You're missing the point. There was a lot of debate around if inference was subsidized or not. And that's a huge point to confirm in the public discourse.
Can I ask what is your opinion about their core CapEx, i.e. model training?
The general trend I observe is that the "shelf lives" of large language models are really short. It costs $1-10 billion to train cutting-edge models at the moment, and they only really last 6 months at best.
There seems to be very little brand loyalty too. Whenever a shiny new thing comes out, people just switch over, which implies that they constantly need to fight the time decay.
It's high, really high. But, that isn't bad. In fact... they are better of with it being extremely high. Then scale matters. They need enough revenue at high enough margins to earn a decent return on that spend, but higher is, from a competitive perspective, better.
I understand your logic ("the high CapEx is the moat"), but on the other hand, isn't it be a bit like multiple high speed railway systems trying to connect San Francisco to Los Angeles?
And there are three internal players chasing the same goal at the moment (OpenAI, Anthropic and Google), and two others (Deepseek and Alibaba/Qwen). What will prevent them from cutting the price floor each other?
Looking from a different angle: Microsoft has been able to maintain its monopoly because it was/is a huge pain for companies to switch the operating system, but do LLMs have that stickiness?
> You're missing the point. There was a lot of debate around if inference was subsidized or not.
To answer that question you have to take into account the cost to produce the thing that inference uses. If you don't, then that's like claiming that the total cost of a car is the cost to keep it on a dealer's lot until it's sold.
"Figuring out how much R&D adds to the total cost of a thing" absolutely isn't a new problem. And given that models seem to get supplanted every year, it's not like you're gonna be able to spread those R&D costs out very much.