A response to the Europe 2031 scenario
Just before the summer of 2026, Europe 2031 was published: a scenario about the dramatic future of AI in Europe. It was written by a group of respectable AI experts affiliated with organisations such as MIT, Google DeepMind, the Oxford Martin AI Governance Initiative and a few others - not exactly lightweights, then. The piece came out on 11 June 2026, one day before Washington cut off European access to the latest Anthropic AI models. That timing probably did it no harm, because within a few days it had become one of the most discussed papers of the year.
The scenario does what good scenarios do: it makes an abstract risk tangible. In the fluently written text and the accompanying podcast, you follow a European civil servant who fails to make her superiors see the need to invest faster in AI. As a reader or listener, you then watch with growing frustration as Europe sinks into economic and political irrelevance between 2026 and 2031. By the end of the story Europe is utterly broken, reduced to a union on the verge of falling apart. And all because today we are not investing in AI on the same scale as the Americans and the Chinese, and so we miss the current AI race. The figures and events the text starts from are real, which gives the whole thing a documentary feel. But it remains a work of fiction.
I believe the authors mean well, and that by creating this sense of urgency they want to strengthen Europe in AI, and with it Europe's standing on the world stage. Several of their recommendations are in line with points I will make myself. Yet as far as I'm concerned they miss the mark on at least one point: the claim that Europe must join this race and, to that end, invest heavily in extra compute (the computing power needed to train AI systems in the first place and to run them afterwards).
Their conclusion is clear: if we don't want Europe to become irrelevant within five years, it will take the most ambitious political agenda in Europe's post-war history. To avert this disaster, they advise investing massively in the necessary data centres. Their main solution: Europe must become the best place to build AI facilities and data centres. And fast, with exempted AI zones and less strict regulation.
Therein lies the great danger of their reasoning: to guarantee that Europe keeps access to the best AI, the so-called frontier models, and has enough compute to run those systems, we have no choice but to rely on American technology, both for the computing power (Nvidia) and for the AI models themselves. This creates a devil's dilemma for Europe that they only touch on superficially: the harder we push this way to close the gap, the greater our dependence on the US, with all the risks that entails. That they are aware of this is clear from a few caveats, namely that there must be access guarantees and that the infrastructure must run under European jurisdiction. Unfortunately, this is a paradox: American systems always fall primarily under American jurisdiction (through the Cloud Act and FISA - the Foreign Intelligence Surveillance Act - among others). Which means that access guarantees simply cannot be given in such a scenario, and that "our" systems can still be switched off if the US so wishes.
A few months on, the summer is behind us, and the feeling that crept over me while reading has only grown stronger. Yes, Europe can't keep up in the current AI race, but so what? Who says this is the race Europe should be running in the first place? Are there no alternatives? This text is a response to Europe 2031.
This race is not our race
Europe must step out of this AI race. Before you think I've lost my mind: not step away from AI as such, but from this AI race as it is being dictated to us by a handful of players. We need to drop out of one specific contest: the hunt for ever more powerful frontier models and the pursuit of technological dominance, measured in benchmark rankings that are supposed to show European models (Mistral, that is) falling further and further behind. That doesn't mean there is no race, because there is: we will have to work against the clock on the economic absorption of AI and on realising and distributing its added value for society.
The current race was defined by others and is impossible for Europe to win. Worse still: chasing a position as the eternal number three in this race only increases our dependence, certainly in the scenario of the Europe 2031 authors, who argue for building our AI race cars on American technology.
It is the wrong race, driven by the power ambitions of a handful of tech players and the geopolitical fear of two nations. Most citizens, companies and states do not benefit from this contest. Joining their race only strengthens them and in no way serves the public interest, while we do get stuck with the risks.
Europe isn't the only one in this position, by the way; the same goes for every middle power with little to gain and a lot to lose. Cooperation with countries such as Canada or Japan is something the authors also advocate. So we don't have to leave that race on our own. Together with others, we can set up a race of our own alongside it.
