(Don’t forget Betteridge’s law!)
Written 2026–10–24.
Please take this all with a spoonful of salt — I am not trained in financio-economico-anything.
I’ve seen on some internet forums this idea that the price of LLMs tokens is gonna go up sharply once the money runs out. I’m starting here with the assumption that this is in a non-catastrophic scenario. That is, if we hit the singularity, or if the global economy crashes in a way that the current definition of “more expensive” becomes meaningless, this is all irrelevant 1.
Anyway, assuming non-catastrophicity — and accompanying that, a petering out of LLM progress — these very large “AI” companies suddenly no longer expecting a return on absurd training costs means that they would switch to focusing on making money from inference.
This is the point where the aforementioned internet commenters assume that the requirement for profit (from investors) requires the cost of these services to rise sharply. I only see that being true in the case where competitors are entirely barred from the market, allowing the big companies to form a cartel 2.
Without that, these companies having very little moat (or distinction) surely means they quickly race to the bottom on margins in order to even make a sale.
1 I almost wrote “moot” there, but the opposing duality of its meanings makes me avoid it.
2 Well, I probably shouldn’t underestimate the chances of that happening.