The US government has ordered Anthropic to block access to its most advanced AI models-Claude Fable 5 and Mythos 5-for all foreign nationals worldwide. This restriction applies even to foreign nationals within the US, including Anthropic's own employees.
Anthropic couldn't even comply selectively. The only way to follow the order was to shut the models down for everyone, everywhere, including American citizens.
Overnight, a model that businesses around the world had built their workflows on simply disappeared.
This is the worst fear of geopolitical experts and tech policy watchers, playing out in real time.
Sovereign AI proponents have long argued: if you don't own the model, you don't own your access to it. A single directive, in a single jurisdiction, can switch off a capability your business, or your country, has come to depend on.
This is no longer a hypothetical situation for a policy paper. It happened. This Friday.
And it makes the case for sovereign AI models trained, hosted, and governed within a country's own jurisdiction not a nationalist talking point, but a genuine continuity and security imperative.
Which brings me to something else that happened this week.
Anthropic founder Dario wrote a policy paper to highlight the gap between how fast AI moves, and how slowly governments respond.
The irony is hard to miss. Dario's essay, "Policy on the AI Exponential," calls for governments to step in with thoughtful, binding regulation, because AI is moving too fast for slow institutions to keep pace.
Days later, the US government stepped in. Just not in the way he was asking for.
This is exactly the kind of blunt, unilateral, knee-jerk, security-driven intervention that founders and policy watchers alike have feared, the kind that smart regulation is supposed to pre-empt, not become.
Still, the essay itself is a seminal moment for all of us in the policy world. Not for what it says about AI, but for who is saying it. The founder himself.
Here is a founder building one of the most powerful AI systems on the planet, saying governments must step in with binding, enforceable law. Not self-regulation.
When I entered public policy over a decade ago, the doctrine was: innovate first, regulate later. I lived it during the Uber era: technology outruns the regulator, and that's fine, regulations will catch up later.
AI has destroyed that philosophy. In four years, AI went from barely writing a line of code to writing most of the code at major tech companies. You cannot afford regulation to be slow when the technology is moving at this pace and rewriting itself every few months.
And yet there are no easy answers. I think about this every time I sit across from a policymaker. In conversations with India's IT Secretary S Krishnan, more recently with UK AI Minister Kanishka Narayan, and with regulators across jurisdictions, the sentiment is the same: It is genuinely hard to be a policymaker right now.
How do you balance twin objectives: Don't throttle innovation. But don't let a civilisation-altering technology, or your access to it sit in someone else's hands without guardrails.
India's position has been considered: pro-AI, legislate when the need is clear.
But these events suggest the need is already here on two fronts. Smart, anticipatory regulation of AI capability. And sovereign AI infrastructure, so that no country is one Friday-evening directive away from disruption.
(The author is President and CEO of Startup Policy Forum, an industry alliance that represents 75+ high growth startups and an active voice on all tech policy issues)
Disclaimer: These are the personal opinions of the author