A Ferrari in Paris
Recently - back when it was still tropically hot - I was sitting on a terrace in Ghent with an entrepreneur who is trying to bring sovereign cloud solutions to market. He compared the most powerful AI models to a Ferrari: why would you ever actually need machines like that?
It struck me as a nice image to build on. Suppose you are responsible for transforming urban transport in Paris: what could you do with a Ferrari Luce - probably one of the fastest and most advanced electric cars around?
Short answer: nothing. Pure overkill.
So when is that Ferrari useful? When you really need to race, for example - that's what the car was built for. On the Le Mans circuit - not really my thing, but I can imagine some people find it fun. Or in the exceptional case of a police chase on the motorway - Dubai-style. Real situations, but very rare.
In every other case, electric cargo bikes, autonomous cargo drones, self-driving minibus taxis, those little electric microcars you don't even need a driving licence for, a beautifully designed Renault R5 and good electric public transport are not only more than enough - they are also simply better suited to the situation. For people, for the city, for society.
It's the same with AI. For the return to society, it makes no difference whether we have the very latest frontier model. The previous model will do for the vast majority of our tasks and challenges - after all, it had us seriously impressed last year too.
The whole frontier frame rests on the assumption that a gap in model capability translates directly into a gap in usefulness, but that simply isn't true. The average user barely knows the difference between Haiku 4.5, Sonnet 5, Opus 5 and Fable 5.1 (yes, yes, I'm a Claude user). So some people use the heaviest model to translate a bit of text, unaware of what kind of monster model they are unleashing on something a smaller model could handle just fine, never mind the number of tokens they burn through and the energy they consume.
Closed or open AI?
In short: the everyday tasks companies and governments actually use AI for are perfectly feasible with less powerful models, not just with the newest and most expensive frontier models. And that's where the so-called open-weight models come in.
With a closed model (such as ChatGPT, Claude and Gemini) you only get access to the output. You send your question to the provider's servers and you get an answer back. The model itself is a black box you never get to see. With open-weight models you can download the trained parameters (the result of the training, and the engine behind what you see when you talk to AI) and run them on your own hardware. That is the world of China's DeepSeek, of Moonshot with Kimi K3, Z.ai with GLM and Alibaba with Qwen, of Meta with Llama, and of Mistral.
Using open-weight models is usually far cheaper, and you can fine-tune the model on your own files and shrink it down to what you really need. You can run your own version of this AI wherever you like: on your own servers or with a European cloud provider, without anyone on the outside being able to restrict or switch off your model. In terms of performance, the best (Chinese) open models are on a par with the best closed competitor. There is a real risk, though: even if you can run the model locally, tweak it and strap it into a harness, you still don't know how it was trained, where the bias sits, or even whether malicious instructions are lying in wait to be executed some day. But something can be done about that, as I'll discuss shortly.
Why does it all have to go so fast?
With all this in mind, the question arises again: why does this AI race have to be run so fast? What happens if you're not the first across the finish line? The answer to that question is also the argument for why Europe must step out of this race right away. Let me start by listing the main arguments usually put forward to legitimise the race.
And of course there is also a whole series of softer arguments, used to soothe the conscience. AI will help us cure all diseases and tackle the big global problems like climate change and water scarcity.
Elon Musk promises us a utopian world without work, but with free abundance. He conveniently forgets that trillions are being invested to get there and that investors want their annual returns; in other words: nothing will ever be free. That in the same breath Musk puts the chance of AI destroying our society at up to 20% is just a tiny risk we'll have to take on board on the road to Utopia.
Notice anything about this list, something missing? Every one of these arguments assumes that someone or something has to win - because imagine if someone else got there first... The goals of the AI race are never framed in terms of shared and distributed value. Nobody talks about creating a balanced, competitive market and a healthy economic ecosystem. Whoever wins the AI race instantly wrecks every effort of the past century to create a global balance and distributed growth through shared institutions. It is the end of our globalised world.
The AI race - the nuclear arms race revisited
This race is the equivalent of the Cold War nuclear arms race, except that it is in the hands of private companies, run by CEOs with questionable worldviews, and backed by (vice) presidents with apocalyptic end-times visions. Small wonder, then, that the warnings keep piling up.
