And, yet, the US AI companies are not actually making a profit, right? They're selling at a loss and trying to make it up in volume (or lock in some kind of monopoly position, in a currently non-sticky product). We're all currently enjoying investor-subsidized tokens from the big guys, and that pushes out the reckoning for US AI. But, I think they're beginning to think maybe they need to ring the cash register. Copilot dramatically reducing usage limits and what models are available on its plans, Anthropic playing games with what's included in the Pro plan, etc. I think they're starting to feel the bleeding.

Not only is the investment that keeps US AI companies flying high slowing, I suspect in two or three years, we'll all mostly be using open models and the people making money will be the hardware manufacturers. Even the small models will keep getting more capable. I'd guess a model you can run on a high end, but not outrageously overbuilt, developer desktop or laptop (something like 128GB of unified RAM), will be competitive with the current frontier when it's allowed to search the web and do research and write test code. You can't fit as much knowledge in a small model (80GB of weights can't store the world's knowledge), but I don't have the world's knowledge in my head, either, and yet I can figure out most problems with a little googling and experimentation. The reasoning and tool use abilities of smaller models is where the gap is closing, and that's what will make the huge models obsolete for huge classes of problem.

Already, there are many classes of problem that the easily self-hostable Qwen 3.6 27B can solve that required a frontier model a year ago. When the self-hosted options reach Opus 4.5-ish levels of capability, the argument for paying for tokens for most work begins to look a lot less compelling. And, looking forward, 1.58 bit models are coming. Incredible intelligence density, and still a lot of improvements happening.