Sweden Is AI's Landlord, Not AI's Builder

The Billions Are Real. The Story We Tell About Them Is Not.

Mistral AI is putting $1.43 billion into a data center in Borlänge right now. NVIDIA is setting up an AI hub in Göteborg. Swedish politicians are on stage talking about how Sweden is becoming an AI superpower. Headlines in Dagens Industri are glowing. Everyone is happy.

I am not happy. I am in Jönköping, running an AI development company in Europe, and from where I sit, what I see is Sweden becoming AI's landlord. Not AI's builder. There is a massive difference. And every CTO in the Nordics needs to understand it before they mistake proximity for competitiveness.

Here is the uncomfortable version of what is happening: foreign companies are buying Swedish energy, Swedish cooling capacity, and Swedish real estate. They are not buying Swedish intelligence. They are not co-developing models with Swedish researchers. They are not funding Swedish foundation model labs. They are renting our cold climate and our cheap hydropower. We are the cloud's basement.

What the Money Actually Buys

Let me be specific. Mistral's deal with EcoDataCenter in Borlänge is an infrastructure play. Compute. Power. Cooling. Racks. This is not an R&D center. Mistral's research team sits in Paris. Their models are trained in Paris. Their intellectual property stays in Paris. Borlänge gets construction jobs and electricity bills.

NVIDIA's Göteborg hub sounds impressive until you read the details. NVIDIA is there as a compute supplier. They are not co-developing Swedish language models. They are not building AI agent development frameworks with Swedish universities. They are selling GPUs and cloud time. Göteborg gets a logo on a press release.

Meanwhile, look at the Swedish AI companies that are actually raising real money. Lovable just hit a $6.6 billion valuation. Legora raised at $5.55 billion. Dentio has a16z backing. These are real companies with real products. But they are all application-layer plays. Every single one of them runs on top of American or French foundation models. OpenAI, Anthropic, Mistral. The intelligence is imported. The value-add is in the interface, the workflow, the vertical integration. That is legitimate business. But it is not sovereign AI. It is not the same thing as building the brain.

Swedish AI-native startups tripled their funding to €454 million according to recent Breakit reporting. Sounds great on a slide. But trace where that money goes. Almost all of it flows into companies consuming foreign AI, not creating it. We are paying API bills to San Francisco and Paris. That is the real flow of value.

The Sweden vs. World Comparison Nobody Wants to Make

Let me put this in context.

France has Mistral. A homegrown foundation model company, now valued at billions, training their own models, controlling their own weights. The French government made this a strategic priority. They backed it with policy and capital.

Germany has Aleph Alpha (pivoted, sure, but they tried) and a growing ecosystem around open-source model development. The German BMWK treats AI sovereignty as industrial policy.

The UAE built Falcon. A state-backed foundation model. Not because they needed to. Because they understood that controlling the intelligence layer is geopolitical power.

China has DeepSeek, Qwen, GLM-5, MiniMax. Multiple foundation model labs, all domestically funded, all operating at frontier scale.

Sweden has... data centers. And application companies. And a government that seems to think hosting foreign compute is the same as building AI capability.

From San Francisco, Sweden looks like a great place to put servers. Cheap energy. Cold air. Political stability. Good fiber. That is it. Nobody in the Valley is talking about Swedish AI research as a competitive force. They talk about Swedish energy prices.

From Jönköping, where we at HEIMLANDR build custom AI solutions and AI agent development for clients across Europe, I see this gap every day. When we build AI automation for a business, we are integrating with OpenAI, Anthropic, or open-source models from Hugging Face's transformers library. The tools are great. But none of the foundational intelligence is Swedish. We are assemblers, not architects. And we should be honest about that.

Why This Actually Matters for Your Business

If you are a CTO or founder in the Nordics, you might be thinking: "So what? We use the best tools available. We build great products on top of them. Who cares where the model was trained?"

Fair question. Here is why you should care.

Dependency risk. Every API call you make to a foreign model is a dependency. Pricing changes, policy changes, export restrictions, geopolitical shifts. You do not control the roadmap. When OpenAI changed their pricing model last year, hundreds of Swedish startups had to rewrite their unit economics overnight. That is not resilience. That is fragility dressed up as efficiency.

Data sovereignty. EU regulations under the AI Act are getting stricter. GDPR already makes it complicated to send personal data to US-hosted inference endpoints. The gap between what regulators want and what Swedish companies actually do is growing wider. Running your AI automation business on American API endpoints while telling customers their data is safe is a position that gets harder to defend every quarter.

Value capture. The biggest margins in AI are in the model layer and the infrastructure layer. Sweden has given away the infrastructure layer to foreign capital. It never built the model layer. What is left? The application layer, which is the most competitive, most commoditized, and lowest-margin part of the stack. We are fighting over scraps while the feast happens elsewhere.

