George Hotz has never been one for subtle arguments. The comma.ai founder and serial provocateur published a new essay today framing the closed-source AI model as something more than a business strategy. He calls it neofeudalism.
The core argument: when a handful of labs, bankrolled by major cloud providers and increasingly intertwined with state interests, control which AI systems get built and how they get deployed, you don't get a market. You get a feudal structure where a small class of custodians decides what forms of dependency the rest of us should accept.
The Structural Problem, Not the People Problem
Hotz is careful to frame this as institutional, not personal. He's not claiming Sam Altman or Dario Amodei are bad actors. His point is that the institutional form of frontier labs - concentrated compute, concentrated talent, concentrated deployment authority - inevitably centralizes power regardless of who's running things. Good intentions don't fix structural incentives.
This is a more interesting argument than the usual "open source good, closed source bad" debate. Most open-source advocates focus on code access or model weights. Hotz is talking about something broader: who gets to decide what intelligence looks like and who can use it.
What He Proposes
The essay outlines what Hotz calls a "free technical order" built on several principles: multiple competing model lineages instead of approved single systems, auditable alternatives where full openness isn't practical, local inference (running AI models on your own hardware instead of renting cloud access), affordable hardware to prevent artificial scarcity, and user rights to inspect, modify, and fork AI systems.
None of these ideas are new individually. The local inference movement is already happening with projects like llama.cpp and Ollama. What Hotz adds is the political framing - connecting the dots between technical architecture choices and power structures.
Where It Falls Short
The essay is long on philosophy and short on specifics. There are no numbers, no concrete examples of harm, no named companies. It reads like a manifesto draft rather than an analysis. The comparison to feudalism is provocative but imprecise - medieval serfs couldn't switch lords, while you can switch from ChatGPT to Claude in about thirty seconds.
Still, the underlying tension is real. The cost of training frontier models keeps climbing, the number of organizations that can afford to build them keeps shrinking, and the regulatory conversations happening in Washington and Brussels tend to favor incumbents. Whether "neofeudalism" is the right label or not, the concentration trend in AI development deserves more scrutiny than it's getting.