The 20% Club: Why Your AI Strategy Is Already Obsolete

Here is a number that should ruin your Monday. PwC says three quarters of the economic value created by AI is landing inside 20% of companies. Not 20% of industries. Not 20% of countries. Twenty percent of companies, full stop. The other 80% are still out there paying for pilots, running workshops, buying licenses, and generating the exact data and behavior patterns that make the winners' models better. You're not adopting AI. You're feeding it.

I've been saying versions of this since we started HEIMLANDR, but it's different when a Big Four firm says it with a straight face in a report CFOs actually read. The line between the companies compounding an unassailable lead and the companies subsidizing that lead is not about budget. It's about whether AI is infrastructure or decoration. Most of Sweden, and most of the Nordics, are still doing decoration.

AI Automation Business Models Have Split Into Two Species

The PwC framing that matters is the distinction between AI as growth infrastructure versus AI as cost-cutting. That's not a subtle difference. It's the difference between a company that restructures how value gets created and a company that uses AI to trim 8% off a support team's headcount. Cost-cutting AI is a one-time gain. You automate a process, you save money, the savings plateau, competitors catch up within eighteen months. Infrastructure AI compounds. Every customer interaction, every production run, every support ticket becomes training signal that makes your next product better than a competitor's. That gap doesn't close. It widens.

This is exactly why Anthropic just pulled in money from Blackstone, Goldman Sachs, and Hellman & Friedman to build enterprise AI services. That's not a startup raising a round. That's private equity buying a position in the infrastructure layer of the next decade of enterprise software, at a scale no Nordic firm, no Swedish bank, no EU sovereign fund is positioned to match. When capital like that moves, it's not betting on a product. It's betting on becoming unavoidable.

The Nordic Playbook Is Still Bolting AI Onto Legacy Ops

Meanwhile look at what passes for Nordic AI news this week. Circuit raises $30 million for "purpose-built manufacturing AI." Microsoft and Tieto announce a partnership. Both are fine, useful, probably profitable for the people involved. Neither is infrastructure. Both are AI bolted onto an existing operational shell, optimizing a process that already exists rather than asking what a manufacturing company looks like if you rebuild it AI-first from the floor up. I don't say this to be cruel to Circuit or Tieto. I say it because this is the pattern across the entire region. We are excellent at incremental improvement and mediocre at structural reinvention, and PwC's 20% club is explicitly built by companies doing the latter.

Sweden and the Nordics: Structurally Positioned to Donate, Not Benefit

Sweden's economy is SME-heavy. That's usually a strength. It's flexible, it's resilient in downturns, it produces companies like ours that can move fast without asking six layers of management for permission. But in an AI economy that rewards scale, data volume, and capital depth, SME-heavy is a structural liability. A manufacturer in Jönköping with 80 employees does not have the data volume to train anything meaningful in-house. It does not have the capital to build proprietary infrastructure. What it has is a vendor relationship with a platform, an OpenAI or Microsoft or Anthropic subscription, and every interaction with that platform is a small, permanent donation of pattern and behavior to a model owned by someone else, sitting in a data center that pays no tax in Sweden. Multiply that by the tens of thousands of SMEs across Sweden, Denmark, Norway, and Finland, and you get a region that is, in aggregate, one of the largest unpaid contributors to a handful of American AI labs' competitive moats. Dagens Industri covers the funding rounds and the partnership announcements every week. Nobody is running the aggregate number on what the region gives away versus what it captures. That number would be uncomfortable, and I think that's exactly why nobody in Rosenbad is asking for it.

Compare This to How Asia and the US Are Actually Positioned

In the US, the capital stack for AI infrastructure is now essentially closed to outsiders. Anthropic, OpenAI, Microsoft, Google, and a handful of private equity players are building vertically integrated moats: compute, model, distribution, enterprise sales, all owned end to end. A European SME buying into that stack is a customer, permanently, with no path to ownership of any layer. China is doing something structurally different. State-directed capital plus a domestic model ecosystem (Baidu, Alibaba, DeepSeek, Zhipu) means Chinese SMEs are, at minimum, feeding data to domestic infrastructure that their own government has leverage over. Whether or not you like that model politically, it means the donor relationship stays inside the country's own economic perimeter. Sweden and the EU have neither the vertical integration of the US nor the sovereign model ecosystem of China. The EU AI Act, for all its good intentions on safety and transparency, does nothing to address who owns the economic upside of AI adoption. It regulates risk. It does not touch the fact that European companies are training American infrastructure for free and calling it "adoption." I've read the EU AI Act framework closely, and there is no article in there about data sovereignty as an economic asset. That gap is the actual policy failure, and it's not getting fixed in this legislative cycle.

