The export control that proved my own thesis - and exposed the hole in it.

Last Friday the US government decided I wasn't allowed to use the most powerful AI model on the market.

Not me specifically.

But me, functionally.

I'm a foreign national in Toronto, and the Commerce Department issued an export control directive ordering Anthropic to cut off Fable 5 and Mythos 5 for any foreign national - inside the US or out, including Anthropic's own foreign employees.

To comply, Anthropic had to pull the models for everyone.

The most capable thing they've ever shipped, dark overnight, by government letter.

My first reaction was the obvious one.

I was pissed.

Then I opened OpenRouter, spent the afternoon hammering DeepSeek across four different providers, looked at the bill, and felt something I did not expect.

I felt fine.

Better than fine.

The whole day cost me about four dollars.

Let me walk through why that four dollars matters more than the export control does.

The model I lost

Fable 5 is the most expensive model Anthropic has ever sold.

Ten dollars per million input tokens, fifty per million output - exactly double Opus, which was already the premium tier.

It was free on the subscription plans for a two-week launch window that closes June 22.

After that, you pay.

So here's the thing nobody's saying out loud.

I was leaving anyway.

The government yanked Fable away from me about ten days before Anthropic was going to start charging me double-Opus rates for it.

The state didn't take something from me.

It accelerated a decision I'd already half-made.

I just didn't know that yet on Friday, because on Friday I was still thinking like someone with no alternatives.

What four dollars buys now

Here's what the afternoon actually looked like.

The same model - DeepSeek V3.1 - served by four different vendors.

Google Vertex. NovitaAI. DeepInfra. AtlasCloud.

Same weights, four price tags, four speeds, all of them charging me fractions of a cent per call.

I wasn't picking a model.

I was picking a vendor for a model that's become a commodity, the way you pick which gas station to fill up at.

And there's a whole column of these.

The workhorse I was running - V3.1 - is roughly sixty times cheaper than the Fable model I'd just lost.

Sixty.

If I want to climb back toward the frontier, V4 Pro sits about twelve times cheaper than Fable, and for the work I actually do, I genuinely cannot tell you it's worse.

The gap between 'most expensive model on earth' and 'good enough that I can't feel the difference' is now a factor of twelve to sixty, and falling.

So the choice the export control forced on me wasn't Fable versus settling.

It was Fable versus two tiers of cheaper, one of which might be just as good.

That's why I stopped being angry.

The confession

I want to be honest about the actual mechanism here, because it's not flattering.

I decided the export control was deserved.

After a day of stewing, I came around to thinking the US is within its rights, the cyber concerns are probably real, fine, I get it.

But I only think that because I have alternatives.

Sit with that, because I had to.

My judgment of the policy didn't move on principle.

It moved on optionality.

The control had power over me for exactly as long as I had nowhere else to go - and the moment OpenRouter handed me four vendors and a sixty-x discount, I could afford to be gracious about being punished.

My ethics were downstream of my menu.

If DeepSeek and Qwen didn't exist, I'd still be furious.

I'd be writing a very different newsletter about American overreach and the weaponization of the AI stack.

The substitutes didn't just save me money.

They changed what I believe.

That should bother you a little.

It bothers me.

Where this breaks my own thesis

I've spent two years saying Own or Be Owned.

Control your stack, or someone else controls you.

So let me hold my own feet to the fire.

I don't own any of this.

DeepSeek V3.1 through OpenRouter is renting, through a broker, a model I didn't train, running on infrastructure I don't touch.

By my own framework that's maximum platform dependency.

It's the exact thing I warn founders about.

And yet it's what saved me.

Not sovereignty.

Not ownership.

Substitutability.

The thing that neutralized a US government export control on a Friday afternoon wasn't me owning compute in Rimouski.

It was a commodity spot market with four interchangeable vendors and a race to the bottom on price.

That's either a hole in my thesis or its next version, and I'm still deciding which.

The honest update might be this: ownership is one form of leverage, but optionality is the cheaper one, and most weeks it's enough.

You don't have to own the model.

You have to make sure no single provider - or government - can be the only door.

The part the government should worry about

The directive was meant to deny a foreign national access to frontier American AI.

On that narrow goal, it worked.

It pushed me off Anthropic inside an hour.

And it pushed me directly onto Qwen and DeepSeek.

Chinese open weights.

At one-sixtieth the price.

That's the whole case against export controls, and I lived it in one terminal session.

The control didn't deny me intelligence.

It transferred my dependency from an American lab to a Chinese one, made me cheaper to operate, and gave me a reason to learn the migration path I'd been too lazy to learn before.

I've now got Claude Code alternatives open in other tabs for the first time.

Friday's directive is the reason.

Multiply me by every foreign founder, every overseas team, every non-American developer who got the same letter's downstream effect.

The policy isn't building a wall around American AI.

It's a customer-acquisition funnel for Beijing's labs.

The call

So here's the prediction I'll put a date on, the kind I can be wrong about in public.

By the end of Q3, an export control intended to protect American frontier AI will be cited - by someone with real numbers - as having measurably grown adoption of Chinese open-weight models among exactly the users it was meant to wall off.

The mechanism that's supposed to be a moat is a migration tool.

I'll come back and tell you if I got it wrong.

Close

The cheapest thing in my business this year is intelligence.

I proved it last week, by accident, for four dollars, after a government tried to take it away.

The expensive thing - the thing I'm still paying for - was believing for a single afternoon that I didn't have a choice.

Own or be owned.

But if you can't own it yet, at least make sure you're never down to one door.

Founder Reality, in long form. No fluff, no recycled takes.