It is 2031, and Europe faces three miserable options. Become an American protectorate. Hand the future to China. Or slowly wither away in isolation. The only card it still holds is the Dutch company ASML, a manufacturer of machines without which the most advanced chips cannot be made. And Washington has just decided that it wants to control the company directly.
That is how Europe 2031 ends, a speculative story written by analysts in Brussels that became a hot topic in European circles within days. It was published just one day before the Trump administration blocked “foreign nationals” from accessing Anthropic’s acclaimed model called Fable. The scenario’s authors feel vindicated in what they predicted.
Three things Europe got wrong
The document goes back to 2025 and argues that Europe made three mistakes at the time. It misjudged how quickly artificial intelligence would advance. It misjudged how much it would change. And it overestimated its own ability to catch up.
When DeepSeek’s Chinese R1 model brought cutting-edge capabilities within reach at a fraction of the cost in early 2025, Europe drew a reassuring lesson from it: catching up would be cheap, and computing power did not really matter that much. But the opposite quickly proved true. Clever shortcuts and computing power reinforce each other. The more chips you have, the easier it is to discover those shortcuts. And DeepSeek’s own progress hit a ceiling precisely where China lacked imported chips.
AI skepticism and problems
At the Paris summit on artificial intelligence, the European Union announced a fund worth two hundred billion euros. It sounded impressive. In reality, it consisted mostly of repackaged money and hoped-for private investment. Compared with what the United States was spending on AI, it was a pittance. “Sovereignty” became a rallying cry. But the concrete measures had no teeth and avoided uncomfortable trade-offs.
Then GPT-5 arrived and disappointed. European skeptics took it as proof that no “AI bubble” would last long anyway, and momentum faltered. Yet something quite different was happening out of sight. Coding agents in Silicon Valley began automating the work of software engineers, while leading labs used their own models to build the next generation.
And there was another problem. Most European officials had no access at all to the most advanced systems; data protection rules barred them from using them. Almost none of them knew how to code. So those tasked with overseeing and reining in the technology often did not properly understand it themselves.
By mid-2026, some European leaders had reconsidered their skepticism. Anthropic’s Claude Mythos model, which was never released to the public, proved powerful enough to transform cyber defense. Europe was initially left out of the defense coalition that formed around it. Soon afterward, a US executive order pushed new cutting-edge models through classified vetting. Washington thus began choosing which “trusted partners” would receive them first.
And this is where the stark truth became clear. Europe controls just five percent of the world’s computing power for artificial intelligence. America controls eighty percent. It is hard to negotiate with a hand like that.
What may come: the years 2027 to 2031
Here, the scenario shifts from the past to what may yet happen. And it grows increasingly dark.
In 2027, a wave of ransomware breaks out, unleashed by an openly available model on the level of Mythos. Germany and France have just pushed through a law requiring important public systems to use only European artificial intelligence. But the organizations that had switched to European providers in advance were running defenses far behind the cutting edge. And those were precisely the ones the attackers locked down and held for ransom. The wave subsided only when both the United States and China banned openly available cutting-edge models. This left Europe even more dependent on closed American ones.
In 2028, artificial intelligence stops reasoning in language that humans can understand. The oversight tools that authorities relied on cease to work. And Washington pressures the Dutch to halt exports of older ASML machines to China. The other member states remain silent. The Netherlands gives in, and Europe negotiates absolutely nothing in return.
The year 2029 brings US country-by-country allocation of cutting-edge artificial intelligence. The shortage of computing power reaches a tipping point, and the United States introduces a tiered system that reserves most capacity for itself and a handful of selected allies. Most of Europe falls into the second tier, and its allocation from American cloud companies is cut in half. When the Union seeks to deploy its “trade bazooka,” the vote fails. European economic growth begins to fall sharply behind America’s.
And then comes 2030. The United States and China are locked in a struggle that both sides see as existential. To prevent a Chinese victory in robotics, a leading US company buys up struggling European carmakers and tool manufacturers. It wants their factories and industrial data. Car factories become robot factories. Unemployment rises, French debt enters a spiral, southern Europe follows, the euro comes under pressure, and the Union begins to crumble. Chinese credit lines appear across the continent, buying favor and seeking to pull Europe away from Washington.
Until finally, in 2031, the White House reaches for ASML. And Europe is left with the three miserable options with which this report began.
A vision being read by lawmakers, but with a catch
The scenario is narrated by Caroline Dubois, a Brussels official and a character invented for the occasion. She has a German friend, Christian Vogt, who has launched a startup in San Francisco. During a visit, she is astonished by America’s seventy- to eighty-hour workweeks and unsettled by the conviction among people there that everything is about to change. Back home, she then tries to convince her well-meaning bosses of what is coming. In vain. Skepticism is too strong, and most people think AI is a bubble.
Members of the European Parliament have read the story. According to the authors, it also came up in conversations between British and German officials at the beginning of that week.
But a skeptical reader will notice a catch. Some of the staggering sums and major projects the authors cite as evidence of America’s rise have already fallen apart in the meantime. The hundred-billion-dollar agreement between OpenAI and Nvidia, last year’s biggest AI deal, evaporated in February. The three-hundred-billion-dollar agreement between OpenAI and Oracle looks dubious, especially when the maker of ChatGPT is reportedly still losing billions. And the bulldozers that were supposed to be clearing land in Texas for a data center may no longer be clearing much of anything. OpenAI pulled out of that flagship project.
Maximilian Negele, one of the authors, takes it in stride. “I wouldn’t rule out that there is some exaggeration in this and that one or two AI companies will go bankrupt,” he says. “But we wanted to convey a sense of what we think will broadly happen.” He joined the project because of what he calls the “incredible translation barrier” between Brussels and San Francisco. He previously worked at the US-based Rand Institute and left this year so that he could focus on this.
Both he and his co-author Alex Petropolous acknowledge that there may be bumps along the way. One example is the growing resistance to data centers in the United States. “A lot of people hate artificial intelligence. People hate data centers. They destroy the landscape. They support big tech companies. It is incredibly unpopular policy,” Negele says.
Build data centers, or foot the bill for someone else’s infrastructure?
Given that opposition, the solution proposed by the authors sounds paradoxical: build more data centers, and build them faster. Ideally in special zones where matters involving energy and permitting can be simplified. “We think the overall supply of data centers is quite inelastic. So only a limited number are built around the world each year, and the question is how many of them you want in the US and how many in Europe,” Petropolous explains.
The authors themselves add further recommendations. Build a coalition of like-minded midsize countries, from Norway and Britain through Canada to Japan and South Korea, and use their positions in the supply chain as collective leverage. Reform the labor market along Danish lines so that companies can deploy artificial intelligence more deeply while protecting the people whose jobs it takes away. And focus on the areas where Europe may still succeed—robotics and industrial AI—because it has most likely already missed the boat on language models.
Not everyone accepts the alarmist tone without reservation. “This scenario, Europe 2031—I believe some parts of it could happen,” says Nicolás Casares, a Spanish member of the European Parliament. “But I think they are turning up the heat a little to get our attention.”
Casares adds a question that stayed with him after reading it. “What value is there for us in having OpenAI or Anthropic data centers in Europe? We are buying into the narrative that we need lots of data centers so that we do not lose the AI race. But that is crazy. We are building infrastructure that they will use, and sometimes they will not even allow us to use it.”
Incidentally, who is behind the Europe 2031 scenario? The main organization is the Brussels-based Arq Foundation, which says it is neither an advocacy nonprofit nor a venture-backed startup. It does not disclose who funds it.
Source: theguardian.com
