SemiAnalysis points out that the computing power allocation of the cutting-edge large model lab shows a clear divergence from the overall industry trend: the training end still accounts for more than 60% of its total computing power, while inference accounts for only slightly more than 30%, and resources have not been massively shifted to inference services for external use as the model goes online; at the same time, the computing power structure within training has also undergone a significant shift, with the focus of resource investment gradually shifting from traditional base pre-training to post-training and reinforcement learning stages.
