The Agent Tax: 2026's AI Gold Rush Builds Their Moat, Not Yours

You are paying an agent tax and you don't even know it yet

Right now, as you read this, 40% of enterprise applications are on track to embed task-specific AI agents by the end of 2026. That number comes from Belitsoft's latest enterprise forecast, and honestly, it feels conservative. Every SaaS vendor I talk to, from Stockholm to San Francisco, is bolting agents onto their product. The question nobody is asking loudly enough: whose infrastructure are those agents running on?

Google Cloud just dropped $750 million to accelerate partner agentic AI development. NVIDIA launched an open agent development platform. IBM is pitching watsonx agents to every enterprise with a heartbeat. These are not gifts. They are moats. And if you are a Swedish company building on these stacks without thinking very hard about what you are signing up for, you are repeating the exact cloud lock-in mistake of the 2010s. Except this time it is faster, deeper, and the strategic cost is an order of magnitude higher.

I run HEIMLANDR.IO from Jönköping, Sweden. We build AI agents and custom AI solutions for companies that want to own their stack. I am not a neutral observer here. But my bias does not make me wrong.

The math nobody wants to do

Let me walk you through the AI agent development cost that nobody puts in their pitch deck.

When your company embeds an AI agent from Google's Vertex AI agent platform, or builds on NVIDIA's NIM agent toolkit, you are making three commitments simultaneously. First, you pay the obvious compute bill. That is the number your CFO sees. Second, you accept a data gravity pull. Your operational data flows through their infrastructure, and moving it later becomes exponentially more expensive with every passing quarter. Third, and this is the one that kills you in three years, you accept their abstraction layer as your reality. Your team builds muscle memory, tooling, and workflows around their APIs, their model formats, their orchestration patterns.

By the time you realize you want to switch, you can't. Not without rebuilding. This is the agent tax.

In the 2010s, moving from AWS to Azure was painful but possible. It was mostly infrastructure. Compute, storage, networking. Portable concepts. AI agents are different. They embed decision-making logic, business rules, customer interaction patterns, and institutional knowledge into a vendor's framework. Extracting that is not migration. It is reconstruction.

What is actually happening in Sweden right now

Look at Scorett. Right now they are ripping out their entire tech stack to build a prompt-driven commerce layer. That is bold. I respect the ambition. But the question I want to ask their CTO is: prompt-driven on whose models? Orchestrated by whose agent framework? If the answer is "OpenAI plus Azure," they have not built a new tech stack. They have rented one.

The Swedish AI startup scene has energy. I see it. But when I look for credible domestic agentic infrastructure plays, I find almost nothing. Skellefteå has its rocket founders making headlines. Stockholm has the usual suspects raising rounds. Nobody is building the substrate. Nobody is building the agent orchestration layer, the model-serving infrastructure, the sovereign compute backbone that would let Nordic companies run AI automation for their business without a permanent dependency on Redmond or Mountain View.

Compare that to what is happening in China. Kimi K2.6's 300-agent swarm architecture is not just a research paper. It is a working system that represents an entirely different philosophy of agent coordination. China is building its own stack from silicon to swarm. Europe is building applications on top of American APIs and calling it innovation.

Sweden gets a lot right. We have strong engineering talent. World-class data infrastructure (thank you, Peter Mattsson era at Telia). A culture that respects privacy by default. GDPR built into our muscle memory. But none of that matters if every AI agent running a Swedish company's operations phones home to a US hyperscaler for every decision.

This is not just a technical problem. It is a sovereignty problem. And the Swedish government is not moving fast enough on it.

The regulatory gap is real and growing

The EU AI Act exists. Fine. It classifies risk levels, mandates transparency, sets compliance requirements. But it says almost nothing about infrastructure dependency. It treats AI agents as products to be regulated, not as conduits of strategic dependency to be governed.

Swedish regulators at IMY (the Swedish Authority for Privacy Protection) do good work on data protection. But they are fighting the last war. The next war is not about where your customer data is stored. It is about who controls the decision-making layer that sits on top of that data. If your AI agent runs on Google's infrastructure, uses Google's model, and follows Google's orchestration patterns, the fact that your data resides in a Swedish data center is a fig leaf.

I want to see the Swedish government, Vinnova, and the broader EU tech policy apparatus wake up to this. Not next year. Now. The window for building European agentic infrastructure is closing fast. Every month, thousands of European companies make integration decisions that lock them in for years.

