Except that’s not what the analysis is. They’re spending < $1 to get $1 from you and the other $9 to figure out how to improve the model further and build up products on top of that to turn that $1 spend into $5 in the future.

In other words, inference is fairly profitable for them and the rest of the money is spent growing revenue as quickly as possible. Building models is still an expensive line item but the costs for that are going down with time.

There is also maybe a “capture the market” mentality but I don’t think that’s necessarily it - the tools and processes are largely fungible and that’s a huge problem. They need to figure out how to make it sticky for “capture the market”, but there’s also a very real “grow as big as possible as quickly as possible to take on Google”; Google has an existential threat here.