Swarm Scaling Law

AI task performance can follow a distinct 'swarm' scaling law: adding more cooperating agents improves performance differently than spending more compute per agent. Early empirical evidence (OpenAI’s multi‑agent Navier‑Stokes effort and Toby Ord’s analysis) suggests a parallelizability exponent around 0.57, which makes large agent swarms much faster but less compute‑efficient than lengthening single‑agent runs. — If true, swarm scaling reshapes AI benchmarking, infrastructure demand, economic growth trajectories, procurement choices, and regulatory priorities because speed‑first multi‑agent strategies produce new winners, risks, and chokepoints.