I wrote up something I’ve been chewing on for a bit.

AI feels very competitive right now with new models, new benchmarks, new products every week. But I’m less sure it stays that way once you look below the model layer.

Training and serving serious AI systems still depends on a lot of scarce infrastructure: GPUs, data centers, cloud compute, power, networking, and the talent that's deeply "in the know".

The awkward bit is that the companies with the best access to that infrastructure are often the same companies building the models everyone else is trying to compete with.

I don’t think the answer is “government should run AI.” That sounds like a great way to make everything worse.

But I do think there’s a real question around infrastructure access. If compute becomes the bottleneck, how do we make sure smaller companies, researchers, and new entrants still have a shot?

Open models help. A lot. But even an open model needs to run somewhere.

This is Part 1, so it’s mostly me trying to frame the problem before getting into possible solutions. I’d genuinely like pushback, especially from people who think I’m overstating the monopoly / infrastructure angle.

What am I missing?