“Coordination on AI safety among major labs is feasible because training the largest models is highly capital-intensive and complex, limiting the number of entities capable of doing so.”