Sweden Exports Compute, Imports Dependency: The Cloud Colony Problem

Mistral AI is building a data center in Sweden right now. They are partnering with EcoDataCenter because we have cheap hydropower, cold air, and political stability. A French AI company looked at the map of Europe and picked Swedish soil to run inference on. Meanwhile, the average Swedish startup I talk to runs its entire production stack on AWS us-east-1. Virginia. A data center 6,000 kilometers away, owned by a company in Seattle, governed by US law.

Let that sit for a second. Foreign AI companies are choosing Sweden to build sovereign infrastructure. Swedish companies are choosing America to rent it. We are becoming a data colony that exports raw compute and imports dependency. And nobody in the Swedish tech scene wants to have this conversation.

The Cheap Energy Trap: Sweden as a Compute Farm

Sweden has some of the cheapest electricity in Europe. We have a stable grid. We have naturally cold climate that reduces cooling costs by 30-40%. We have a functioning rule of law and relatively sane permitting processes for industrial builds. This makes us extremely attractive for data center operators.

Mistral and EcoDataCenter see this clearly. So does Microsoft, which has been expanding Azure capacity in Gävle and Sandviken. So does every hyperscaler scouting Nordic locations. They are not here because they love fika. They are here because our energy is cheap and our air is cold.

This is the part Swedish politicians celebrate. Jobs. Investment. Green compute. The narrative is that Sweden is becoming a hub for AI infrastructure. And technically, that is true. But whose AI? Whose infrastructure? Whose models are running on those racks?

Not ours.

The value chain in AI looks like this: data collection, model training, inference, application. Sweden is positioning itself at the bottom of that chain. We provide the electricity and the real estate. The models, the APIs, the billing, the terms of service, the kill switches. Those stay in San Francisco, Seattle, and now Paris. We are the mine. They are the mint.

The Swedish Startup Stack: Made in Virginia

I run HEIMLANDR.IO from Jönköping. When I talk to founders and CTOs at Swedish companies, the infrastructure story is almost always the same. AWS. Sometimes Azure. Occasionally GCP. Almost never anything European, let alone Swedish.

And I get it. The developer experience on AWS is good. The ecosystem is massive. The tooling works. When you are a five-person startup trying to ship a SaaS product and you need to move fast, you do not want to evaluate Nordic cloud providers. You want to type terraform apply and have it work.

But here is what is happening at scale. Every Swedish company building on us-east-1 is:

• Subject to US jurisdiction, including CLOUD Act subpoenas
• Dependent on pricing decisions made in Seattle
• Sending Swedish customer data across the Atlantic
• Paying an infrastructure tax that leaves the Swedish economy entirely
• Building on a platform that can change its terms, deprecate services, or raise prices at any time

This is not theoretical. AWS has raised prices. They have deprecated services. They have changed terms. And when the US government decides it wants access to data stored by a US company, CLOUD Act gives it a clear path. GDPR says otherwise, but good luck enforcing that from Jönköping.

Swedish software development is world-class. Our engineers are some of the best. But we are building castles on rented land, and the landlord speaks English and answers to a different legal system.

The Open Source Funding Boom: Necessary but Not Sufficient

There is real money flowing into open-source AI right now. Featherless.ai just raised $20M, backed by AMD and Airbus, explicitly to build open-source AI infrastructure free from Big Tech control. RadixArk launched with $100M to democratize frontier AI inference via SGLang. Sweden's own 2026 open-source prognosis calls this AI's next golden age.

I am glad this money exists. Open weights matter. Open training code matters. But here is what the open-source evangelists keep glossing over: the inference layer is still controlled by three American hyperscalers.

You can have the most open model in the world. Mistral's weights are open. Llama's weights are open. But when you run inference, you need GPUs. You need networking. You need a data center. And right now, the economically viable way to do that at scale is AWS, Azure, or GCP. Maybe Oracle if you are feeling adventurous.

Open-source solves the software layer. It does not solve the hardware layer. It does not solve the physical layer. And the physical layer is where the real lock-in lives. You cannot fork a data center. You cannot git clone a GPU cluster.

This is why Mistral building in Sweden is interesting but not liberating. They are building their own inference capacity on Swedish soil, but they control it. Swedish companies using Mistral's API are still renters. Better renters, maybe. Renting from a European landlord instead of an American one. But still renters.

Sweden vs. The World: How Other Countries Handle This

France gets this. The French government has been aggressive about AI sovereignty. They backed Mistral early. They are building national compute capacity through programs like Jean Zay. They treat AI infrastructure like they treat nuclear energy: a matter of national independence.

China obviously gets this. They have been building domestic AI infrastructure for years, partly by choice, partly because US export controls gave them no alternative.

The US does not need to worry about this because the hyperscalers are American. The dependency runs in their favor.

