When the government can flip the switch

At 5:21 p.m. Eastern on Friday, June 12, 2026, Anthropic got a letter. It came from Commerce Secretary Howard Lutnick. Within hours, two of the best AI models in the world went dark for everyone on Earth.

The order targeted Fable 5 and Mythos 5. It blocked foreign nationals from using them, even foreign nationals working inside the United States, including Anthropic's own employees. Anthropic has no way to check the citizenship of every user in real time. So it disabled both models for all customers to comply with the US directive.

They launched and became fully available to people. The trigger was reportedly only verbal evidence of a potential narrow jailbreak, with no formal written disclosure provided. I think that detail matters more than the shutdown itself.

Here is the new truth every builder has to face. Your model provider is now entangled with its government's national security politics. That is no longer a footnote. That is a line item in your architecture.

The Availability Tax

Builders price three things into their stack: latency, accuracy, and cost. There is now a fourth. Call it the Availability Tax.

The Availability Tax is the regulatory risk that your model goes dark with hours of notice, for reasons you cannot predict and cannot appeal. It is not a bug. It is a tripwire that lives outside your code. A single jailbreak claim against one vendor can take down your entire product.

The math is simple. Before June 12, your worst-case outage was vendor downtime, maybe a few hours, with an SLA behind it. Now your worst-case outage is a government order with no SLA, no timeline, and no published standard. Anthropic itself said it is "working to restore access as soon as possible," but nobody knows if that means days or weeks.

The lesson is not panic. The lesson is to treat single-provider dependency as a concentrated bet. When the state can flip the switch, concentration is the risk you are not pricing.

What the Export Control Logic Actually Reveals

Let me pull this apart the way I would any long-term structural shift. The surface story is a jailbreak. The deeper story is a change in category.

For decades, U.S. export controls targeted physical things. Chips. Supercomputers. Specific foreign end-users, mostly in China. The Anthropic order applied that same logic to a live cloud model serving everyone at once. That is the shift. A frontier model is now treated less like software and more like a controlled dual-use weapon.

Consider the contrast pair. A product gets recalled when it is defective. A weapon gets controlled when it is dangerous. The government did not say Fable 5 was broken. It said the capability was strategic. The capability was never the novelty. The category change was.

Now hold the contrarian view in your other hand, because honesty demands it. This may not be a stable regime at all. Anthropic was already in a public fight with the Department of War, which had reportedly pushed the company to drop safeguards on domestic surveillance and autonomous weapons. The shutdown could be leverage in that specific standoff, not a neutral policy for all builders.

The evidence on durability is mixed. The LA Times reportedly called it an "extraordinary directive" and noted nothing like it had happened before. First strikes get litigated. Courts may narrow how far "national security" stretches to justify a blanket shutdown. Congress may step in once the cost to U.S. competitiveness becomes obvious.

So here is the honest framing. Treat this as a real, quantified risk, not a permanent law of physics. The asymmetry is what should worry you. The cost of one shutdown is high. The cost of building a fallback is low. When one side of a bet is cheap and the other is catastrophic, you hedge.

There is a deeper principle here for any long-game thinker. Only resilience is real. The rest is convenience. A workflow that depends on one frontier model from one vendor in one country is a beautiful house built on someone else's land.

2031

Three signals inside the same shift

Pull the lens back five years. By 2031, "regulatory availability" will be a standard column in vendor comparison sheets, sitting right next to price per token.

I think the export control tripwire becomes a permanent feature, not because every shutdown sticks, but because the precedent now exists. Once a government proves it can do this, the option lives forever in the toolbox. Smart founders plan for the option being used, even if it rarely is.

This is why several analysts now argue every nation wants its own frontier model. Relying on a foreign cloud means your capability can vanish based on a politics you do not vote in. What follows is predictable. More sovereign models, more regional deployments, more fragmentation.

That fragmentation cuts both ways, and this is where I hedge hard. It is unclear whether a splintered model ecosystem makes safety better or worse. More diversity means more competition and less single-point risk. It also means harder global coordination and more places for bad actors to route around the rules.

The strategic read is this. The winners of the next five years will not be the builders chasing the single best model. They will be the builders who designed for impermanence. Capability is rented. Resilience is owned. That distinction will separate the companies that survive a 2026-style shutdown from the ones that go dark with their vendor.

What to Build This Weekend

You do not need a policy team to start hedging. You need an afternoon and a willingness to break things on purpose. Build a model abstraction layer this weekend.

A model abstraction layer is a thin piece of code that sits between your app and any AI model. Your app talks to the layer. The layer talks to the model. When one model dies, you swap the model behind the layer without touching your app.

First, pick two providers from different vendors. Route your main traffic to one. Keep the second wired up and tested, not just listed as a backup. A backup you never test is not a backup, it is a hope.

Then run a fire drill. Manually cut off your primary model and watch what breaks. Time how long your app takes to fail over. I promise things will break the first time. That is the point. You want to learn this on a Saturday, not at 5:21 on a Friday when the order lands.

Want to practice the abstraction muscle on something low stakes first? Tools like Showrunner, which generates voiced and animated episodes from text prompts, and StoriesForKids.ai, which helps families co-create illustrated stories, both sit on top of swappable models. Study how they treat the model as a replaceable part. If you want to scan for fallback options fast, There's An AI For That maps specific tasks to specific tools, so you can build a shortlist in minutes.

Start small. One layer, two providers, one fire drill. Get your reps in this weekend, and you will sleep better the next time a government letter goes out at 5:21 on a Friday.