DeepSeek's newest model costs something on the order of 1/100th of OpenAI's, while being slightly dumber. Even still, DeepSeek is actually able to make a profit on their margins because they focus so many resources on making their models as efficient as possible.
OpenAI and Anthropic are betting on simply pumping more compute into the AI to get to superintelligence first. They don't want to waste time on optimization because they think once they get the superintelligent AI, it can just optimize itself after the fact. Any time you think that OpenAI and Anthropic won't make money or have some problem, they will just say that the superintelligence will fix everything like a deus ex machina.