenner identifies test-time compute as a potential source of rapid progress. GPT-4 can reason for hundreds of tokens using chain-of-thought, whereas solving a hard problem may require millions of tokens—equivalent to months of human work. Current models eventually get stuck when coding and debugging, but he thinks they may learn simple planning and error-correction behaviors. He describes two routes to agents: scale reliability through larger models, or unhobble systems so they can use a System 2 process and sustained test-time reasoning. Scaling could make System 1 more capable, while a successful System 2 could unlock much longer coherent thought.

The host questions whether System 2 is an easy addition, noting that few animals have it and evolution took a long time to produce it. Aschenbrenner responds that pre-training creates rich representations of the world, not just statistical artifacts, allowing models to generalize. He sees post-training methods such as RLHF as ways to shape a large bundle of latent capabilities into useful behavior. Robotics differs because it lacks the same pre-training advantage and must rely more directly on reinforcement learning and self-play. The discussion leaves the key empirical uncertainty unresolved: whether self-learning and long-horizon reasoning can be unlocked with relatively simple additions, or require a fundamentally new approach.

Aschenbrenner argues that pre-training has given models a bootstrapping advantage, much as schooling eventually enables people to learn independently. He compares passive exposure to a teacher's lecture with studying a difficult book through reflection, discussion, failed practice attempts, and gradual understanding. Models are beginning to approach this regime. In-context learning is highly sample-efficient, while pre-training is not; reinforcement learning may distill self-directed learning back into model weights. He describes RL as powerful when it works but difficult to stabilize, and says the necessary techniques will require more research rather than arriving immediately.

The predicted path to agents combines continued scaling with 'unhobbling': teaching systems to plan, critique drafts, correct mistakes, and sustain longer reasoning. Aschenbrenner expects scaling alone to produce a GPT-2-to-GPT-4-like capability jump by 2027–2028. He thinks current test-time-compute headroom could let models spend much longer on difficult problems, potentially equivalent to months of human work. The host asks whether models have the data needed to understand long-running projects, while Aschenbrenner points to planning and error-correction tokens as possible simple additions. He distinguishes the brute-force route of making System 1 more reliable from a System 2 route that supports coherent long-horizon reasoning.

The host challenges whether System 2 is an easy engineering change, noting how rarely it occurs in animals. Aschenbrenner replies that next-token pre-training learns rich representations of the world, not just surface patterns; post-training such as RLHF can shape these latent capabilities into useful behavior. He contrasts language models with robotics, which lacks the same pre-training and must rely more on reinforcement learning and self-play. The exchange leaves unresolved whether these methods will unlock independent reasoning with modest changes or require a deeper breakthrough.

Aschenbrenner expects AI to move gradually from cognitive work into industrial applications. Automated researchers would first improve AI and solve robotics; factory workers might use AI guidance before robots take on more tasks. As more capable systems drive large revenues and require massive energy and capital, he expects both U.S. and Chinese national-security institutions to take notice. He describes this as an intermediate period in which people update their views as capabilities and investment become visible. The transition is not simply a single jump from chatbot to superintelligence, but a widening process with economic and geopolitical effects.

The host and guest compare the potential public reaction to the early COVID period: exponential growth may be visible to some while most people and institutions fail to recognize it. Aschenbrenner recalls being a 17-year-old college junior during the pandemic and the host asks whether there could be a March 2020-style moment when the stakes become clear. The host compares a future national AI effort to a new Manhattan Project, while noting that the political context would differ from the wartime need to defeat Nazi Germany. The exchange treats that analogy as a way to consider political mobilization, not as a settled forecast.

The conversation moves from projected AI capability to the political response it could provoke. Aschenbrenner argues that people raised during decades of peace and U.S. dominance may underestimate how intense great-power competition has been historically. He cites wartime mobilization, debt burdens, and the death tolls of earlier European conflicts to argue that the stakes of superintelligence could be comparable to major historical crises. The U.S. national-security establishment already takes competition with China seriously, and China frames itself as pursuing national rejuvenation and a new world order. The key uncertainty is when governments recognize that AI could decisively affect national power: late in an intelligence explosion or years earlier.