The Age of Private AI

The best AI is going private, behind national walls and company walls. The rest of us buy whatever comes through the slot.

On a Friday, Anthropic was given about ninety minutes to comply. Washington had decided that Fable, the company’s most capable model and one that is REALLY cool, could no longer be touched by foreigners. Not foreigners abroad. Any foreigner, including the people who work at Anthropic and happen to hold the wrong passport. Researchers at Amazon had found a way around Fable’s safety limits to reach its hacking abilities, and instead of patching the hole for everyone, the government shut the door on everyone who isn’t American. Anthropic took the model offline while it figured things out. At least that was the official story.

The same thing happened a few days later with GPT 5.6, which launched only for a short list of “trusted partners.”

Officially, this is about safety. The reasons nobody says out loud are simpler. One is to delay China and everyone else from copying a top model to train their own, which the industry calls distillation. The other is to keep America’s lead while it still has one. Take away the safety language and what is left is a wall, with the best models on the inside and the rest of the world on the outside.

That ban has already been lifted, which, as I will get to at the end, proves the point in this blogpost more than it undoes it.

We’ve been here before

Not the first time. In the 1990s the US government tried to ban strong encryption. Encryption is the math that scrambles a message so only the right person can read it. It is what keeps your bank login and your WhatsApp chats private.

Back then Washington treated that math as a weapon, the same legal category as a missile. When a programmer named Phil Zimmermann released PGP, free software that let anyone encrypt their email, the government opened a criminal case against him for “exporting munitions.” His supporters printed the code on T-shirts and inside a book, because a book is protected speech and you cannot un-publish math. By 2000 the government gave up, and losing turned out to be the best thing that could have happened. Once encryption was free and everywhere, the whole internet got safer, online shopping and banking became possible, and ordinary people got a kind of privacy that used to be a government privilege. The thing Washington tried to bottle up became the plumbing the digital economy now runs on, and the country that fought hardest to keep it locked gained the most when it spread.

The case for trying again with AI is not stupid, even if today it sounds a bit like a marketing stunt by Mr. Wolf-Is-Coming. A model that is very good at writing cyberattacks or designing a virus is closer to a weapon than to a word processor.

AI is a little different from encryption, though. The software still leaks (as Mark Zuckerberg keeps pointing out), a top model fits on a thumb drive, but training one from scratch takes rare and expensive hardware. So a government can make it slower and more costly for everyone outside, and pick who falls behind. That buys time and chooses winners. It does not stop the technology.

How much that matters depends on one thing I will come back to, whether AI keeps getting cheap or the very best pulls away into its own league. In industry jargon, whether AGI is coming.

Who is impacted

Start with the obvious loser. AI labs make money selling their models to the world. Tell them that shipping the model triggers a government lockdown, and for sure they will ship less, or won’t ship at all. The logic of the ban is shaky too. If the model were truly that dangerous, you would block Americans from it as well, because an American can do damage just as easily. Blocking only foreigners is not really safety. It is the chip-ban playbook, the same move Washington already ran on advanced chips to China, aimed now at software, and like those chip bans, it mostly teaches your rival to get good without you.

Then there is everyone else. Normally a country that cannot build the best models just uses them, the way you use electricity without owning a power plant. That works only while the supplier is reliable. The moment the US shows it will cut you off on a Friday afternoon with ninety minutes’ notice, every other country has to assume it could be next, and starts building its own worse version, because you cannot run your economy on a supplier who might pull the plug without warning. The problem is that with continuously improving capabilities, not having the latest model, even if only temporarily, could mean the difference between unblocking an insane amount of value or getting stuck. Let’s go back six months into the past. In November 2025 we were on the brink of software engineers becoming 10x, 100x more productive, just a simple model update away. If that model update were only available to some companies, we would not have earned all those gains in the last six months.

Europe is one of the most impacted. We are trying to innovate (Luzia, ahem), with no top model of our own, not even one competitive with Chinese open source as of this writing. We are the most exposed to a cutoff, and the most likely to find that “just use the American one” has quietly turned into “build your own or pick a new patron.” And a patron is waiting. If Washington will not sell and Beijing will, the world drifts toward Chinese models. This is Made in China pointed at AI, something I flagged a while back. The stated goal was to contain China, and the method ends up handing them the floor, the mass market of everyday users. More of that usage means more acceleration, which makes containment harder still. I am not saying that is a bad outcome, just an ironic one.

Do these bans really matter?

The wall only works if what is locked inside is clearly better than what the rest of the world can get cheaply. Whether that is true is the real question.

