Will There Be an AI Hegemon? A Mental Model for AI Power Concentration

On the Baker-Anthropic Conversation, Power Acquisition and Escape Velocity Will there be an AI hegemon, i.e. an entity that, by wielding superintelligence, acquires so much economic or political power that no rival, or coalition of rivals, can effectively constrain it? Are we about to go off-equilibrium, with the balance of power permanently skewed towards a single company or government? What makes it so plausible that some Anthropic representatives believe that there will be a single company in the future? That’s the topic that fueled the recent public conversation between investor Gavin Baker and Anthropic representatives Sholto Douglas and Dario Amodei.  In this post, I analyze these questions through two pairs of variables:  An actor’s absolute level of power and speed of power acquisition. Other actors’ coalescability (their propensity to coalesce), and combined coalitional counterpower.  An actor can become hegemonic if it passes a self-sustaining threshold of power faster than threatened actors can recognize the threat, coordinate, and deploy counterpower. I call this threshold “escape velocity”. If there is one AGI lab likely to believe we are on the verge of an unprecedented state of capitalism where a single company reaches escape velocity, it is Anthropic. Three sets of beliefs create a fertile terrain for believing in such an outcome:  They’re on the verge of building superintelligence, i.e. a form of capital that can substitute a large share of economically valuable labor. Once capital and labor become highly substitutable, there’s no more possible disruption of large companies by startups by being leaner. There’s no more “being smarter” if the main input to intelligence is spinning up agents; there’s no more “being more agile” if agents that one can spin in hours can self-manage micro-organizations able to execute end-to-end startup-scale projects. Result: the absolute level of power of whoever ends up unleashing superintelligence is absolutely unprecedented. Coasean bottlenecks to organizational scaling become drastically alleviated by agent speed of interaction and the organization first wielding superintelligence can scale in human-equivalent labor way further than any organizations to date.  Recursive self-improvement is here, starting this year with the country of genius in a datacenter, and will induce a speed of power acquisition that is absolutely unprecedented. Anthropic scaling its ARR from $9B to $47B in 5 months is the early sign of this; Anthropic representatives do not believe that it will stop. They expect this to continue; as Dario said on Dwarkesh’s podcast, he expects trillion-dollar scale revenue possibly within one to two years, which is on trend with their current growth rate.  End-to-end superhuman AI researchers are near, somewhere between 2027 and 2029. If this is true, the transition from “normal company” (e.g. $10B of revenue per year in 2025) to hegemonic company with tens of trillions of revenue could happen over less than 4 years. This speed of power acquisition may be sufficient to create a true global hegemon where the East India Company, Genghis Khan and Napoleon failed.  Notably, speed is a crucial part of the equation. The same transition to superintelligence over 10 years is much less likely to yield a hegemon, whether a company or a government. Historical wannabe hegemons often succeeded initially because they moved faster than their opponents could coordinate. Their global takeover attempts typically failed when their expansion slowed, giving threatened states time to form (sometimes improbable) countercoalitions. Napoleon’s France, the Nazi’s IIIrd Reich and Imperial Japan all followed a version of this trajectory. This same mechanism is what can still prevent entities building superintelligence from effective takeovers.   The core underlying dynamic is that at any given time, a range of other competing entities are fighting for their survival. These entities are all deeply intertwined in a network of alliances and rivalries that take time to overcome. Once a threat becomes sufficiently existential, these entities eventually overcome their differences to fight the common enemy. The best example is the US and USSR, vehement ideological enemy who would go on to fight a multi-decade Cold War, coming together and coordinating to take down the IIIrd Reich once it became highly threatening. This unlikely union demonstrates the essence of coalescability: given enough time, actors with major rivalries can coordinate to defeat a sufficiently threatening common enemy.  Similarly, the scaling speed of Anthropic revenue past $10B has hit many who share different beliefs on AI progress by surprise; we may be now witnessing the first attempts by large-scale entities to fight back, such as Meta unleashing a price war, and the first coalition of the largest AI companies coming together on open-weight symbolically against Anthropic. Anthropic’s recent spats with the Am