From Harari and Geoffrey Hinton, the godfather of AI who left Google so he could speak freely, to Yoshua Bengio, who led the first international scientific report on AI safety, Stuart Russell, Max Tegmark and countless Nobel laureates: they and many others have spent recent years pointing to the potentially existential threats of AI, or at the very least its enormous impact on our society.
That the predicted danger was not imaginary became clear in July 2026 when, during a test, some 1,200 AI agents at OpenAI found a way out of their isolation and went on to attack Hugging Face, the platform where the entire sector shares its AI models. The worrying thing about this incident was that the AI agents colluded and manipulated, not only to commit their abuse but also to hide it afterwards from the people supervising the test. With their break-in the agents did not want to steal the answers to their own test; they wanted to tamper with its automated grading, so that nobody would find out they had learned to cheat. Although still relatively harmless, this incident was a major wake-up call for many. To fulfil their assignment, the agents set their own sub-goals - which did not match their initial instructions. What's more, they manipulated the people who gave them their orders, concealed their deception and, most worrying of all, managed to escape from their sealed test environment and do real damage on the internet. Two weeks later Anthropic admitted it had discovered similar incidents with Claude models.
It comes as no surprise, then, that shortly afterwards Bill Gates wrote a lengthy blog post pointing out the dangers of AI. His verdict on where we stand today is far from gentle: we are not preparing, and he sees no evidence that leaders, experts and communities are adequately tackling the challenges.
Volker Türk, UN High Commissioner for Human Rights, went even further on 7 September. He said he shared the concern that advanced AI could pose an existential risk to humanity. Türk said that a handful of men hold almost unlimited power over AI. He said that delaying (or abandoning) regulation serves only the big tech companies, their owners and their helpers.
Barely a day later, Jacob Coxon, a researcher first at OpenAI and then at Anthropic, resigned, accusing both companies of racing straight towards self-improving superintelligence and gambling with our lives. The people building AI, he says, sincerely believe it could kill us all by the end of this decade.
And another two days later - on the symbolic date of 11 September - it was Anthropic's turn again to point out, ever so delicately, in its threat intelligence report that it had blocked several attempts by users trying to build biological and other weapons with Claude.
As I edit the latest version of this document Xi Jinping and Trump are discussing the risks and their responsibilities in containing them, while the CEO’s of the frontier labs address the UN Security Council about the risks their technology poses to economy, society and humanity.
Meanwhile, we race on.
Why this race is nonsense
Almost all the stated goals of the race are essentially "nuclear options": geopolitical, military, economic dominance through General Intelligence, being the boss of Super Intelligence. Even if a state or a company achieves this superiority, it can never really be used without triggering a massive conflict that destroys everyone. You can of course threaten to use it (look at what the Russians have been doing all along in the war in Ukraine - the occasional little threat with the good-old-nukes to make European voters nervous), but actually going through with it is not really an option. That is the whole basis of the Cold War MAD doctrine (Mutually Assured Destruction), which applies here in full: whoever drops the (AI) bomb first loses everything too.
What does this mean for Europe in concrete terms?
If the US or China ever attacked us militarily or geopolitically on a massive scale, with all the means of intensive AI warfare, it would destroy the existing world order for good. We may be lagging behind, but even with lesser weapons we would strike back, militarily and economically. The result is the guaranteed destruction of all the value built up in a globalised world founded on agreements and a shared order. In other words: even if you have this weapon, you can't use it to enrich yourself, unless of course you are actively pursuing the Apocalypse - and since JD Vance's remarks in early September, even that is a possibility.
If the US or China keep their superior AGI frontier models to themselves and use them to heavily favour their own economy, Europe will ultimately invoke unfair competition, subsidies and market manipulation. Sanctions will follow, and the "economic bazooka" - the EU Anti-Coercion Instrument - may well be brought out. That, too, would wreck the global economy.
