We are still in the Uber-subsidized part of the relationship.
It’s not comparable, the costs for Uber are fixed development costs which decrease per customer with more customers. AI is strictly more expensive the more users there are since the cost is per use. The financial outlook for the industry is very bad, it will probably never be profitable except maybe in some extremely niche situations.
Yeah, but they never really systematically subsidised fares as such, except for some one-off bonuses to attract divers and stuff like that, in the sense that they didn’t systematically pay drivers more than people were paying per ride. They subsidised it in the sense that the cut they took didn’t cover their development costs, servers, marketing, etc. But those costs don’t increase linearly per customer and they also plateau as the software stack matures, so there’s a path to just raising prices and getting to profit, and each additional customer brought them closer to profit even while the subsidies existed.
OpenAI and Anthropic are paying something like 10x the amount for inference as they get from subscriptions, let alone free usage and training costs. So each new customer is taking them further from being profitable. And if they jacked up the prices 10x, so that the basic subscriptions were a few hundred and the pro ones a few thousand a month, they would still be in the position that Uber was when they were doing the subsidies for their infrastructure. I think it’s fairly obvious that they wouldn’t be acquiring customers very quickly at those prices.
I think they mean the company since they’re the most famous example of subsidising the product to gain market share and their name is invoked all the time in discussions of AI economics
It’s not comparable, the costs for Uber are fixed development costs which decrease per customer with more customers. AI is strictly more expensive the more users there are since the cost is per use. The financial outlook for the industry is very bad, it will probably never be profitable except maybe in some extremely niche situations.
Uber fares are also per use…?
Yeah, but they never really systematically subsidised fares as such, except for some one-off bonuses to attract divers and stuff like that, in the sense that they didn’t systematically pay drivers more than people were paying per ride. They subsidised it in the sense that the cut they took didn’t cover their development costs, servers, marketing, etc. But those costs don’t increase linearly per customer and they also plateau as the software stack matures, so there’s a path to just raising prices and getting to profit, and each additional customer brought them closer to profit even while the subsidies existed.
OpenAI and Anthropic are paying something like 10x the amount for inference as they get from subscriptions, let alone free usage and training costs. So each new customer is taking them further from being profitable. And if they jacked up the prices 10x, so that the basic subscriptions were a few hundred and the pro ones a few thousand a month, they would still be in the position that Uber was when they were doing the subsidies for their infrastructure. I think it’s fairly obvious that they wouldn’t be acquiring customers very quickly at those prices.
I don’t think they mean uber the company, I think they meant uber as in over or very.
I think they mean the company since they’re the most famous example of subsidising the product to gain market share and their name is invoked all the time in discussions of AI economics
Isn’t most of the cost in the model training?
The cost of inference has passed the cost of training quite a while ago and it’s by far the majority of expenses
Thanks for the link, but disagree with your summary.
The Spend on inference has passed the Spend on training.
Training is the monopoly product and even the cheapest known model (deepseek) cost 10s of millions to train.
That’s got nothing to do with how expensive it is to serve models to customers which is what was being discussed.