Here is my conspiracy theory about why the strongest AI models suddenly need a government's blessing.
Last week OpenAI's best model launched to about twenty customers approved by name by the US government, and Anthropic's new flagship lived three days before a directive switched it off. The official story is cybersecurity. Here is my conspiracy theory about the second engine under the hood.
Here is my conspiracy theory, and I want to be honest from the first line that it is a theory. I have no inside source and no leaked memo. I just have a nose for the moments when the official story and the profitable outcome point in suspiciously the same direction, and the last few weeks have been full of them.
The facts first, because the facts are strange enough on their own. Last week OpenAI launched its strongest model yet, Sol, to roughly twenty customers, each name individually approved by the US government, with access granted, in Sam Altman's own words to staff, "customer by customer." A couple of weeks earlier Anthropic launched a new flagship to enormous fanfare, and three days later a Commerce Department directive forced it to switch the thing off for every single customer. The headline reason for all of it is cybersecurity. I think there is a second engine running under that hood.
What actually happened
The chain is short and worth laying out. On 2 June an executive order asked AI companies to "voluntarily" hand new high-capability models to the government up to thirty days before anyone else gets them. The order goes out of its way to say that nothing in it creates a mandatory licensing or preclearance regime. Then it grew teeth anyway. Anthropic's Fable and Mythos launched on 9 June and were dark by 12 June, because a directive said no foreign national could touch them and the only way to comply that fast was to pull the plug on everybody. A former White House AI adviser, now heading to OpenAI, called the result a "de facto involuntary licensing regime," which is a careful way of saying the voluntary part is a formality.
OpenAI, to its credit, said out loud that it does not think a government access list should be the long-term default. It also cooperated fully, hired the adviser, and shipped Sol to twenty vetted partners anyway. That is what managing a relationship from the inside looks like.
The part the security story leaves out
Now the theory. There is a problem that nobody running a flat-rate AI subscription wants to say into a microphone: the strongest models lose money on the heaviest users. A flat monthly fee was always a land-grab bet that the average customer stays cheap and subsidises the few who do not. The power users living on the flagship, running agents all night, are the ones who blow a hole in that math. The better and more expensive the model, the bigger the hole.
If you are the company, you have two honest ways out and both are ugly. You can raise the price, which means admitting the thing was underpriced and watching the internet set itself on fire. Or you can quietly throttle the best model, which means admitting the product you advertised as unlimited is rationed. Both are bad mornings.
Then a third door opens. The government, citing national security, asks you to limit your most powerful model to a small vetted list. And suddenly limiting the flagship to a handful of approved customers stops looking like a pricing problem and starts looking like national responsibility. The security explanation and the margin-protecting outcome want the exact same thing: far fewer people on the most expensive product you make. When the noble story and the profitable one line up that cleanly, I do not assume a conspiracy. I just stop assuming a coincidence.
There is a less cinematic version of the same story, and it points in the same direction. Google has reportedly started limiting how much of its Gemini models Meta can use, because it could not hand over as much compute as Meta wanted, and clients have been told to spend their tokens more carefully. That is Google rationing its best model to one of the richest companies on earth. Around the same stretch of weeks it agreed to pay SpaceX roughly $920 million a month for raw computing power. When the supply is that tight at the very top, a flagship only twenty approved customers can touch starts to look less like a pure security decision and more like the honest amount of compute there was to go around. Governance is a respectable word for rationing something you could not have supplied to everyone anyway.
The clearest tell, for me, is the shape of the Anthropic episode. Three days. Long enough for the launch, the benchmarks, the breathless threads. Short enough that almost nobody had time to build a workflow that actually depended on it before it vanished. Whatever the legal reason, the shape of the thing is a sneak peek, not a product. You were shown the good model, you were allowed to want it, and then it was put back in the case.
What I would actually do about it
I am not telling you to panic, and I genuinely might be wrong. Maybe it is cyber threats all the way down and the economics are a happy accident. But the trend line does not care whether I have the motive right. The direction is the same either way: the best models drift toward a vetted few, with the rest of us renting a watered-down tier, or waiting for a preview window that opens for three days and closes.
So the practical advice is boring and a little urgent. Use the flagship models now, while "anyone with a credit card" is still the access policy. Build the thing you keep meaning to build. Run the experiment you have been saving. The window where the strongest tool is simply available may turn out to have been the unusual period, not the permanent one.
And then there is the version that no directive can switch off and no pricing memo can ration. The model on hardware you own. It is not going to win any benchmarks, and it will not be Sol. What it will do is keep working on the morning a directive somewhere turns the good one dark, on a box where nobody approves your access customer by customer. I have watched enough all-you-can-eat deals quietly become all-you-can-eat-for-approved-eaters to keep cooking in my own kitchen, even when the restaurant down the road is clearly better.
That is the whole theory. The official reason is security, and it may even be the real one. I just notice that the people who would have had to explain why the best model costs more than you pay never had to give that speech. Convenient. That is all I am saying. Convenient.
Buy Me a Coffee
Fuel my creativity with a coffee — every sip keeps this blog running!
Support This Blog — Because Heroes Deserve Recognition!
Whether it's a one-time tip or a subscription, your support keeps this blog alive and kicking. Thank you for being awesome!
Tip OnceYou read this far.
I write up the moments where the official story and the profitable one line up a little too neatly. One post a week, by hand. Subscribe.
SubscribeDOGE: DSYxsbfWKAX8wWED9aWeqLEVXU7KihKk6h
Canary: pro-it.rocks-canary-3683c6e6
