Training compute builds AI models. Inference compute runs them — repeatedly, at global scale, serving millions of users billions of times daily.