What the open-source world is telling us

The good news is that the tools exist for companies that want to build with more independence. The open-source community is not sleeping.

Ollama now supports Kimi-K2.5, DeepSeek, Qwen, Gemma, and a growing roster of models you can run locally or on your own infrastructure. That is 171,000+ GitHub stars and growing. The community is voting with their keyboards.

n8n continues to mature as a self-hostable workflow automation platform with native AI capabilities. 187,000+ stars. It is not perfect, but it represents a real alternative to vendor-locked agent orchestration. You can build serious AI automation for your business on top of n8n without handing your operational logic to a hyperscaler.

Langflow is another one I keep coming back to. It is a visual builder for AI-powered agents and workflows with nearly 148,000 stars. Good for prototyping, increasingly viable for production.

And then there is AutoGPT, still going strong at 184,000 stars, providing the kind of accessible agent framework that lets smaller teams punch above their weight.

None of these are drop-in replacements for a fully managed Google or NVIDIA agent platform. But that is the point. The managed platform is the trap. The open-source stack is the escape route. It costs more up front in engineering time. It saves you everything in the long run.

Where this goes: 2027 to 2030

Here is what I see coming.

By 2027, the first wave of "agent remigration" projects will start. Companies that went all-in on a single vendor's agent platform in 2025-2026 will hit a wall. Model capabilities will shift. Pricing will change. A vendor will deprecate a key API. And these companies will discover that their agents are not portable. They will spend millions rebuilding what they thought they had already built. I have seen this movie before. It was called cloud repatriation, and it happened to half the companies that went all-in on AWS in 2014.

By 2028, the conversation around AGI will force a much harder reckoning. If general-purpose agents become viable, the company that controls the agent layer controls the business. Full stop. If your AGI-adjacent agent runs on someone else's platform, you do not have an AI strategy. You have a subscription. The path toward AGI makes infrastructure sovereignty not just smart but existential for any company that wants to exist independently in the 2030s.

By 2029, I expect the EU to finally pass something resembling a Digital Sovereignty Act that addresses AI infrastructure dependency directly. It will be too late for companies that locked in during 2025-2026. It will help the next generation. Maybe.

The smart play right now is to build your agent layer with portability as a first-class requirement. Not as an afterthought. Not as a "we'll deal with it later." As the foundation.

What to actually do about this

If you are a CEO or CTO reading this, here is what I would tell you over coffee.

One: Audit your agent dependencies today. Every AI agent your company uses or is building, map where it runs, whose model it calls, and what it would take to move it. If you cannot move it in under a quarter, you have a strategic risk on your books that nobody is accounting for.

Two: Invest in model-agnostic agent architecture. Build your agents to swap models. Use abstraction layers that let you switch between OpenAI, Anthropic, open-source models on Ollama, or whatever comes next. This costs maybe 15-20% more in initial development. It saves your company.

Three: Self-host what matters. Not everything. But your core decision-making agents, the ones that touch customer data, pricing logic, supply chain decisions. Those should run on infrastructure you control. At HEIMLANDR, our AI solutions work starts with this principle. We help companies build agents they actually own.

Four: Look at these tools seriously.

• Ollama for local model serving. Run frontier-capable models on your own hardware.
• n8n for self-hosted agent workflow orchestration. The Swiss Army knife of AI automation that you actually own.
• Langflow for visual agent building. Useful for rapid prototyping and increasingly for production deployment.

If your team is not already evaluating these, you are behind.

The uncomfortable truth from Jönköping

I sit here in Jönköping, not in San Francisco, not in Stockholm even. And from this vantage point, the pattern is painfully clear. Swedish companies, European companies, are sleepwalking into the deepest technology dependency in history. Not because they are stupid. Because the on-ramp is so smooth. Google makes it easy. NVIDIA makes it easy. That is the whole point. Easy adoption. Impossible extraction.

The AI agent development cost is not what you pay to build the agent. It is what you pay when you realize the agent is not yours.

We are at HEIMLANDR because we believe an AI development company in Europe should help companies build things they own. Not rent things that own them. That is not a sales pitch. That is a conviction. And if the next five years play out the way I think they will, it is the only position that makes strategic sense.

So. Are you building your moat, or are you building theirs?

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.