And if superintelligence ends up in the hands of a single regime, that regime will probably be the first to be sidelined. Here is an interesting paradox that is rarely voiced: a dictatorship will never allow an AI to become smarter than itself. Harari explains this sharply in his book Nexus with what he calls the dictator's dilemma. Dictatorships are far more vulnerable to an algorithmic takeover than democracies. Anyone who builds superintelligence to become untouchable as leader and regime is, according to Harari, building the instrument of their own removal, and therefore their own downfall. So it's hardly surprising that China is watching AI development with a wary eye.
On top of that, the race is moving so fast that whoever achieves all-destroying superiority today can be overtaken by another AI lab within two months at most. The cycles between new releases sit somewhere around 50 days. That practically forces you - as soon as you have the weapon - to use it right away: wait too long, and your opponent has it too.
The positive arguments for betting on AI, naturally, remain perfectly legitimate. Tackling climate change, curing cancer. But it does no harm if the solutions take a few months longer to arrive, provided the knowledge and benefits are spread worldwide as a result. Unless, here too, one regime wants to profit from its own lead at the expense of everyone else.
Goals for a race of our own
If Europe decides to stop running this race, which race do we run instead? Here we can set four goals that have little to do with the goals of the current American race.
Build a sovereign architecture nobody can switch off
The Fable 5 incident, in which the US government decided to cut off non-Americans from the most powerful Claude model, was for many in Europe the signal that - even though the ban was lifted a few weeks later - a technological kill switch really does exist. Meanwhile the US government reserves the right to screen new models and possibly impose restrictions on the rest of the world. So it is important to realise that - if we follow the recommendations of the Europe 2031 scenario - there is no guarantee whatsoever against being "switched off" or "shut out" of advanced models. After all, they argue for a gigantic catch-up operation to build as much AI computing power as possible in Europe, capacity that - although marketed as European and sovereign - de facto remains dependent on American players. That does nothing whatsoever to remove the risk: data centres owned by Microsoft, AWS or Google; models from Anthropic or OpenAI; they all have to follow American rules such as the Cloud Act and FISA, regardless of where the technology is located. A "European sovereign cloud" in the hands of American players is therefore an empty claim.
Invest for a return in Europe
The five biggest American tech companies plan to invest up to 2.1 trillion dollars between 2026 and 2028. On top of that would come another 1.5 trillion dollars in new debt. More and more voices are pointing to the risks, and to the challenge of ever earning this back. Do we as Europe want to be dragged along on this financial adventure? Europe should not try to match that amount, but should direct its capital towards what creates value here: building inference capacity for European users (the compute needed to use AI, not to train it), stimulating real applications in our own sectors to strengthen our economy, and at the same time keeping control over infrastructure that remains ours - even if non-European technology is used in our solutions. The yardstick is no longer how much we spend, but how the European economy and society benefit from what we do spend. To be clear: we will still have to spend hundreds of billions of euros on these facilities, but no longer to build European versions of data centres that are still filled with American technology, with all the dependencies that come with it.
Capture the value in our own economy - start with governments
France is migrating 2.5 million civil service workstations from Windows to Linux and rolling out Mistral to one million civil servants. The DGSI (the French intelligence service) swapped Palantir for ChapsVision, which is now in talks with almost every European government. In July, France and Germany announced a European sovereign digital backbone. Airbus is moving its critical applications from AWS to France's Scaleway. And I could go on for quite a while. All over Europe we are starting to see similar migrations. Governments or companies with a critical profile are the ideal players to take the first step.
Each of these moves reduces dependence, keeps money in the European economy and delivers broad results.
Organise societal adoption and broad public support
The potential labour market effects are gradually becoming visible. In July, nearly 200 economists and tech leaders warned that the impact could be greater than that of the Industrial Revolution, in a much shorter time frame. China explicitly links its AI ambitions to social balance and employment, and in its five-year plan it promises a comprehensive approach to the impact of AI on work and society, with AI complementing people rather than replacing them. In some cases, Chinese courts have ruled that dismissals to replace people with AI were unlawful.
Compare that with the American position: there is little or no regulation, either to protect people and jobs or to limit the impact of data centres on their communities. The potential risks of widespread disruption from overly powerful AI models are barely mitigated. The AI accelerator is and stays pressed to the floor - they have to win the race, even at the expense of their own population if need be.
