It seems that each model generation, and each generation of GPUs, has a limited shelf life, so you need to make more money from each generation of models/GPUs t

It seems that each model generation, and each generation of GPUs, has a limited shelf life, so you need to make more money from each generation of models/GPUs than it cost you to build/buy them. User pricing to make a profit depends on inference volume. You need to make back a fixed dollar amount before breaking even, which could come from high volume and low prices, or lower volume and high prices. The trouble is that everything is changing so fast that any kind of forecasting is extremely error prone, especially when one forecast builds on another. First you need to guess how much more capable the models are going to get (at things where people will pay more for better performance), in what time frame, then guess what level of demand (inference volume) will exist with that level of capability... The LLM developers like OpenAI and Anthropic like to tout things like Math Olympiad and Competitive Programming results as signs of progress, but there is no guarantee that they will be equally successful in applying RL to more general areas of commercial value where RL rewards are harder to define. These companies also like to talk about "scaling laws" is if there was some inevitability about investing more money & compute and getting better results, but this only works until it does not, and they replace one broken "law" with another. Right now it's all about scaling of RL-training and test time compute, but how long will that last, and what type of problems will benefit? The profit model here seems a bit like the Drake equation for calculating the probability of other intelligent life in our galaxy... it may be possible to define the equation, but the outcome depends on having the right values for all the variables, which are largely unknown.