Sweden's sovereign AI push just got the most concrete product it's ever had. Berget AI, the Kista-based infrastructure startup, launched Berget Code: an agentic coding service that keeps your source code physically inside Swedish borders while still delivering something close to the experience of Claude Code or Codex.

That isn't a small thing. Most Swedish developer teams use American coding assistants. Most of those assistants run inference in the United States. For organizations bound by GDPR, the Data Privacy Framework, or sector-specific rules in finance, healthcare and public administration, that pipe is a problem. Many of those teams have been blocked from using AI coding tools at all.

Berget Code's pitch is the obvious one. Run the same workflow against open models hosted in Sweden, with the same VS Code or terminal-style developer experience, and your repos never leave the country.

Berget AI was founded in 2024 by Christian Landgren and team, operating mostly out of view in Kista, Sweden's deeptech hub. The company raised €2.1 million in seed funding in February from investors who were already convinced that Europe's AI infrastructure can't depend permanently on US hyperscalers. Berget Code is the company's first real consumer product, and it's an aggressive one.

Landgren's framing in his own launch post is worth quoting. American tools have, in his words, started raising prices, tightening availability and deprioritizing sustainability. Whether or not you buy each of those claims, the strategic story for Berget Code lands cleanly. European developers want optionality. Berget is the first Nordic company actually offering it as a usable product, not a roadmap slide.

The product launches with Kimi K2.6 as the underlying model, paired with a 256,000 token context window. That's a meaningful number. It's enough to load most enterprise repos into a single conversation, which is the unlock that turned coding assistants into production tools in late 2025.

The honest objection is the obvious one. Open models have historically been worse at coding than proprietary frontier models. Claude and OpenAI's coding tools weren't winning developer love because of marketing. They were winning because they wrote better code. Berget's bet is that, as the Berget team puts it, in December 2025 open coding models crossed the quality threshold where they're useful for real development work, and they've kept improving fast.

If that thesis holds, the strategic ground shifts. Suddenly you don't have to compromise on quality to get sovereignty. You just have to wait for the open models to catch up, then bundle them with EU-hosted inference and a credible developer surface. Berget Code is the cleanest implementation of that strategy on the market right now.

Berget Code's headline rivals are obvious. Claude Code and Codex CLI set the bar. But Berget's actual battle is closer to home. France's Mistral has its own coding strategy. Aleph Alpha in Germany has long pitched sovereign AI to enterprise buyers. Even hyperscalers are now offering EU residency for inference. None of them, though, has shipped a real coding-first product that runs entirely on European infrastructure and treats sovereignty as the headline feature instead of a footnote.

That distinction matters. Buying a generic LLM API and then telling your developers to figure out the workflow is what European AI infrastructure has looked like for two years. Berget is, finally, presenting a packaged experience that a regulated team can adopt without inventing a new developer story from scratch.

Three forces are converging. First, open models hit useful quality for coding sometime in late 2025. Second, EU regulatory pressure on data residency has gotten sharper, especially in healthcare and public sector. Third, American coding assistants started behaving more like infrastructure than software: pricing changes, regional restrictions, capacity rationing. Each of those, on its own, would be a manageable inconvenience. Together, they create a procurement opening for Berget that didn't exist a year ago.

There's an unexpected wrinkle. Berget Code's launch model is Chinese-origin (Kimi K2.6, from Moonshot AI), even though the inference itself runs in Sweden. That's a quiet reminder that the geography of model weights is now decoupled from the geography of inference. Sovereign AI in Europe doesn't necessarily mean European models. It means European compute, European data handling, European jurisdiction. The model could be from anywhere, so long as the weights stay where they're told.

Sweden has spent the last 18 months trying to assemble the components of a sovereign AI stack. EuroLLM, the various supercomputing initiatives, and a steady drumbeat of policy work. Berget Code is one of the first products you can actually use today that closes the loop. Compute in Sweden, model running locally, developer workflow ready to adopt.

If even a few hundred Swedish engineering teams switch over in the next quarter, that's a leading indicator for the entire European sovereign AI thesis. If it doesn't happen, you'll know the demand for sovereignty was softer than the policy conversation suggested. Either way, Berget Code just made the experiment runnable in production. Worth watching closely.