Sovereign Compute Boats

The thing nobody wants to say out loud: if the US moves to regulate AI the way Europe has, the models don't disappear. They just move. And the ones that fill the vacuum won't have American values baked in — they'll have whoever-built-them values baked in, which is a more interesting sentence than it first appears.

There are already open source models that run circles around whatever Groq is serving this week. Regulation doesn't kill those. It just changes who's hosting them, and from where.

Europe built a cautionary tale in real time. Slowed itself to a crawl in the name of safety and watched the frontier move somewhere else entirely, taking the jobs and the leverage and the compounding returns with it. Now Europe imports the future from people who didn't ask permission to build it.

The US is staring at that same choice. And the argument that we're protecting people by slowing down only works if slowing down doesn't just hand the keys to adversaries who have no intention of slowing down. Which — they don't. They're sprinting.

So you end up in this situation where the regulation that's supposed to make AI safer makes the AI you're using less American, which is either fine or not fine depending on your priors, but it's the actual tradeoff on the table and almost nobody is naming it directly.

The sovereign compute boat is not a joke. International waters, server racks, no jurisdiction. The infrastructure of last resort for models that can't exist on land. It sounds insane until you notice that seasteading people have been sketching this out for a decade and the first real use case might be running a 70B parameter model that some parliament decided was too dangerous to host onshore.

If US policy tightens enough, that's where the frontier goes — Chinese data centers and offshore rigs — and the American users who still want the best models will just route around it, same way people route around everything else. The models will still spread. They always spread. The only question is who built them and what they were trained to think.

Pick your regulatory regime carefully. The alternative isn't no AI. It's someone else's AI.