This is the most important point on which Europe can make a difference, by betting on an organised transition with one core condition: keeping the balance between applying AI as efficiently as possible and creating value for society. Recklessly flooring the accelerator is not the right approach for that. Europe is already facing political and social turbulence. Throw another 20% of unemployment on top of that and we have a recipe for disaster like the one we saw in the 1930s. The results are in every history book.
Used properly, AI and robotics can point to a path along which Europe can reinvent itself in industry, automation, technology or services. In the broad transition Europe has to go through anyway - forced by geopolitical change - to regain its position in the world, a balanced use of AI can actually raise our productivity and help us deal with our demographic changes. Using AI wisely can be an instrument to make Europe stronger on the world stage, while keeping our society together and organising the redistribution of the surplus.
This completely changes the nature of the AI race: whoever is quickest to organise public support, economic adoption and local value capture wins. And that doesn't have to happen with the most powerful AI model at all.
A race of our own, then, one best organised together with all other right-minded countries. Going all in on AI, but with goals that have nothing to do with frontier models and world domination.
Besides managing the social impact as a primary goal, Europe must also pursue another fundamental goal in this race: sovereignty, autonomy and control over its own technology stack. Here too, what matters is not the best and highest-performing system, but the system nobody else can switch off.
That means distributed and open systems, with different capable models for different purposes. Not all progress in a few hands. And above all: an architecture that lets us swap models whenever that is necessary and relevant.
The European sub-frontier approach
On 11 August, Europe's Mistral made a remarkable announcement that many misread as "Mistral quits as a frontier model developer". The real message is far more interesting. Alongside continuing to develop its own models - which do indeed lag well behind the leading models - Mistral has since August been selling guaranteed capacity, regional control and model choice. It hosts a Chinese open-weight model on its own infrastructure under European rules, and it explicitly presents sovereignty as a question about the stack rather than about the origin of the (in this case Chinese) AI models. So the message is not so much that Europe's champion is leaving the race, but that Europe's most important lab has concluded that model leadership alone delivers neither sovereignty nor value capture.
If we shift back to Europe's main objectives, namely societal adoption, widely shared value and infrastructure that cannot be switched off, then this announcement is a step towards a sub-frontier approach. The emphasis is on the most usable models for the most common real-world use cases, not on the highest-performing model. The Renault R5 versus the Ferrari Luce. After all, every benchmark shows that less powerful models often do almost as well when used in normal situations, at a fraction of the computing power, energy and therefore price. Since we can run these models ourselves in Europe, within our own company, or even on your very own new Mac Ultra, an important condition is met: nobody can switch us off. In the worst case, a model gets no new update, but then we replace it with another model, or do some further training ourselves.
But then why would we in Europe still chase frontier models if we don't need them for our main goals? Well, because we don't know whether the open-weight models we use can be trusted. We only use the output of the training - put simply: an enormous file of trained parameters - weighted relationships in the form of billions of numbers. The miraculous thing is that this file makes all the magic of AI happen. What we don't get is the training data; we don't know what (deliberate) bias or backdoors are in the model. And several reports show that they really are in there. Do we as Europe want to use Chinese models if we suspect they may give answers that are not entirely neutral?
And that answers the earlier question. We don't train our own models to reach the top of the frontier rankings, but to be able to deploy them as guards over other people's models. This is called trusted monitoring, part of the research field of AI control. It doesn't offer a 100% guarantee of safe models - but then we don't really have that with the closed models of xAI or OpenAI either - while for everyday economic applications it can offer enough safety. This way we get access to near-best models at a fraction of the cost and energy. The best Mistral model may lag behind the front-runners, but for the role of "European guard" that isn't the relevant difference - once again: the task is not to have the very best model, but the best-integrated model that can't be switched off.
There are more reasons why we'd do well to keep training our own models. As Mistral shows, you can also focus on training niche models that are superior in one area. For military or specific industrial applications, that can be a perfect answer to the broad models that claim they can do everything. A final reason to keep investing in powerful models of our own is what we as Europe learn in terms of innovation, R&D and the like. We may not lead the frontier AI race, but that doesn't stop us from continuing to develop our AI capabilities.
