Sweden Is Europe's GPU Colony. Nobody Asks Who Owns Inference.

Sweden is becoming the EU's favorite place to plug in GPUs. That is not the same as becoming an AI power.

This week, Mistral AI and EcoDataCenter announced a billion-dollar AI infrastructure push in Sweden. HIVE Digital is upgrading its AI cloud infrastructure here. NVIDIA is all over Europe talking about its "next industrial transformation," and Sweden is exhibit A. The headlines are glowing. Politicians are proud. The narrative writes itself: Sweden, land of cheap hydro, cold air, and stable grids, is becoming the backbone of European AI.

I run a software development company in Jönköping, Sweden. I should be thrilled. More compute in the country means more ecosystem, more jobs, more proximity to the action. Right?

Here is my problem: I look at this and I see a colony being built. A very well-cooled, very green-powered, very politely negotiated colony. But a colony nonetheless.

The landlord trap

Let me be blunt about what is actually happening. Foreign companies are coming to Sweden because we have three things they want: renewable energy that is cheap enough to make GPU clusters economically viable, ambient temperatures that reduce cooling costs by 30-40%, and a political environment that does not ask hard questions. They get the compute. They train the models. They run inference. They capture the margin. Sweden gets construction jobs, electricity revenue, and a warm feeling about being "part of the AI revolution."

This is the Nordic resource curse dressed up in a rack-mount chassis.

We have been here before. Sweden mines iron ore and ships it south for others to turn into cars. We generate some of Europe's cleanest electricity and watch it power someone else's intelligence stack. The pattern is consistent: we export raw inputs and import finished products at a markup.

The difference this time is that the "finished product" is not steel or manufactured goods. It is intelligence. And intelligence compounds. The company that owns inference owns the customer relationship, the data loop, the feedback cycle that makes the next model better. The company that owns the building where inference happens owns a building.

What Stockholm is not asking

I keep waiting for someone in the Swedish government, at Vinnova, at the AI Commission, to ask the following questions:

• When Mistral trains models in Swedish data centers, who owns the trained weights?
• When HIVE sells GPU-as-a-service from Boden, where does the revenue go? Where are the taxes paid?
• If these companies decide to move their compute to Iceland, Norway, or Finland next year, what is left?
• What percentage of the inference revenue generated on Swedish soil stays in Sweden?
• Are we building sovereign AI capability, or are we building someone else's?

Nobody is asking. The EU talks about ending its reliance on US internet technology, but the actual policy response is to subsidize foreign compute providers to build on European soil. That is not sovereignty. That is hosting.

The Nordic picture vs. the global reality

Let me zoom out. In the US, the AI value chain is vertically integrated. NVIDIA makes the chips. OpenAI, Anthropic, Google build the models. Microsoft, Amazon, Google run the cloud. American companies own every layer. The margin stays in America at every step.

In China, the government mandates that AI infrastructure built on Chinese soil serves Chinese strategic interests. The models trained there are Chinese models. The data stays in China. It is heavy-handed, but the value capture is intentional.

In Sweden, we are building the physical layer and hoping the value trickles down. It will not. Value in AI flows to whoever owns the model, the inference endpoint, and the customer relationship. Not to whoever owns the concrete and the copper wiring.

Compare this to what Skellefteå is doing with Northvolt and now Lyten's new industrial-tech hub. That corridor has a different logic. Battery manufacturing keeps IP in Sweden. The factory is not just a building. It produces a product that Swedish and European companies own, brand, and sell. There is a clear value chain where Sweden captures margin at multiple layers.

The AI data center play has no equivalent structure. EcoDataCenter builds the facility. Mistral fills it with GPUs. Mistral trains and serves models. Mistral's customers pay Mistral. EcoDataCenter gets rent. Sweden gets a line item on its renewable energy export balance sheet.

What actual software development in Sweden looks like from here

I am writing this from Jönköping, not from Stockholm or San Francisco. From here, the view is different. We at HEIMLANDR build AI solutions and custom SaaS development for clients who need to own their own technology. That is the whole point. Ownership. Control. Margin capture.

The same principle that applies to a mid-market company building its own SaaS product instead of renting someone else's applies to a country building AI infrastructure. If you do not own the intelligence layer, you do not own the business. You are a cost center for someone who does.

This does not mean Sweden should try to build its own foundation model from scratch. That ship has sailed, and the cost structure does not support it. But there is a massive space between "build GPT from scratch" and "rent out buildings to people who build GPT." That space is where Swedish software development companies, and Swedish policy, should be operating.

