Ultimately it's marginal, meaning that every $ of revenue that Google and Meta make have an associated compute cost with it, but scale your revenue down and you scale compute down also. Moreover this compute cost per revenue is fairly low. A small cloud instance equivalent's of compute can generate a ton of business value.

The big foundational labs are not like this. Every model they release takes a ton of data infrastructure and training (pre, train, post) compute. Then inference itself requires compute. Even if the labs sell tokens, each token embodies a lot of training and inference compute. This is a much more capital heavy business, so I can see revenue multiples being smaller than pure SaaS businesses. A model release bust (like Opus 5) ends up leading to big losses because of the necessary training compute spent on the model and the labor spent on training. A bust SaaS product ends up being wound down quickly, and only costs the labor spent on building it and the compute spent running it.

A good analogy is chip companies before all of the AI related compute pressure. These were largely low margin businesses because of the fixed capital cost associated with each new chip gen, architecture, process, etc