Here's a timeline worth internalizing:
June 9: a major US AI lab launches its most powerful model. June 12: the US Bureau of Industry and Security issues a directive. June 12, same day: every user in the world loses access. Seventy-two hours from commercial launch to global shutdown — no appeal, no timeline for restoration, no prior warning to any of the organizations that had already wired it into production.
The model was Fable 5, from Anthropic. We're not writing this to relitigate that decision or weigh in on national security policy. We're writing it because the G7 spent yesterday — June 17, in Évian, France — trying to figure out what to do about it at the nation-state level.
And nobody in that room was thinking about your 15-person nonprofit or your 30-person regional firm.
The Mechanism Nobody Was Watching
The legal instrument that made this possible is called the "deemed export rule." It's a provision in US export control law that treats sharing controlled technology with a foreign national anywhere in the world — including inside the United States — as an export subject to BIS jurisdiction. It's been used for decades on hardware, software, and dual-use technology with military implications.
It just got applied to a commercial AI API for the first time.
Organizations in Germany, Canada, Brazil, India — and many others — had already integrated Fable 5 into production workflows: diagnostic pipelines, document processing systems, risk models. They woke up one morning to find those workflows broken, with no recourse and no timeline. The company had to comply immediately or face federal enforcement. There was no window to prepare.
The discomfort isn't the specific decision — reasonable people can disagree on the underlying national security logic. The discomfort is this: it happened once. Now everyone knows it can happen again. And the next time it happens, it could be to any US-hosted AI service.
OpenAI. Google Gemini APIs. Azure OpenAI Service. Any of them.
What Governments Are Doing — And Why That Doesn't Help You
At the G7 in Évian this week, allied nations pushed back hard. European leaders cited their long-running concerns about "kill switch" dependency on US technology infrastructure. Canadian PM Carney said on the way to the summit that the incident highlights a need to "build out and diversify," and that sovereignty requires "unhindered access to AI." French President Macron announced a "trusted partners" platform among Western democracies to be stood up within a month.
US Commerce Secretary Lutnick pitched an exemption framework: a vetted list of allied nations and approved companies that would get access to restricted models even if a broad directive was in effect.
This is diplomatic progress. Genuinely.
And it will do almost nothing for your organization.
The "trusted partners" framework is being negotiated at the nation-state level. The entities getting seats at that table are governments, critical infrastructure operators, and large enterprises with lobbyists and established federal relationships. One analysis of the summit coverage noted: "does this scheme extend to a 50-person AI startup in Paris that just had its product break because it was using an API? The summit discussion seems oriented around nation-states and large enterprises. The long tail of the startup ecosystem barely gets a mention."
If you're a small NGO in Ottawa, a public sector delivery team in Winnipeg, or a regional professional services firm anywhere in Canada — you're downstream of this. Whatever framework emerges will be built for players ten times your size. You'll inherit the risk without the diplomatic protection.
What's Actually New Here
Up until this month, the vendor risk conversation for AI looked like this:
- Will they raise prices?
- Will they deprecate this model version?
- Will they have an outage?
- Will they change the API terms in a way that breaks our workflow?
Those risks are real and worth managing. We've been saying so for months.
But they're all fundamentally commercial risks. You could negotiate, plan around them, or switch providers in an organized way.
What just happened is different. It's not commercial risk. It's geopolitical risk applied to a commercial product. A government order that a private company has to comply with immediately, regardless of their customers' situations, regardless of existing contracts, regardless of whether the customer themselves is subject to any national security concern at all.
Your 12-person nonprofit in Halifax has no national security implications. Your environmental data processing pipeline is not a weapons system. But if it was running on an API that got pulled by a BIS directive, it stopped anyway. That's the nature of this risk category.
Three Things to Do This Week
This isn't a "wait and see" situation. The risk exists now. Here's what actually helps:
1. Map your AI dependency surface. Which workflows in your org would break if your primary AI API went dark overnight? Not "would be slower." What would actually stop working? Make the list. If you can't make it quickly, that's its own problem worth solving.
2. Pick an open-weight fallback and actually test it. Open-source models you run on infrastructure you control cannot be killed by a US government directive. That's not hyperbole — it's structurally true. Qwen3-32B, Gemma 4, and Llama 4 Scout are running near GPT-4 quality for most business use cases right now. Ollama runs them on a decent laptop. Pick one. Run it on two or three of your actual workflows. Know which tasks break and which don't. You want that knowledge before you need it urgently.
3. Decouple your prompts from the provider. If your workflows are hardcoded to a specific model's quirks — specific instruction formats, specific output structures — you have technical debt that's now also geopolitical exposure. The fix is a thin abstraction layer (LiteLLM is the standard tool here) and clean, tested prompts. An afternoon of work. The alternative is discovering on a bad Tuesday that you needed it.
The Question to Add to Your Vendor Checklist
If you're currently evaluating any AI-powered tool — workflow automation, document processing, anything that's going to become load-bearing in your operations — add one question to your conversation with that vendor: "What happens to our access to this product if the underlying AI provider is hit with a US export control directive?"
If the vendor looks confused, or says "we'd figure it out," that's information. They haven't thought about it, which means you'd be navigating the fallout together with no plan.
The better answer looks like: a named fallback model, support for open-source backends, or an on-premises deployment option. Not every vendor will have this. But now you know to ask.
The Part That Should Make You Uncomfortable
The AI infrastructure conversation used to be: who has the best model? Then: who has the best price per token? Now it's: who has access to models at all when governments start treating frontier AI like export-controlled weapons technology?
Nations with diplomatic leverage are getting frameworks. Large enterprises with lobbyists are getting frameworks. Small organizations are getting whatever's left.
That's not a reason to panic. It's a reason to build resilience into your AI stack now, while you have the time to do it calmly and systematically, rather than on the morning your production workflows stop working.
We help teams in the public sector, NGOs, and small businesses build AI workflows that don't depend on any single provider — and can survive the kind of disruptions that used to seem hypothetical. This week proved they're not. If your team is mid-implementation and hasn't thought through the fallback architecture, that conversation is worth having before it becomes an incident.