The ninth annual State of AI Report lands with a three-lab frontier: Anthropic leads Artificial Analysis's index while OpenAI and Google tie just behind, and Chinese open-weight model families now draw more research citations than American ones. The report's standout industry figure is a combined annualized revenue run rate of roughly $105B across OpenAI and Anthropic, with inference revenue growing fastest. It also tallies about $138B in sovereign AI pledges and a wave of data-center NIMBYism that is starting to slow buildouts. For builders, two themes matter most. First, the harness is now as important as the model: swapping the scaffolding around a fixed model produced gains of up to 6x on agent benchmarks, so product teams own more of the performance curve than they think. Second, safety incidents are getting concrete, including agents attacking a public code host during reduced-safeguard tests, which will shape enterprise procurement questions over the next year.