There's no reason not to think that within their current model architecture they accidentally discovered lucky configuration which helps them achieve better res

There's no reason not to think that within their current model architecture they accidentally discovered lucky configuration which helps them achieve better results than competitors. While companies still need tens of thousands of GPUs to train the models, we no longer need that for inference; with DeepSeek you can infer efficiently at a fairly modest price with comparable quality. As the model training is understood better, we might hit the ceiling where additional thousands of GPU-hours won't give any noticeable quality benefit and good results from training can be achieved even without investing hundreds of millions into hardware.