> what breaks the spell?

If throwing more compute at the problem keeps only resulting in incremental gains, I think that should do it. It goes one of 2 ways, really. Either we can throw enough compute at pre-training that results in infinitely more capable models to the point that the cost is now justified [1], or, we hit a scaling wall, get stuck with what we have now (or at that time) and the valuations crash knowing that "this is it" for the foreseeable future without a big breakthrough.

The labs go bankrupt or get acquired by the typical giants (Google, Microsoft, Amazon), the models get rolled into GCP, Azure, and AWS as a service, and that's it. It becomes another dev tool, much like a new IDE.

[1] cost being justified I'd rank as "your average non technical PM can now end to end develop robust, production software free of most serious vulnerabilities." model & tool capabilities that would allow you to hire a small team of non-techincal roles, for half the salary, that can produce the output of a large engineering org. If that doesn't happen, I don't see how the current buildout is sustainable.