“The proportion of compute dedicated to post-training is expected to increase, based on the reasoning that model-generated output is often of higher quality than the average content on the web.”