Strongly dispute this. Compute very depreciates rapidly. Inference is cheaper than training. DeepSeek was the warning shot across their bow, but the big AI firms can't afford to change course without jeopardizing their "Wile E. Coyote off the cliff" economics.

LLM performance is already plateauing; models will get more efficient. Good-enough models will be deployed on chips, the same way H.264 is a good-enough video codec but used ubiquitously.

More than your points, I'm very curious how these AI companies are going to turn profit without making using AI insanely expensive. Some time ago, each prompt was highly subsidized, I doubt the picture has changed much.

Edit: maybe the model efficiency you mentioned is the key, we'll see.

I suspect they just won't. First-mover disadvantage is real for many markets. Everyone knows Amazon, but how many remember Kozmo.com?

My assumption is that OpenAI, Anthropic, etc will go bankrupt and eventually be subsumed into Microsoft/Google/ByteDance & friends. New entrants will take their pioneering work and sell inference for pennies on the dollars without investing in massive R&D spend.

Supposedly they could make money, if they wouldn't have to burn a lot on the research. There was an interview with Dario where he stated this and hinted to the fact that a monopoly would not have the research problem, and thus could start making money.

> Whoever has compute, will have the power.

Nonsense. There's a temporary shortage, but even with the shortage it's still a commodity.

No. There will never be an oversupply again. Tokens = work in the same way that humans = labor.

> No. There will never be an oversupply again. Tokens = work in the same way that humans = labor.

So? Electricity = work as well, and yet it's a commodity.