And Sweden? Sweden has Vinnova. Sweden has AI Sweden. Sweden has lots of reports and strategies and roundtable discussions. What Sweden does not have is a national compute infrastructure that Swedish companies actually use. We have no sovereign cloud that competes on developer experience. We have no government mandate that sensitive Swedish data must run on Swedish or EU-controlled infrastructure. We have the theoretical protection of GDPR but no practical enforcement that makes American cloud providers sweat.

The EU AI Act is now live. It has rules about high-risk AI systems, about transparency, about data governance. But it says almost nothing about where the compute happens. You can be fully compliant with the AI Act while running your entire inference stack on AWS us-east-1. That is a gap you could drive a data center through.

Even within the Nordics, there is variation. Finland has invested heavily in national AI compute through CSC and the LUMI supercomputer. Denmark has been more pragmatic about cloud procurement requirements. Sweden, as usual, has written a strategy document and formed a commission.

Where This Goes: 2027-2030 and the AGI Question

Let me project forward.

If current trends hold, by 2028 Sweden will host significant AI data center capacity. Most of it will be owned by foreign companies. Swedish companies will be more dependent on American cloud infrastructure than they are today, not less. The open-source model ecosystem will be thriving, but the inference bottleneck will remain.

As models get larger and inference gets more expensive, the concentration of compute power matters more, not less. The path toward AGI, whatever form it takes, is a path that requires massive compute. Whoever controls that compute controls the pace, direction, and access to the most powerful technology ever built.

Sweden's position in that future is not determined yet. But the window is closing. Every year that Swedish companies default to us-east-1, the switching costs get higher. The expertise in running European infrastructure erodes. The economic argument for "just use AWS" gets stronger because the ecosystem effects compound.

There is also a geopolitical dimension that Swedish founders do not think about enough. What happens if US-EU relations deteriorate? What happens if a future US administration decides to use cloud access as a trade weapon? What happens if CLOUD Act requests become routine rather than exceptional? These are not paranoid scenarios. They are scenarios that any serious risk analysis should include.

For companies doing blockchain development or building decentralized systems, the irony is even sharper. You are building trustless protocols on trusted American infrastructure. Your smart contract is immutable but your cloud provider's terms of service are not.

What To Actually Do About This

I am not going to tell you to migrate everything to a Nordic cloud provider tomorrow. That is not realistic, and some of those providers are not ready for serious production workloads. But there are concrete steps.

1. Build infrastructure-agnostic from day one

If you are starting a new project, use Kubernetes (k8s) seriously. Not as a resume item. As an actual portability layer. Combine it with Traefik for ingress and Ansible for provisioning. Make your infrastructure describable and portable. The cost of doing this upfront is low. The cost of doing it later is enormous.

2. Know your self-hosting options

The awesome-selfhosted repo on GitHub has over 290,000 stars for a reason. There are production-grade self-hosted alternatives to almost every SaaS tool you use. Workflow automation? n8n is real, it works, and you can run it on your own metal. Monitoring? Netdata is excellent and self-hostable. You do not need to self-host everything. But you should know what you could self-host if you had to.

3. Demand European inference options from your AI providers

When you evaluate AI APIs, ask where inference runs. Ask which jurisdiction governs your data. If the answer is "us-east-1" and the vendor cannot offer a European alternative, that should factor into your decision. Market pressure works. If enough European buyers demand European inference, providers will build it.

4. Audit your actual dependency map

Most CTOs I talk to underestimate their AWS dependency. It is not just EC2 and S3. It is Cognito for auth. It is SES for email. It is CloudWatch for monitoring. It is SNS for notifications. Each one is a thread in a web that gets harder to untangle every month. Map it. Know what you are actually dependent on. Then make conscious decisions about what you keep and what you make portable.

The Uncomfortable Truth

We at HEIMLANDR build AI solutions and custom SaaS development for clients across Europe. We think about this every day. Where does inference run? Where does data live? What jurisdiction applies? These are not abstract policy questions. They are engineering decisions with real consequences.

The uncomfortable truth is this: Sweden has everything it needs to be sovereign in AI infrastructure. Cheap energy. Cold climate. Great engineers. Stable institutions. Political stability. What we lack is the will to use those advantages for ourselves instead of renting them out to foreign companies.

Mistral building in Sweden is a compliment. It is also a mirror. They see the value of what we have more clearly than we do. They are building sovereign infrastructure on our soil for their benefit. The question is whether Swedish companies will keep paying rent to Virginia or start building something of their own.

The answer, right now, is rent. And that bothers me more than it probably should.

I am writing this from Jönköping. The air is cold. The energy is cheap. The fiber is fast. And my deploy target is Stockholm, not Virginia. It is a small thing. But small things compound.

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.