Which is why I said that it can be roughly estimated. And it can be roughly estimated even without those numbers assuming a fleet of some size X and assuming the number of hours this fleet is utilized per day, for the whole year. Either way, you will end up with a hefty number. Do the math and you'll see that inference is far from being cheap.

Of course all the requests is a hefty number, they're serving nearly a billion weekly active users. What else would you expect ? Google search, Facebook - those would all be hefty numbers. The point is that inference is pretty cheap per user, so when they get around to implementing ads, they'll be profitable.

Again, there are many indicators that inference per user is cheap. Even the sheer fact that Open AI closed 2024 serving hundreds of millions of users and lost 'only' 5B is a pretty big clue that inference is not that expensive.