Open-Source AI Is Sovereignty Theater If You Don't Own the Silicon

Every week another headline tells me enterprises are "fleeing to open-source AI to protect proprietary data." I read that and laugh, then I get angry, because it's a lie most CTOs are telling themselves to sleep at night. You can run Llama or Mistral or DeepSeek all day long. If that model is executing on NVIDIA silicon sitting in IBM Cloud, or Oracle's freshly bought 50,000 AMD chips, or some "fully managed inference" wrapper from Gcore running NVIDIA Dynamo underneath, you have not achieved sovereignty. You have achieved a different vendor lock-in with better marketing.

This is the part nobody wants to say out loud in the boardroom: the model license was never the risk. The infrastructure was.

The Open-Weights Illusion

Open-source models solve a real problem: you can read the weights, you can audit the architecture, you can fine-tune without asking permission. That matters. But "open" stops mattering the moment your inference runs on infrastructure you don't control, priced by someone else, governed by someone else's terms of service, sitting under someone else's jurisdiction.

Look at what actually happened this week. IBM and Together AI signed a multi-year deal to scale open-source inference, on NVIDIA hardware, hosted on IBM Cloud. Read that again. The industry's flagship "open-source sovereignty" story is three layers of dependency stacked on top of each other. Open weights, rented chips, rented cloud. That's not sovereignty. That's a subscription with extra paperwork.

Oracle just bought 50,000 AI chips from AMD, according to Computer Sweden. Great for Oracle's balance sheet. Irrelevant to your sovereignty if you're renting time on those same chips through an Oracle Cloud contract with a termination clause you didn't read carefully enough. Gcore is bundling NVIDIA Dynamo as "fully managed" inference, per Aktiellt. Managed by whom? For whom? Under what data residency guarantee that survives a subpoena from a foreign court?

The entire AI stack right now, model layer to silicon layer, is consolidating around four or five American hyperscalers and their chip suppliers. Meanwhile every enterprise procurement deck in Europe has a slide about "EU cloud sovereignty" that means absolutely nothing if the actual FLOPs are executing in a US-controlled data center running on US export-controlled chips.

What Real Sovereignty Actually Requires

Sovereignty is not a license file. It's four things, and you need all four or you have none:

1. You own or lease dedicated hardware you control

Not shared tenancy. Not "your instance." Hardware where you decide the firmware, the network path, the physical location, and who has root.

2. The model runs entirely inside your perimeter

No API call leaves your network for inference. This is achievable today with open weights on self-hosted infrastructure. It is not achievable renting inference-as-a-service from a hyperscaler, no matter how "open" the model card says it is.

3. You control the update and patch cycle

If your inference provider pushes a change to their managed layer and your model behavior shifts overnight, you don't have sovereignty. You have a dependency you're calling something nicer.

4. Jurisdiction actually matches your legal exposure

If your data is processed by hardware inside US jurisdiction, US law applies regardless of where your company is headquartered. The CLOUD Act doesn't care that your servers say "EU region" on the invoice. This is the single most misunderstood fact in every "sovereign AI" pitch deck I've seen this year.

Sweden, Jönköping, and the Nordic Delusion

Here's the part that should embarrass us. Sweden loves to talk about digital sovereignty. We had the Schrems II ruling, we had years of GDPR debate about US cloud providers, we have a whole cottage industry of consultants selling "EU-compliant AI" workshops. And then most Swedish enterprises, including plenty in Jönköping's own industrial and logistics sector, still run their AI workloads through Azure OpenAI, AWS Bedrock, or now these new "open-source but hyperscaler-hosted" inference deals. We are performing sovereignty while renting the exact same compute stack as everyone else.

Compare this to what I see in smaller, more paranoid markets. Companies in the defense and critical infrastructure space in Finland and Norway are further ahead on this than most Swedish enterprises, because their threat model was never theoretical. When your neighbor is Russia, "we trust the cloud contract" isn't a real security posture. Sweden, frankly, has had it too easy for too long and it shows in how casually we outsource compute.

The EU's response so far, Gaia-X, the various sovereign cloud initiatives, is well-intentioned and mostly toothless. It hasn't produced infrastructure at the scale or price point that competes with NVIDIA-backed hyperscale deals. Until it does, "EU AI sovereignty" is a policy paper, not a deployment option. I'd love to be wrong about this. I'm not.

Where This Actually Goes

Here's my read on the next two to five years, and it's not comfortable for people who like clean narratives. Compute becomes the actual battleground, not models. Model weights are already commoditizing fast, open-source is catching frontier labs within months instead of years. What doesn't commoditize is who owns the GPUs, who owns the power contracts, who owns the fabs. That's where the real power concentrates. NVIDIA, AMD, TSMC, and the handful of hyperscalers who can secure allocation at scale. Everyone else, including most "sovereign AI" providers in Europe, is a reseller with a nicer flag on the homepage. As we get closer to AGI-level systems, this gets more dangerous, not less. If frontier capability requires massive compute clusters that only a handful of entities can build, then "open weights" becomes a distraction from the real chokepoint. You can have the smartest model on earth sitting on a hard drive and it means nothing if you can't run it without asking permission from whoever controls the chips it needs to think. Regulators are not ready for this. The EU AI Act focuses heavily on model risk classification and transparency requirements. It says almost nothing meaningful about compute concentration or infrastructure dependency as a sovereignty risk. That's the actual gap, and I haven't seen a single Brussels policy document treat it with the urgency it deserves. Someone is going to write that paper in 2027 and act shocked that it wasn't obvious in 2026.

What to Look At If You're Serious

If you actually want to move toward owned inference instead of rented sovereignty theater, here's where I'd start looking this week:

• awesome-selfhosted, a genuinely massive catalog of self-hostable software. Before you sign another managed inference contract, check whether the capability you need already has a self-hosted answer.
• Dify and n8n, both let you build agentic workflows and automation on top of models you control, deployable on your own VPC or bare metal, not locked to a hyperscaler's managed layer.
• Kubernetes, still the right foundation for running your own GPU clusters at scale if you're serious about owning the inference layer instead of renting it by the token.
• Netdata, because if you're going to run your own inference infrastructure, you need real observability, not a dashboard from the same vendor selling you the compute.

What You Should Actually Do

Stop asking "which open-source model should we adopt." Start asking "where does the inference actually execute, under whose jurisdiction, on whose hardware, and what happens if that vendor changes terms tomorrow." If you can't answer that clearly, you don't have a sovereignty strategy, you have a marketing slide. At HEIMLANDR we build this the unglamorous way: infrastructure first, model second. When we do AI solutions for clients who actually mean it when they say data control matters, we start with where the GPUs physically sit and who has root access, not which model card looks best in a pitch deck. Same logic applies when we build AI agents for enterprise workflows, the agent framework is irrelevant if the underlying inference is a black box you're renting by the API call. And if your team is still stuck deciding between "buy the SaaS" and "build it properly," that's exactly the conversation worth having before you sign another multi-year cloud commitment. Our custom SaaS development work exists because most off-the-shelf platforms bake in the exact dependency chains I'm describing here, quietly, in the fine print.

This isn't anti-cloud dogma. Sometimes renting compute is the right call, especially early. The problem is calling it sovereignty when it isn't. Say what it actually is: a vendor relationship with better documentation. Then decide, with clear eyes, whether that's a risk you can live with.

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.