i believe the article is making the case that the market has already priced in the expectation that training (but especially training which is what NVIDIA is useful for) and inference will be continually more expensive and deepseek implies that the hardware required to train and infer really good models has a lower ceiling.