Companies doing foundational models need to cover the cost of training which is much more expensive than training something like kimi.

Yes. I would not consider Kimi a particularly good model relative to its size, and making a SotA model is a lot more expensive. But training costs are explicitly excluded when talking about the cost to serve tokens

>Companies doing foundational models need to cover the cost of training [...]

But that's moving the goalposts? The original claim was on inference itself, not the whole company.

> The cost to serve tokens is absolutely profitable today and that’s been true for at least a year.