I think you’re right - it’ll be interesting to see how the behemoths navigate when new and cheaper memory manufacturing capacity comes online. Currently predicted to happen starting 2028. That is where most of the current hardware costs are concentrated.

On that front, I think DeepSeek is a dark horse. They are really odd - not chasing revenue, care more about unit profitability, are obsessed with driving down costs (like Elon back in the day). As the founder says, they are content with where they are, being the model that does 80% for a fraction of the cost.

I mention them because they are playing a different game and might be able to out-innovate everyone else.

Deepseek has been a game changer for me. I don't use them right now but the model is very capable.

What a lot of people don't seem to realize is how much of the edge of the 'frontier' is in the tooling, the prompts and the harness. The LLMs themselves are already all reasonably close if you keep all the other stuff constant. Sure, one will have an edge in one domain or another. But that edge mostly translates to a degree of performance rather than some absolute. And if you allow for a few more toolcalls, a web search or a longer context then that difference is much smaller than it might seem at first glance.

The calculus may be that DeepSeek (or similar) + human is good enough and that the frontier simply won’t have the growth to get to the next level that differentiates them and justifies the token cost. They get stuck in a race to the bottom which is opposite of what they want.