You got it backwards. Training is extremely costly in many different ways, while running inference is much less expensive. Meta, Google, Alibaba, and some other

You got it backwards. Training is extremely costly in many different ways, while running inference is much less expensive. Meta, Google, Alibaba, and some others train their models and put them into the open for free burning through their cash reserves trying hard not to lose it all to OpenAI (and MSFT).