Avery Pennarun
When you choose an AI provider, your first question is whether it can do the job you need. There are other decisions too: where your data goes, where your software runs, what else you depend on, and what happens if you want to change providers later.
Those decisions determine how much control you keep.
And that’s the real question behind sovereign AI. Sovereignty makes it sound like something mainly for governments, but I think the question of control applies to companies and individuals. You should be able to use the best technology without handing one provider control over everything, including your data. Sovereignty should mean getting the best AI and ecosystem, on your own terms.
Today we announced that Tailscale is making a multi-million-dollar, multi-year investment in sovereign AI. Patricia Prewitt has joined us as General Manager of Sovereign AI, to lead the work.
I first met Patricia when she was on the Rapid Evaluation team at Google X, Google’s Moonshot Factory. She later became General Manager of another moonshot project, developing AI to design physical structures for silicon photonics. There isn’t a tidy specification for sovereign AI yet, and Patricia knows how to take an ambitious technical idea, figure out what would have to be true for it to work, and organize the work to find out.
Her job at Tailscale is to turn digital sovereignty into something concrete that anyone can use, anywhere in the world: deciding what we should build, where we should partner, and how to make the pieces work together. Fast.
What sovereignty really means
A lot of discussion about sovereign AI starts with national borders. Keep the data in the country. Buy local compute. Run your own models. And most importantly, control your entire system.
Trying to eliminate every external dependency doesn’t quite work. You’d need your own chips, servers, operating systems, and models, plus everything that manufactures them. Eventually you’re building the machines that build the machines. In today's world, no country can do all that alone.
Modern computing depends on other people’s work. The question is how much control you keep over those dependencies once you rely on them.
You should be able to inspect them, change them, replace them, or choose something else without rebuilding everything around them.
This applies at every scale. A country wants domestic compute without also having to train its own model. A company needs data to stay in one datacenter while using compute elsewhere. A person using an AI agent to build an app needs to own the resulting code, run it where they want, and switch AI providers at any time.
You need to own your data. But ownership only matters if you can do something with it. You need to use your data with the software you choose, under rules you control.
Choice requires interoperability
An alternative compute provider isn’t real if using it requires a six-month rewrite. The capacity may exist, but in practice you can't move. The same is true for models, storage, and applications. If changing one part means rebuilding around it, you don’t have real freedom.
That’s why interoperability matters so much. Complex systems need to work together even when parts come from different providers.
Tailscale is already the interconnect of choice for nearly every AI company. Our mission is to make safe connection effortless, so where something runs doesn’t dictate what it can do. You can host your own AI infrastructure using Tailscale and Aperture today, across any locations you want, with control over who can reach what and when.
We’ve already been invited to sovereignty discussions across Canada, Europe, and several countries in Asia, including projects where Tailscale is in use today and others where it has been proposed. As a Canadian company, we know a lot about how to connect across vendors, countries, and politics without creating unhealthy dependency.
Tailscale isn't going to build models, GPUs, or datacenters. That would recreate the problem we’re trying to solve. Our job is to help glue together an ecosystem of the parts you already want, with the freedom to change any of the parts, whenever you want.
An open alternative has to be so real you actually use it.
Turning this into something useful
We’re not announcing a finished roadmap today. Patricia will work with our internal teams, customers, and partners to figure out where the barriers to choice are, how we can refine our own product, and where somebody else is better placed to solve a problem.
“Usually my question isn’t ‘Can we do this?’ It’s ‘What will it take to do this?’” — Patricia Prewitt, GM of Sovereign AI
Tailscale, and the world, already make the components you need for sovereignty. What's left is the fine tuning and integration, to create complete solutions. We'll have more to announce later this year.
For now, we want to hear where your needs break down today. Which model can’t you use? What can’t you do without becoming dependent? Where does changing one provider force you to rebuild everything? Tell us what you’ve tried and where you’re stuck.
You shouldn’t have to, as Carl Sagan said, invent the universe to make an apple pie. A whole industry is here to make that work. Let's do it together.