Where This Goes: The Path to AGI Makes the Gap Structural, Not Temporary

People want to believe the 20/80 split is a phase, that the gap closes as AI gets cheaper and more accessible. I don't believe that, and I'd tell you to stop believing it too. Cheaper models make the gap worse, not better, because they lower the floor for adoption without changing who owns the infrastructure layer. Everyone gets access to the same commodity model. The differentiation moves entirely to who has proprietary data, proprietary workflows, and the organizational will to restructure around AI rather than just use it. That differentiation is exactly what the 20% club already has and the 80% does not. As we move closer to genuinely general-purpose AI systems, that dynamic compounds harder. An AGI-adjacent model that can reason across a company's entire operational history is only as good as the operational history you feed it. Companies that have spent five years building infrastructure AI have five years of high quality, structured, proprietary signal. Companies that bought a chatbot license in 2024 and called it done have nothing comparable to feed a more powerful model in 2029. The gap PwC measured today is the floor, not the ceiling. For regulators, the honest question nobody in Brussels or Stockholm wants to ask out loud is whether "AI sovereignty" needs to mean something closer to what happened with energy security after 2022. If your economy's productivity gains are flowing structurally to infrastructure you don't own and can't tax properly, that's not a competitiveness problem anymore. It's a sovereignty problem. I don't expect Rosenbad to frame it that way this year. I expect it to take another PwC-style report and a worse number before anyone does.

What Founders and CEOs Should Actually Do About It

Enough diagnosis. Here is what I tell founders and CTOs who come through HEIMLANDR asking how to not end up in the 80%. First, stop treating AI as a subscription. If your AI strategy is "which vendor do we license," you are in the donor class by definition, because you own nothing and every interaction trains someone else's model. The move is AI agent development that runs on your own data, your own workflows, deployed in a way that compounds your advantage instead of a vendor's. That's the actual difference between infrastructure and decoration, in practice, not theory. Second, build for self-hosting and data control wherever you can. The trending tools on GitHub this week aren't accidental. Dify and Langflow both let you build agentic workflows without handing every byte of proprietary process data to a closed API. awesome-selfhosted is a good reminder that self-hosting is a legitimate strategic choice, not a hobbyist preference, especially if you're a manufacturer or logistics company whose operational data is your actual moat. And if you're evaluating what "agentic coding" even means for your engineering org, Claude Code is worth studying not because you should adopt it blindly but because it shows what a serious agent-native workflow looks like from the inside. Third, get honest about restructuring versus bolting on. If your AI initiative can be described as "we added a chatbot to X," it's cost-cutting, it will plateau, and PwC's data says you're in the 80%. If your AI initiative changes how a decision gets made, how a product gets built, or how a customer relationship compounds over time, you're building infrastructure. That's the whole test. It's brutally simple and most companies fail it because it requires admitting the current org chart is the problem. This is the work we do at HEIMLANDR when we build custom AI solutions and rapid MVPs for clients who refuse to be donor-class. It's not glamorous. It's rebuilding the actual plumbing of how a company operates so the data stays yours and the compounding advantage stays yours too.

What to Look At This Week

• Dify: build agentic workflows and RAG pipelines on your own infrastructure, cloud or self-hosted. The right starting point if you want to own your stack instead of renting someone else's.
• Claude Code: study how agent-native engineering actually works before you decide whether to build or buy your automation layer.
• PwC's report itself: read it with your own P&L open next to it. Ask which column your company sits in, honestly, not aspirationally.
• SVT Ekonomi: keep an eye on how Swedish coverage frames the Nordic AI investment story, and notice what questions never get asked about data ownership.

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.