What a weird thing to say: a business is profitable if we ignore a key part of its cost base. That’s true of any business, you know. Even if they make a 100% coverage of every request served, it’s still a far cry from the usual internet business model where responding to a request is almost free and most of the expense is the initial investment in software and its upkeep. This is why Alphabet used to be flush with cash and we get free services.

The trouble with Anthropic and OpenAI is, that not only do they need to get operationally profitable, they also need to be sufficiently profitable to provide a return on capital on all the money they have burned to date. Much of which has gone (and keeps going) not to assets but to expenses like the electricity bill. Thus, once the models are efficient enough to be served profitably, a relatively small amount of assets needs to be generating income for a relatively large amount of capital. It is my understanding, that currently the efficiency of the models does not grow even fast enough to outpace the need for new capital.

One more thing, that gets conveniently ignored when talking about AI cost is the cost of training the models. From one hand, there’s but a few institutions capable of building them and, from the other, it’s like the models grow on trees. Yes, you could run your open-weight model on your own hardware and pretend it’s profitable because you don’t have a comprehensive costing model and the infrastructure costs from security to electricity to risk are mixed in with your general cost base. But someone has spent astronomical amounts of money training that model you run and would probably like to get it back with some interest. Let’s hope that return is just money and not, say, the ability of a certain country to control the global software development process.