Today, OpenAI's latest article, *Research acceleration: The view inside OpenAI*, indirectly confirms what Dylan Patel previously mentioned. Dylan Patel had earlier stated that, for cutting-edge model companies, in the long run, allocating more compute resources to next-gen model research might be more valuable than immediately using it to serve users, sell tokens, or generate short-term revenue. OpenAI's article today makes this logic even clearer: The previous generation of models is becoming the research infrastructure for the next generation. Models are now participating in writing research code, debugging infrastructure, analyzing experiments, generating synthetic data, and completing long-cycle research tasks. This means that the compute required to develop the next generation of models in the future will go far beyond just that one massive frontier training run. A significant amount of compute will also be consumed by AI researchers, RL rollouts, synthetic data generation, verifiers, evaluations, simulations, and numerous parallel experiments. This is why Dylan previously said that the inference compute used for researching next-gen models could, over the course of a year, surpass the compute used for that single frontier training run itself. It also means that the definition of how much compute an AI company has for research needs to be redefined. It should include both training compute and the inference compute required for research. Dylan believes the latter could be several times the former, which is why OpenAI might be allocating more than half of its compute to new model research. If we look at it economically: Compute → Inference → Short-term revenue. But one GPU for an AI researcher: Compute → Research → Better Model. And stronger models will, in turn, create stronger AI researchers: Better Model → Better Research → Faster next-gen models. Thus, internal research compute starts to take on the nature of a "compounding asset." Therefore, if today’s cutting-edge labs are allocating roughly 50-60% of their compute to internal research and 40% to external inference, it’s not hard to imagine that in the future, as we enter the automated researcher phase, the internal allocation could rise to 70%, 80%, or even higher.