Anthropic's S-1 breaks down compute use between model serving and non-serving activities such as internal R&D, training, safety, and alignment.

“Yeah, exactly. All right, here's a chart, and this is broken down in their S-1, that shows the percent of their compute that goes towards basically serving models, so inference revenue generating, or internal R&D kind of training or maybe safety and alignment or, you know, non-serving of, uh, of, of tokens.”