Fable is really good. Not just good, really good, a step-function change. But how useful the ban is does not come down to the model’s capabilities in absolute terms. It comes down to relative position: how far ahead Fable is of the next best option, and whether, for your particular use case, the model is good enough.

And today, the bar is lower than it sounds. Fable does not have to be the single best model on earth. The ban does its job as long as it widens the gap between the locked top and the open models everyone else builds on, today or six months from now. A wall around a slightly better model still leaves the outside world a step behind, and a step behind has a way of growing.

The best AI goes from rare to ordinary the same way every technology does (differentiation then commoditization). A new thing starts out differentiated, owned by whoever built it first, and then it commoditizes: it gets cheap, everyone has it, and the advantage slides from the people who make it to the people sitting closest to the customer. AI is running that loop faster than anything before it.

AI is not one thing improving at one speed. For the everyday stuff most of us use it for, it keeps getting cheaper and good enough, and the free open models are only months behind the best paid ones. For a few high-stakes things, like hacking, designing viruses, or building better AI itself, the very top might pull away and stay ahead. So the likely future is cheap, plentiful AI for almost everything, and a small locked vault for the handful of things that decide real power. The wall matters as much as that vault does, and no more.

Two things stand between that vault and a permanent American monopoly. The first is China (and other competitors, but let’s be honest, today it’s really only China), which the ban only strengthens: lock the world out of American models and you push it toward Chinese ones. The second is that AI might simply stop getting much better. If progress hits a wall, the locked models stop pulling ahead, the open ones keep closing the distance, and the lead behind the wall quietly stops being worth guarding (and today there are no signs of this happening at all).

The second wall

There is a second wall going up, and this one needs no government. A company with a model good enough to do real work has a choice. It can rent you the model, or it can use the model itself and sell you the result. As the models get better, especially if we start putting government restrictions around them, the second option wins more often. A firm that could rent you an AI lawyer will likely make more money using that AI to sell you the legal work directly. Companies exist because some things are cheaper to do in-house than to buy. AI makes a huge amount of work cheap to do in-house. So companies pull more inside, the best AI stays behind their doors, and what reaches you is the finished service, never the engine that made it.

I made the hopeful version of this argument in an earlier piece, “AI Is the Org”: inside a company, AI pushes people up a level, the way the spreadsheet pushed accountants up instead of out. This is the dark side of the same idea. Inside the company, a person moves up. Outside it, the rest of us move down, into the seat of a customer who only ever sees what comes out of the box. You keep getting cheaper magic, and you slowly lose the ability to make the magic yourself. Preventing models from reaching the market makes this scenario even more likely.

Is it worth it?

So the wall is a bet. For it to pay off, two things both have to hold: the lead has to last, and the game must not change underneath it. History has been rough on both. The strongest defense is that a year or two of lead buys time to make this technology safe before a reckless race makes it dangerous, and that is worth taking seriously. But it needs the lead to be both controllable and long enough to matter, which is exactly the bet the encryption wars lost, and every jump to a new kind of computing lost before them.

So America pays now, with a slower industry at home and allies drifting toward Beijing, for a lead that only pays back in the one future where it was going to win anyway. The specific ban may have gone, but unless something big changes, this is the new direction. The best AI is going private, behind national walls and company walls, and the rest of us are ending up on the outside, buying whatever comes through the slot.

For somewhere like Europe, this becomes a brutal little bet. If the wall ends up not mattering, building your own models was wasted money. If it does matter, being stuck outside the garden means being cut off from the technology that sets the next few decades, and that loss has no floor. When one side of the bet is some wasted budget and the other is permanent irrelevance, you build. One of those Pascal’s wagers I love. The asymmetry, not the odds, is the argument.

You can lock up today’s AI. You cannot lock up the one that comes next.

Postscript

I had this written and scheduled. The morning it was due to go out, the ban lifted (and then life got complicated and took me 48 hours to re-adjust)

Fable is back, worldwide, and the reason it came back is the argument I tried to make here. The capability that set off the panic was never unique. Anthropic’s own testing showed that Opus, GPT-5.5, and Kimi, a Chinese open model, could all find the same flaw. The wall had gone up around something that was already on the other side of it.

So this round was a false alarm, but a telling one of what’s to come. The models do not stop getting better (rumour is that Fable 6 is already available internally), and the next one will not be a false alarm. Fable was the drill. While everyone argued about the drill, the machine got built: an executive order, government sign-off before a model ships, an industry system for scoring how dangerous a jailbreak is. None of that machinery cares why it was built. A switch installed for safety works just as well for trade, and a switch that exists gets flipped. The ban was temporary. The machine is not, and it is waiting for the model that actually earns it.