It is not certain that superhuman AI will arrive in 2027. But it already looks certain that the AI race is not about chatbots, but about industrial, geopolitical and institutional power. For Central and Eastern Europe, and Hungary within it, the question is not whether we can beat Silicon Valley. We cannot. The question is whether we can find a battlefield of our own before others decide for us which systems the economy and public administration of the next decade will be embedded in.
It is worth starting with two texts that look at the same phenomenon from two opposite points.
1. What are AI 2027 and Europe 2031?
AI 2027 is a detailed, month-by-month scenario written by Daniel Kokotajlo and colleagues. It writes from the center: a fictional frontier lab, OpenBrain, builds ever more capable coding and research agents that accelerate its own development. From this unfolds a self-reinforcing loop (the AI R&D feedback loop), a US-China arms race, model theft, cyber capabilities and, in the background, alignment, that is, the question of whether the systems really want what their designers want. By the end of 2027, the story reaches the threshold of superintelligence.
Europe 2031 is its mirror image, seen from the periphery. Here Europe is the protagonist, and the picture is gloomy: technological lag, a shortage of compute, dependence on American infrastructure, and a Union that, unless it acts radically, will lose the ability to shape its own future. According to the text, the stakes are nothing less than whether there will be values we can defend and welfare systems we can finance.
The two texts are best treated not as predictions but as strategic stress tests. The question is not whether exactly this will happen, but which of their claims already hold up against data, and where speculation begins.
2. What is already visible?
The movement of money is not speculation. Private AI investment in the US reached 285.9 billion dollars in 2025, twenty-three times China's 12.4 billion, and more than three quarters of it came from a single state, California (Stanford AI Index 2026). On top of this comes hyperscaler investment: the combined capital expenditure announced by the big four (Amazon, Alphabet, Meta, Microsoft) for 2026 is between 630 and 725 billion dollars, compared with about 390-410 billion in 2025, which already counted as a record (Financial Times, CNBC). Lest this look like a castle in the air: NVIDIA's data center revenue grew from 15 billion to 115 billion dollars in two years. Demand is therefore driven by real purchases, not merely by presentations.
This has physical consequences. According to the International Energy Agency, global data center electricity consumption will roughly double between 2025 and 2030, from 485 to about 945 terawatt-hours, driven by AI (IEA, Energy and AI). The United States today accounts for 45% of global data center consumption. Compute and energy are therefore linked, and both are heavily centered on the US.
There is real movement with coding agents too. 84% of developers already use AI tools (Stack Overflow 2025), and according to Microsoft's CEO, 20-30% of the company's code is generated by AI. In the labor market, a concrete, measured effect is visible: according to an analysis by the Stanford Digital Economy Lab based on payroll data, employment of software developers aged 22-25 has fallen by about 20% from its peak at the end of 2022, while that of those over 30 has grown. A Harvard study of 62 million workers confirms this: at companies that adopt AI, junior employment falls by 9-10%, while senior employment barely changes.
Europe's response so far has been mainly regulation (the AI Act), with more determined infrastructure building only from 2025. More on that shortly. But first it is worth looking at where the scenarios' narrative turns from provable to assumed.
3. Where does speculation begin?
The boldest claim of AI 2027 is not the growth of compute but the self-improvement loop: that AI systems substantially accelerate their own research, and that this leads to an exponential explosion. Today there is no public, independent evidence for this. The labs talk about internal use, but claims of the kind "algorithmic progress one and a half to three times faster" cannot be verified from the outside. This is the speculative core of the scenario.
The 2027 superintelligence deadline is also aggressive, and not only skeptics say so. The authors themselves indicate that their uncertainty grows considerably after 2026. The sharpest counterargument comes from the Princeton researchers Arvind Narayanan and Sayash Kapoor in their thesis of AI as normal technology: the impact of technologies is set not by the pace of invention but by the pace of diffusion, and diffusion moves at human speed. A telling figure of theirs: although 40% of adults already use generative AI, this affects only 0.5-3.5% of working hours. There is a gap between capability and actual economic impact.
This gap was shown most starkly by a randomized trial. In METR's 2025 study, experienced developers finished their tasks 19% slower with AI tools, while they had expected a 24% speedup and afterwards still felt 20%. A benchmark, then, is not the same as real productivity, and even users' self-assessment is unreliable. For the sake of balance, though, it matters that according to METR's early-2026 update, a speedup is already visible in certain subsamples. The effect is neither zero nor uniform. It depends on context.
Europe 2031 exaggerates from the opposite direction: it is deterministic. The lag is real (70% of foundation models are American, and three hyperscalers dominate two thirds of the European cloud market), but marginalization is not an inevitable conclusion. The path is more open than the scenario suggests.