Talent drain. Swedish AI researchers with real model-training expertise go to London, Paris, San Francisco. Because that is where the frontier labs are. We produce excellent ML graduates at KTH, Chalmers, Lund. Then we watch them leave. Data centers in Borlänge do not bring them back. Research labs would.

Where This Goes: 2027-2030 and the AGI Question

Let me look forward, because the trajectory here is what should scare Swedish policymakers.

We are maybe two to four years from AI systems that can do most knowledge work. Not all of it. But most of it. The path toward AGI, whether it arrives in 2028 or 2032, passes through a phase where controlling the intelligence layer becomes the most important strategic asset a country can have. More important than oil. More important than manufacturing.

Countries that control foundation models will set the terms. They will decide what languages are supported, what values are embedded, what use cases are enabled or restricted. Sweden, without its own model capability, becomes a policy-taker. Not a policy-maker.

The EU AI Act is supposed to address some of this. But the act was written for a world that already does not exist. It regulates deployment, not dependency. It says nothing meaningful about what happens when your entire economy runs on two or three American model providers. Swedish regulators at Integritetsskyddsmyndigheten (IMY) are already overwhelmed by GDPR enforcement. Adding AI governance on top, without increasing capacity or expertise, is a recipe for regulation that exists on paper and nowhere else.

Here is my prediction: by 2028, at least one major European country will launch a state-backed foundation model initiative comparable to what the UAE did with Falcon. It will not be Sweden. It will be France or Germany. Sweden will still be celebrating data center investments.

For the AI agent development cost equation, this means something concrete. Right now, building an AI agent on top of GPT-5 or Claude 4 is economically rational. The models are good, the APIs are accessible, the cost per token keeps dropping. But the long-term cost is dependency. And dependency always gets more expensive over time, never less.

What Builders Should Actually Do

I do not write articles just to complain. Here is what I think Nordic builders, founders, CTOs should actually be doing right now.

1. Invest in open-source model competence internally. Even if you use closed APIs today, your team should know how to fine-tune, deploy, and run open models. Tools like Ollama make it trivial to run models locally. It now supports Kimi-K2.5, GLM-5, DeepSeek, Qwen, and others. This is your insurance policy.

2. Build the orchestration layer, not just the prompt layer. The real value in AI agent development is not in calling an API. It is in building robust orchestration, memory, tool use, and multi-agent workflows. Frameworks like Langflow (147k+ stars on GitHub) and n8n (185k+ stars) give you visual workflow building with AI capabilities built in. The companies that own the orchestration will survive model switches. The ones that just wrap APIs will not.

3. Push your government representatives. If you are a Swedish founder or CTO, the next time someone from Vinnova or Tillväxtverket asks what the AI ecosystem needs, do not say "more data centers." Say "fund a Swedish open-source foundation model project." Say "create tax incentives for AI R&D, not just AI hosting." Say "match what France is doing." Be specific. Be loud.

4. Think about data as a strategic asset. Swedish companies sit on incredible domain-specific data. Healthcare data, industrial data, environmental data. That data is more valuable as training material for specialized models than as rows in a database. Start thinking about how to use it. Not by sending it to OpenAI, but by building local model capability around it.

What to Look At

If you are technical and want to go deeper, here are concrete things worth your time this week:

• Ollama for running models locally. The project has exploded to 170k stars because people understand that local inference matters. Run DeepSeek, Qwen, or Gemma on your own hardware. Understand the tradeoffs.
• Langflow for building AI agent workflows visually. If you are evaluating AI agent development cost for your organization, start here. It dramatically reduces the time from idea to working prototype.
• Hugging Face Transformers remains the definitive open-source model library. If your engineering team is not comfortable working with this, fix that.
• HEIMLANDR's Rapid MVP service if you need to test an AI-powered product idea fast without burning six months on architecture. I am biased, obviously. But we built this service specifically because Nordic companies keep over-planning and under-shipping.

The Real Question

Sweden is good at building products. We have Spotify, Klarna, King, a hundred other companies that turned Swedish engineering talent into global businesses. We are not bad at this. We are good at this.

But we are making a strategic error by confusing infrastructure investment with capability investment. A data center in Borlänge does not make Sweden an AI power any more than a parking garage makes you a car manufacturer. The steel is not the engine.

I write this from Jönköping, where we at HEIMLANDR build AI agents and custom software for companies across Europe every day. I see what our clients need. I see what models they depend on. I see the invoices going to San Francisco. And I think Sweden deserves better than being the continent's server room.

The question is not whether AI investment in Sweden is good. It is. Jobs are jobs. Tax revenue is tax revenue. The question is whether we are going to be satisfied with that, or whether we are going to build something that actually belongs to us. The window for that decision is closing. Foundation models are getting bigger, more expensive, and harder to catch up to. Every year we wait, the gap gets wider.

So the next time you see a headline about billions flowing into Swedish AI, ask one question: whose AI? If the answer is always someone else's, we have a problem. And no amount of cheap electricity fixes it.

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.