Control over AI systemsâwho can develop, deploy, and direct powerful AIâis highly concentrated along multiple dimensions. A handful of companies (OpenAI, Anthropic, Google DeepMind, Meta) control virtually all frontier model development. Two countries (United States and China) dominate global AI capability. A few cloud providers control most AI compute infrastructure. And within these organizations, small groups of executives make decisions affecting billions.
This concentration has both risks and potential benefits. On the risk side: concentrated control means few actorsâ values are embedded in systems affecting everyone, single points of failure could have catastrophic effects, and power could be abused for private benefit. On the potential benefit side: fewer actors may be easier to coordinate and regulate, and concentrated resources enable safety investment that distributed development might not.
The trend is toward increasing concentration. Capital requirements for frontier models are rising rapidly ($1B+ per training run), creating insurmountable barriers for most actors. Talent pools are limited and increasingly captured by major labs. And first-mover advantages compound, making leaders harder to challenge.