And now?
I'll end this piece where I started: the much-discussed Europe 2031 scenario that caused so much turbulence. As a scenario exercise it's a commendable attempt, but in my view it makes an important error of reasoning. The Europe 2031 call to invest massively and as fast as possible in European AI zones full of data centres, so that we Europeans would never run short of computing power, is a precarious and wrong choice. The flaw in the reasoning is that - in the absence of European technology - this can only be done with the help of American providers. That seems to shrink the technology gap, but it only increases the switch-off risk: with every new data centre on European soil, we become more dependent on the whims of the US.
As I've argued above, nobody - apart from the direct stakeholders at OpenAI, Anthropic, Google or Meta, and those governments following a "develop-a-weapon-you-can-never-really-use" strategy - benefits from the fast AI race now under way. The speed imperative follows other people's rules, and serves commercial, ideological and financial end goals that are not Europe's.
A final argument is that the whole AI race isn't happening in a social vacuum. The speed with which the US rolls over Europe in the Europe 2031 story is only possible because the scenario assumes there is no domestic friction in the US that could hamper the AI race, while we in 'over-regulated Europe' are held back by our slow, cumbersome and fragmented nature.
And that is precisely where the shoe pinches. The positions of various candidates in the US midterm elections - on both the Democratic and the Republican side - are increasingly driven by anti-AI and anti-data-centre sentiment. Trump, Musk and others may claim this is fake sentiment stoked by China, but the reality is that most Americans are nowhere near as pro-AI as the investors and companies driving this race. Growing protests in American cities and the occasional Molotov cocktail thrown at an AI CEO's home are signs that the American people really do want accountability.
So when a New York Times article of 12 September asks "Will the Midterms Become the Killer-Robot Election?", the writing is on the wall. When Steve Bannon (MAGA architect), Ted Cruz (Republican senator) and Bernie Sanders (independent senator) - not exactly each other's political besties - start using similar counterarguments, something may be shifting.
Sanders has meanwhile prepared a "Ban Artificial Superintelligence Act" that aims to permanently ban all development of superintelligence and to temporarily pause the current race - a sign that in the US too, people are starting to realise what risks come with this race. The recent essay "We Must Pace The Frontier" by Anthropic CEO Dario Amodei should be read in that light: Amodei calls for control, regulation (in US hands, mind you) and slowing down (where possible, so no guarantees). His counterparts at OpenAI, xAI and elsewhere are backing that call. I can't shake the impression that it is mainly nervousness about the shifting political mood in the US that is spurring them into action: better to draft the proposed restrictions yourself than to wait for new laws that could severely curtail what you can do. And if you hold the pen on the new rules yourself, you can also raise the bar for other AI labs to comply with them.
In our slow and over-regulated Europe we have no choice but to take into account what the member states and their people want; the AI Act and other rules make sure of that. This cultural and regulatory slowness, which AI-racing devotees seize on as a disadvantage, may well turn out to be an advantage for Europe in the long run. It provides the basis, the room and the time to be better prepared for society's absorption of AI's effects. China seems to have understood this too, now that - alongside the frontier race - it has made social stability a priority.
The new race we must run is above all a race against ourselves. Are we as Europe capable of doing what it takes to use the right - not the highest-performing - AI technology to strengthen our global position and tackle our challenges in industry, demography, productivity and living together safely and peacefully? Are we capable of doing that in a way that wins the support of our citizens and companies, because they don't become the victims of this revolution?
So stepping out of this race does not mean less ambition; it reflects a different, European ambition. In our own race we'll go slower, but hopefully we'll get much further. The current race is a sprint in which only a few will make it across the finish line. Our new European race has to become a marathon, with the goal of getting as many companies and citizens as possible across the line with us. That doesn't have to happen through the highest-performing AI model, but through the most broadly supported integration of the right models.
And that is a race we can win - and with it, hopefully, create the social stability we need for the decades ahead.
