Point the Compute at the Machine That Decides Who Gets Care

The same day some of the smartest people alive were arguing over who deserved credit for a fluid dynamics breakthrough, a friend who works in AI research asked me what they should go after next. I didn’t have to think long. “I got a better one,” I said. “Model the full US healthcare system.”

I want to be fair about this. Millennium problems are real problems, and the people chasing them aren’t wasting their lives. But if you’re sitting on the largest concentration of compute in human history, I’d point it at something with a shorter fuse.

Model the American healthcare system. All of it: provider networks, utilization management, revenue cycle, Medicare, the private payers, and the PBMs sitting in the middle taking a cut nobody can find on a statement. Build it as one system and run it forward.

Then answer one question. Which part is load-bearing?

(Yes, I hear it. I’ve spent so much time talking to AI that I now say “load-bearing” with a straight face.)

We do this kind of thing everywhere else. We model hurricanes days out and evacuate a coastline on the output. We model the grid well enough to know which handful of substations could take a whole region down with them. We rate bridges for a load before the trucks ever roll across. After 2008 we started stress-testing banks every year on purpose, because finding the weak wall after it fails is the expensive way to learn where it was. Healthcare is about $5.3 trillion a year and 18% of GDP. There are good models of who’s covered and what it all costs, but as far as I can tell nobody has built one of where it breaks.

It isn’t because it’s impossible. I think it’s because nobody holding the compute has wanted the answer, and everybody holding the data gets paid by the spread that answer would expose.

I didn’t get the ask right the first time. My first draft ended with “tell us when the catastrophic failure will happen.” I caught that one myself, and even as I cut it, it still felt weird to say not when it breaks. A date is what everybody wants.

But dates are guesswork, and a year mostly gives people a number to argue with. So I rewrote it. Skip the when. Tell me what breaks first, and what it takes down with it. Is it rural hospitals? Is it the float in a revenue cycle that’s been running on borrowed time? Somebody find it and say it out loud, so we can spend our time arguing about the fix instead of the diagnosis.

We can simulate a galaxy, and we can fold a protein, but we can’t say which part of American healthcare fails first, and we’re all standing inside it.

If I had one year of frontier compute and one system to point it at, I’d skip the proof and go for the answer.