4. What does this mean for Europe?
Europe is in trouble if it only regulates and imports. But it is not lost if it builds capability in industrial, regulated, domain-specific AI adoption.
The turn has already begun. The 2024 Draghi report said what had long been knowable: the EU's productivity gap with the US is largely explained by the tech sector, and without investment the welfare model will become impossible to finance. The response to this is InvestAI: 20 billion euros in gigafactory grants, with a mobilization target of 200 billion euros. Here, however, disciplined reading is needed. The 200 billion is not committed capital but a hoped-for envelope multiplied by private capital. The EU's actual contribution is about 17% of the investment, and construction of the first gigafactory begins only in 2027. Mistral, Europe's best model company, has raised a total of roughly 2.8 billion euros so far, and in 2025 the chipmaking giant ASML joined as lead investor. This is strategically significant (the chip and AI supply chains meet), but in scale it matches a few weeks of spending by an American hyperscaler.
The sober European conclusion is therefore not that it must catch up with the US at the frontier. The Bruegel think tank poses exactly this dilemma: catch up with the frontier, or thrive below the frontier in the market for derived AI products and services? Europe's real capital is its industrial depth, its regulated sectors (healthcare, finance, energy, public administration) and the demand for reliability. Here it is not the biggest model that wins, but the one that is auditable, compliant on data protection, and built into existing processes. Energy reinforces this too: the IEA itself notes that grid and permitting bottlenecks, with power plant permitting taking 10+ years in both the US and the EU, slow down the most aggressive scenarios. The constraint, then, does not hit Europe alone, and that also means some room for maneuver.
5. What does this mean for Central and Eastern Europe?
The region will not win in frontier training. The numbers are clear: total venture capital investment in CEE in 2024 was about 2.3 billion euros, less than a hyperscaler spends in a few days, and 48% of successful scale-ups move their headquarters abroad, typically to the US (Dealroom/Vestbee). More than two thirds of European AI venture capital flows into ten large cities, and less than 1% lands outside the capitals. Building a frontier lab from here is not realistic, and it is not the task either.
The real opportunity lies in the second and third waves: when the models have to be built into real institutions, industries, data flows and decision systems. This layer is AI adoption engineering: building usable systems out of messy institutional data. Concretely: RAG and GraphRAG architectures, ontologies and taxonomies, evaluation frameworks, human-in-the-loop processes, and domain-specific vertical AI tools. The region's strengths (industrial and automotive AI, manufacturing, shared service centers, fintech, cybersecurity, govtech, healthtech, edtech) fit exactly here.
And there is a concrete Hungarian confirmation as well. According to Microsoft's 2025 adoption report, Hungary is among the 20 most advanced countries in the world in AI use, ahead of the US, Germany and Poland on some indicators. A consortium led by HUN-REN SZTAKI won the EU AI Factory and AI Antenna call, through which the country can become an intermediary hub in the European AI network. Roland Jakab, CEO of HUN-REN and president of the Hungarian AI Coalition, summed up the region's logic concisely: AI does not create value by itself; it benefits those who use it deliberately and well. This is not a consolation prize. This is the strategy.
6. What does this mean for a Hungarian professional?
In the next two to five years, the strong ones will be those who can combine domain knowledge, data engineering, cloud, AI workflows, a culture of measurement and governance into a single competence package. Not the person who can type the newest model into a chat window, but the one who knows which problem is worth solving with AI at all, and how to prove that the solution works.
In practice, this is a well-defined toolkit. At the foundation: SQL, Python and PySpark; on the backend, FastAPI and asynchronous processing; in operations, cloud, Docker, CI/CD. The AI layer builds on this: RAG and GraphRAG, ontology and taxonomy management, and (this is the most undervalued part) evaluation frameworks and human-in-the-loop design. All of this is worth something only if it is framed by product thinking and real domain expertise.
There is also a deeper, almost personal truth in this. The path of the region's most valuable people is often not the classic engineering track. Someone who has come to AI engineering from a social science or other domain background brings exactly the bilingualism this market needs: they understand the data, and they understand the context in which the data is created and in which it becomes a decision. The second wave does not belong to purely technical geniuses. It belongs to those who do the translation between the model and reality, and this translator's role will not be automated away any time soon.
7. Conclusion
The question is not whether Central and Eastern Europe can win the first AI wave. Probably not. Training frontier models, compute worth hundreds of billions, Californian capital density: this match has largely been decided, and not here.
The question is whether we recognize the second one in time: the wave in which the models have to be built into real institutions, industries, data flows and decision systems. There, it is not the biggest model that wins, but the system that is best embedded, most auditable and best understands the local context. That does not require a gigafactory, but AI-ready institutions, clean data architecture, domain knowledge and measurement discipline. The region, and Hungary, can build this. But only if it does not try to win the race it has already lost, but the one that is beginning now.