The most common concern about AI is about speed and jobs: how fast is it moving, and what disappears? Those are reasonable questions. But there is another one, harder and less discussed: the question is not primarily whether AI moves quickly or slowly, but for whom it moves quickly.
Call it Divergence. It is the trajectory in which frontier capability keeps rising — models get better at code, analysis, documents and tools — but access to the best becomes uneven. The most advanced models, the lowest usage costs, the best integrations and the most robust safety functions are controlled by a limited number of companies and cloud ecosystems. The capability exists. It is simply not available on equal terms.
This is what separates Divergence from the pure speed trajectory. When the concern is pace, speed is the main problem. In Divergence, control is the main problem. And the two need not look different on the surface. A country — or a company — can have very high AI adoption and still have little influence over the technology itself. It can gain strong productivity benefits while becoming increasingly dependent on external suppliers, pricing models, terms of use, interfaces and safety decisions made somewhere else. Use and sovereignty drift apart. That is the core of Divergence.
Inside the economy, the pattern looks the same. Large and digitally mature organizations build internal capability, negotiate with suppliers, secure data quality and redesign processes. Smaller actors more often lack specialists, time, capital and procurement competence — the ability to know what to buy and how it should be used. The statistics already show the gap: large companies use AI at far higher rates than small ones, and the most common reason among those that have abstained is lack of relevant expertise. In Divergence, that difference does not remain a transitional detail. It becomes a structural pattern.
The public sector can end up in the same position. Some agencies, regions and municipalities build advanced capability; others get stuck in pilot projects, uncertain procurement or shortages of technical and legal competence. The result is different service levels within the same public system. At that point, Divergence is no longer only about business competitiveness. It is about equal treatment — whether citizens receive different quality depending on which public authority happened to build capability in time.
Behind all of this lies a concentration that is easy to evidence. According to the OECD, AI firms in the United States attracted around 75 percent of global AI venture capital in 2025; the EU27 attracted 6 percent. Epoch AI reports that the United States holds roughly three quarters of global GPU cluster performance. For a small, open and digitally advanced country, it is entirely rational in the short term to import capability as a service — small countries rarely build the whole technology stack themselves. But dependency has consequences. Pricing, access, security, data protection, continuity and bargaining power are then shaped by markets and decisions one does not control. And one thing is worth saying plainly: a country can have data centers on its territory without thereby controlling model capability, data flows or value creation. The facility brings investment and jobs, but it does not solve the dependency.
This is where the EU often enters as an answer, and it is half an answer. Common rules can protect rights, competition and trust, and give a small country more bargaining power than it would have alone. But regulation is not capability. If the EU regulates without also building or securing capability, regulatory strength can coexist with strategic dependency. Europe can have the world's most advanced AI legislation and still depend on external models, clouds and chips. That is Divergence's most important European point: one can govern the conditions for a technology one cannot itself produce, but only up to a limit.
All of this changes the question one should ask about every proposal, every procurement, every initiative. It is not enough to ask whether something increases AI use — it almost always does, and it almost always sounds good. The questions that actually matter in Divergence are different: Who gets capability? Who controls the infrastructure? Which actors can participate, and which are shut out? Which workers get transition pathways? And how can coordination reduce dependency rather than merely regulate it? A proposal that increases use but deepens dependency and widens the gap between large and small can look like progress and still move in the wrong direction.
At the same time, it is important not to exaggerate. Divergence is not a catastrophe scenario in a simple sense. It can contain strong productivity, rapid innovation and real benefits across large parts of the economy. The problem is not that things go badly — the problem is distribution and control. And the answer is not technical self-sufficiency; the idea that Sweden or Europe should build the entire stack themselves is neither realistic nor necessary. The point is narrower and more manageable: the more socially critical AI capability becomes, the more important it is to know which parts are external, which can be built nationally or at European level, and which dependencies have to be handled politically rather than wished away.
The warning, then, is not to close oneself off. That would be to trade one problem for a larger one. The warning is that rapid adoption without dependency analysis makes the transition narrower, more uneven and harder to steer — and that this is easy to miss, because Divergence can look like success from the inside. High adoption, visible benefits, satisfied quarterly reports, while control quietly moves out. The question one must keep asking is therefore not how much AI we use. It is how much of it we actually own.
Sources and checkpoints
The sources below cover the central factual claims about capital, compute capacity and Swedish business adoption. The conclusion about dependency and control is the author's.
- OECD: Venture capital investments in artificial intelligence through 2025, 17 February 2026. OECD reports that AI firms in the United States accounted for about 75 percent of global AI VC deal value in 2025, while the EU27 accounted for about 6 percent.
- Epoch AI: The US hosts the majority of GPU cluster performance, followed by China, 5 June 2025. Epoch reports the United States at 74.5 percent and the EU at 4.8 percent in its GPU cluster dataset.
- Statistics Sweden: Artificiell intelligens i Sverige 2025. SCB reports AI use of 30.8 percent among small firms, 49.6 percent among medium-sized firms and 71.9 percent among large firms; lack of relevant expertise was the most common reason among firms that considered but did not adopt AI.
- Statistics Sweden: AI-användning i företag 2025. Deeper report on firm-size classes, barriers and the risk of persistent gaps.
- AI-skiftet: AI Future Scenarios 2026–2040. This essay develops the Divergence scenario: uneven access and control.