Expensive part is providing the services at loss to get the biggest part of the market, which is the current state of affairs. At some point the market leader will emerge and it will price things according to the actual cost of running this thing + markup, and ALL efforts would be going into squeezing as much profit as possible from the leading position. At the same time usability and quality would not be priority anymore. We are enjoying the late-capitalism at it's peak here.
I'm just saying what Anthropic is saying: individual models are printing money. The problem is all the capital needed to get the next model before the previous ones are obsolete; investors capitalize the business because the future models are expected to print even more money, at least a few months ago. It's increasingly obvious that the new models won't print as much money as expected to justify the expenses, but if they all stopped training today they'd be swimming in cash in a couple years.
>if they all stopped training today they'd be swimming in cash in a couple years.
If they stopped training, then the data cutoff date is one year farther each year? How do you make money from model which is stale on data and doesn't include any recent stuff?
Not staying behind is the expensive part.