The best objection to everything I’ve argued comes from my own data, and I want to meet it head on, because an argument that hides from its strongest counterexample isn’t worth your time.
Here it is. Broadcast television, in the mid-twentieth century, scored a 5 on capability. The same maximum I’m giving AI. Three networks, ABC, NBC, CBS, owned the studios, the cameras, the towers, the talent, basically the entire capacity to make and send the moving images that reached the public. By my own instrument, television concentrated capability just as hard as AI does. And television reopened. The cycle turned, like Wu said it would. So why am I treating AI as a structural break instead of just the next lap around the same track?
The answer is that two technologies can land on the same score for completely opposite reasons, and the reason is what tells you whether it will ever reopen. There are two kinds of concentration, and they behave nothing alike.
Television’s concentration was reach-driven. To put a signal into fifty million living rooms at once, in 1965, you needed enormous radio towers, government-licensed spectrum, and a fortune in national distribution. The networks had a monopoly on the reach. But the actual physical act of recording a moving image was never the thing they had locked up. Cameras were going to get smaller and cheaper no matter what, because nothing about the physics of a camera demanded a studio. And they did. Studio rigs became shoulder-mounted news cameras became camcorders became the thing in your pocket. The capability to create video kept leaking out the whole time. So when the internet finally erased the reach barrier, that capability was already sitting in millions of hands, just waiting for a road. You didn’t need a tower to reach fifty million people anymore. You needed a YouTube account. The cycle reopened because the car was already ours. Television only ever had a chokehold on distribution, not on creation.
AI’s concentration is capability-driven, and that one word changes everything.
There is no reach barrier with AI. A new model reaches a billion people the day it launches. The road is wide open. The thing that’s locked up is the capability itself. The actual physical capacity to train a frontier model sits behind a wall of scarce, ferociously expensive compute, and you cannot route around it the way the internet routed around the broadcast towers. You can’t invent a clever protocol that makes a hundred thousand specialized chips and a city’s worth of electricity unnecessary. The bottleneck isn’t the road that delivers the output. It’s the multi-billion-dollar factory that makes the intelligence in the first place. The capability is welded to the capital.
So: two identical scores, opposite mechanisms, opposite fates. Television’s 5 was a chokehold on the road, and the road can always be rerouted. AI’s 5 is a chokehold on the car, and there is no historical guarantee that ever comes back, because Wu’s cycle only ever reopened things by giving an already-existing capability a new channel to flow through. It has no move for a capability that was never in your hands to begin with.
If you want the mirror image that makes this unmistakable, look at the Arab Spring. That was actually the subject of the unfinished thesis I started back in 2016, and getting it wrong then is part of why I think I have it right now.
In 2011, citizens across the region used ordinary phones to organize, to film, to report, to coordinate the toppling of regimes. The capability sat squarely with the users. They had the phones, the voices, the footage. What they lacked was a safe road. Authoritarian governments owned the internet providers and could throw a master switch and black out an entire country. High capability, fragile channel.
AI is the exact inverse of Tahrir Square. The channel is flawless. The interface works perfectly on every device on earth. What’s missing is the capability, hoarded at the root. The protester could make the media and struggled to send it. The AI user can send anything instantly and cannot make the engine. Same era, same devices, opposite structure. Cheap access hid the difference both times, which is exactly why the access trap is so dangerous.
Now the part I owe you, because I’ve been promising it since Post 1. This claim can be wrong, and I can tell you precisely how. The Inversion dissolves, and AI rejoins Wu’s ordinary cycle, if a real substitute path to frontier capability appears. Three things would do it. One: algorithmic efficiency improving so fast and so durably that training a frontier model stops requiring a fortune in compute. Two: open-weight models reaching genuine parity with the closed frontier and holding it, not trailing a year behind. Three: distributed or federated training maturing to the point that people who don’t own a data center can pool their way to frontier scale.
If any of those lands and sticks, the car comes back into our hands, and I am happily, publicly wrong. I’m watching all three. So should you, because they are also the most important things anyone could be building right now. That’s not a hedge. That’s the watchlist, and it doubles as the to-do list.
Which leaves the only question that actually matters. If the thing won’t reopen on its own, what do we do before it locks?
Next: Nobody’s Coming. Why the fix has to be built now, by us, and what building it actually looks like.