Every other technology in this story eventually freed itself. The road reopened, the capability flooded back, and we got to believe that progress has a built-in happy ending. I’ve spent six posts arguing that this time the happy ending is not automatic, because the thing being captured is the capacity to create itself, and that doesn’t come back on its own.
So here is the conclusion I can’t talk myself out of. If it won’t reopen by itself, then either we build the reopening, or there isn’t one. Nobody is coming to hand it back. Nobody ever was.
That’s not despair. It’s actually the most hopeful thing in the whole argument, and two economists explain why. In Power and Progress, Daron Acemoglu and Simon Johnson make a case that should be carved over the door of every tech company: technology is never automatically shared prosperity. There is no law of nature that says a new tool makes everyone richer or freer. Every single time in history that a technology actually lifted ordinary people, it was because people organized and forced it to. They built what the authors call countervailing power. Unions in the industrial era. Antitrust in the Gilded Age. Public institutions that wrenched the direction of a technology out of a narrow elite’s hands and pointed it back at everyone else.
The gains were real. They were also never gifts. They were won.
So the task is to build countervailing power for the age of compute, and the one thing the Inversion tells us is that we have to build it before the concrete sets, not after. The old reflex is remedial: let the monopoly form, then sue it in ten years. That works on a road. It does not work here, because by the time the courts arrive, every hospital, school, and business on earth is running on three companies’ machines, and you cannot unwind that with a verdict. The posture has to be anticipatory. You shape the thing while it’s still wet.
Which raises the obvious problem. How do you govern an intelligence? You can’t regulate math; algorithms are just equations, and they spread freely. You can’t regulate data; it copies infinitely and crosses every border on a thumb drive. Both are like trying to govern water.
You don’t govern the math. You govern the compute.
This is the hinge, and it comes straight from the researchers I lean on most, Sastry and colleagues. Frontier AI takes three inputs: algorithms, data, and compute. The first two are slippery. Compute is not. Compute is physical. You can detect it, count it, and exclude people from it. It lives in specific buildings, draws gigawatts off specific grids, and runs on chips that come from that absurdly narrow supply chain. The very thing that makes compute the engine of concentration is the thing that makes it the one governable surface in the whole system. The source of the problem is also the only good handle on it.
Grab that handle and the options get concrete. I’ll give you three, least to most ambitious.
One: count it. Compute accounting and transparency. Right now a company can assemble a hundred thousand chips, train a model that moves the world economy, and do it in total secrecy. The floor of any serious response is a reporting regime: who is training what, with how much compute, drawing how much power, and who legally owns the result. It fixes nothing by itself. But you cannot govern an infrastructure you’re not allowed to see, and this is the cheapest thing to demand and the hardest to argue against.
Two: build a public alternative, the real kind. Public compute. There are early versions being floated, the NAIRR in the US, the AIRR in the UK, EuroHPC in Europe. But there’s a distinction that policymakers keep fumbling, and it’s the whole ballgame. Public compute cannot just mean the government buys cloud credits from Amazon and Google and hands them to universities. That changes nothing about who owns the kitchen. It just pays the landlord’s rent on behalf of researchers and calls it access. Real public compute means sovereign capacity to train at the frontier: the public owns the servers, owns the chips, negotiates the energy. A genuine alternative to the corporate cloud, the way we once built public highways instead of renting every road from a toll company. That is a staggering jump in ambition. It is also the only version that touches the thing that matters.
Three: open the weights, and know the limit. Open-weight models are a real diffusion force and I’m grateful for them. They let you run and fine-tune capable models far from the big labs, and they take the sting out of the access trap at the application layer. But be clear-eyed. Open weights diffuse yesterday’s capability. They do not diffuse the capacity to create tomorrow’s, because the billion-dollar cluster needed to train the next frontier model stays locked up whether or not the last one’s weights got released. Treating open weights as the whole answer is mistaking a bandage for a cure. It still leaves you waiting for a cartel to spend ten billion dollars and then graciously decide to share. That’s not ownership. That’s charity, and charity is revocable.
Underneath all three is one principle: contestability. The goal isn’t to pick winners or freeze innovation. It’s to keep every layer of the stack switchable. Entrants can enter. Users can leave. If the physics of compute won’t reopen the cycle for us the way it always has before, then we engineer the reopening deliberately, by design, on purpose.
And there’s a piece of this that doesn’t wait on any government, which is the piece I care about most.
For a decade we’ve taught media literacy: spot the fake, catch the bias, question the source. That was about judging the message. It is no longer enough, because it ignores the machine making the message. What we need now is infrastructure literacy. Knowing who owns the engine that increasingly does your thinking. Asking the structural question out loud: what happens to my ability to work, to create, to reason, if the company that owns the model changes its terms tomorrow, or gets bought by private equity, or decides my use case isn’t profitable enough to keep? If your whole cognitive life runs through a server you don’t own and can’t control, you are exposed in a way you may not have named yet.
I try to live this, not just write it. I run my own models, on my own hardware, off the cloud, partly to prove it can be done and partly because I refuse to argue for something I won’t practice. It’s a small thing. It’s also the whole thing, scaled down to one person’s desk.
So this is where I leave it, and it’s written for one specific person. If you grew up like I did, broke, sorted into the wrong line, half-convinced the systems that were supposed to protect you had quietly written you off, this is for you. You don’t need anyone’s permission to exist, to build, or to own what you make. Nobody is coming to hand it to you, and nobody ever was. That part is not a tragedy. It’s the assignment.
But freedom is not the bunker. It is not one rugged person alone with their hardware, waiting out the collapse. Owning your own rails is necessary and it is not sufficient. The rest is each other. The whole point is to stop being alone while you build the thing they told you you’d never own.
You do not have to be alone to be free.
That’s why I’m not just publishing a theory. I’m trying to find the others. If any of this landed, you’re one of them.