That space includes:

• Fine-tuned, domain-specific models for Nordic industries (forestry, mining, maritime, energy, medtech)
• Inference infrastructure owned and operated by Swedish and European companies
• Application layer AI, the agentic systems and vertical SaaS that actually touch customers
• Smart contract development and blockchain development for transparent, auditable AI supply chains
• Self-hosted, sovereign deployments where data and models stay under your control

Where this goes in 2-5 years

Here is the trajectory I see. And it is not comfortable.

Compute becomes more commoditized. The current GPU shortage eases as NVIDIA, AMD, and custom silicon from hyperscalers ramp production. When that happens, the "Sweden has cheap power" advantage shrinks. The data centers will not leave overnight, but the leverage shifts. New facilities will go wherever the next incentive package is sweetest.

Meanwhile, the inference layer, where the real money is, consolidates around 3-5 global players. If Swedish companies have not built positions in the application and model layer by then, they will have no seat at the table. They will be tenants in their own country's infrastructure.

The path toward AGI makes this worse, not better. As models become more capable, the gap between "owning compute" and "owning intelligence" widens. A general intelligence system running on Swedish power but controlled by a French or American company does not make Sweden an AI leader. It makes Sweden a power plant.

The EU AI Act, for all its ambition, does not address this. It regulates what AI can do. It says nothing about who captures the economic value of AI built on European resources. That is a massive gap. Swedish regulators are not filling it. They are too busy celebrating the inbound investment numbers.

What to look at

If you are a founder, CTO, or senior engineer in Sweden or Europe, and this resonates, here are concrete things to look at right now:

awesome-selfhosted (300k+ stars on GitHub). This is the single best directory of self-hosted alternatives to every SaaS service you currently depend on. If sovereignty matters, and it should, start here. Every service you self-host is one less dependency on someone else's stack.

n8n (193k stars). Fair-code workflow automation with native AI capabilities. Self-hostable. This is how you build AI-powered workflows without routing everything through US cloud providers. We use it internally at HEIMLANDR for client automation pipelines.

Daytona (72k stars). Secure, elastic infrastructure for running AI-generated code. If you are building AI agents or AI-native applications, you need sandboxed execution environments that you control. Daytona gives you that without handing everything to a hyperscaler.

Netdata (79k stars). If you are going to run your own infrastructure, which you should consider, you need observability. Netdata is fast, open-source, and does not require a PhD in Prometheus to set up.

So what should you actually do?

If you are a Swedish tech company: stop celebrating the data centers and start asking what layer of the stack you own. If the answer is "none," you have a strategic problem that will get worse every quarter.

If you are building SaaS: build it so it can run on European infrastructure, owned by European companies. Your customers will increasingly demand this. GDPR was the appetizer. Data sovereignty for AI workloads is the main course.

If you are a founder: the opportunity is not in competing with Mistral or OpenAI on foundation models. It is in the vertical application layer, domain-specific models, agentic systems, and the custom software that connects AI capability to actual business value. That is where the margin lives. That is where we build.

If you are in government: wake up. Ask who owns what. Tie infrastructure incentives to value-capture requirements. Norway does this with oil. There is no reason Sweden cannot do it with compute.

The uncomfortable comparison

Norway created the Government Pension Fund Global, one of the largest sovereign wealth funds on earth, by insisting that the extraction of Norwegian resources benefit Norwegian citizens. They did not just let Shell and BP pump oil and say "thanks for the jobs."

Sweden is letting Mistral, HIVE, and NVIDIA plug into our grid and saying "thanks for the jobs." There is no AI sovereign wealth fund. There is no requirement that models trained on Swedish power contribute to Swedish AI capability. There is no strategic thinking about value capture at all.

We are a resource-rich country behaving like a resource-poor one. Cold air and clean electrons are resources. They have value. We should be extracting concessions, equity positions, model access, inference quotas, technology transfer agreements. Instead, we are rolling out the red carpet and hoping for the best.

From Jönköping, that looks like a mistake. A quiet, well-intentioned, very Swedish mistake. But a mistake.

The data centers are not the problem. They are fine. Let them come. The problem is that we are treating them as the goal instead of as the foundation for something that Sweden actually owns. Until that changes, we are Europe's GPU colony. Clean power, cold air, and someone else's margin.

Fredrik Brunnberg is the CEO of HEIMLANDR.IO, building AI and software solutions from Jönköping, Sweden. This is the daily HEIMLANDR briefing. If you found this valuable, share it with someone who builds things.