Talk to European companies about AI right now and one word keeps surfacing: sovereign. They want a local player — not just for their cloud, but for their AI.
“Which models are we allowed to use?”
“Where does the data actually go?”
“Is the provider European?”
A year ago these were compliance footnotes. Today, they’re often the first question in the room. Cloud and AI have picked up a geopolitical dimension, and European customers increasingly want European answers.
That’s a reasonable thing to want. It’s also, right now, a genuinely hard thing to deliver. I’d rather tell you that honestly than sell you a comfortable story. The problem shows up on two fronts.
Most SaaS Platforms Aren’t Fully EU (Yet)
The first is the cloud platform itself. Some applications still don’t run fully in the EU. And, on August 17th, Atlassian began using cloud-product data to help train its AI in the US. Customers can (and generally should) switch off in-app data contribution, and for most organizations the “this is a GDPR breach” framing is vastly overstated. But, there’s a real, unresolved problem underneath it: data-residency settings cover data at rest, and it isn’t yet clear if they exclude your data from the AI training pipeline, which can route through US processors. For customers with strict residency requirements, that gap matters. And some are voting with their feet and leaving the ecosystem over it. Our job isn’t to tell you Atlassian is wrong; it’s to help you get a written answer and make an informed call before the deadlines that matter to you.
Frontier AI Models Run on the Vendor’s Servers
The second front is AI itself. The frontier models — the best ones — are essentially SaaS: you send your data over an API to run on the vendor’s servers, and almost all of those vendors are American. The alternative is open-weight models you run on your own hardware, where nothing leaves your building. But here’s the catch European customers keep hitting: the strongest open-weight models are mostly Chinese. And for a lot of organizations, that’s a red flag.
A European Model That Competes Is Nowhere to Be Seen
So they ask for a European model. And here I have to be blunt, because I’d rather be useful than diplomatic: there isn’t a serious one yet. Mistral is the only real contender, and it’s both behind on quality and expensive to run — to the point that they’ve started renting their compute out to Microsoft. I genuinely root for the European players; I would love a European model trained on European data to work, and we’ll deploy it the day it does. But wanting it doesn’t make it exist. Europe accounts for something like 5% of global AI compute. That’s the honest starting point.
I’m not saying this from the sidelines. We run open-weight models in production on our own hardware. We replaced a US cloud AI service with a self-hosted model for our internal support system, and we deliberately run sensitive material through it because nothing leaves our servers. That includes testing Chinese models directly. Which brings me to the part people find uncomfortable.
We Evaluate Every Serious Option on Its Merits
We test Chinese open-weight models. Not because we’re promoting them — I’m not here to wave a flag for anyone’s models — but because they’re a real option in the market. They aren’t banned and they aren’t illegal; they just feel strange to use in the current climate. And assessing them honestly is exactly what responsible advice requires: we evaluate every serious option on its merits, including the ones that feel politically uncomfortable, because ruling something out on reflex instead of evidence doesn’t serve you. What we won’t do is pretend the map is simpler than it is.
Because it rarely is black and white. A European model that’s “behind” might be exactly right for your use case: if your requirements fit what it does well, the fact that it doesn’t top a leaderboard is irrelevant. It’s the same call we help customers make between Atlassian and open-source alternatives like XWiki or OpenProject: sometimes the “lesser” option is the right one, when it fits your situation. Sovereignty, quality, cost, and use case are dials, not a single switch; the right setting looks different for a public-sector body handling citizen data than for a team summarizing meeting notes.
Our Advisors Find the Setup That’s Right for Your Situation
So here’s what we actually do. We track this market constantly and test the models ourselves — frontier and open-weight; American, Chinese, and European — so that when you ask “what can we rely on for this use case?,” the answer comes from the bench, not a brochure.
European alternatives exist. They’re behind and they’re expensive, but depending on what you’re doing, that might be perfectly fine. Talk to our advisors and we’ll help you find the setup that’s right for your situation, sovereignty included.
tl;dr: Sovereignty is a real requirement. Pretending it’s already a solved one helps no one.