Ben Thompsons's Who's Afraid of the Chinese Models is the best piece of analysis on token economics I've come across.

The tech industry just assumes zero marginal costs.

With tokens, the marginal cost of inference is non-trivial.

The frontier labs have presumed that they'd do significant capex to generate proprietary differentiated models that they could then make a lot of margin on inference.

But it turns out that the models aren't actually that differentiated, so despite the capex it doesn't turn into sustainable margin.

Instead, the margin comes primarily from the garden-variety economies of scale of doing inference at scale.

A real sustainable competitive advantage, but perhaps 100x less powerful than a "we have the differentiated models everyone needs" that their trillion dollar valuations presume.