A new scenario report from MIT researcher Michiel Bakker and six co-authors argues that Europe still has time to narrow its AI gap, but not much.
The warning in Europe 2031 is simple enough to understand and uncomfortable enough to matter. Europe is still talking about AI sovereignty as a policy objective, while the United States is treating compute, power, chips, and model training as an industrial race. That difference is now showing up in the numbers.
The report, released June 11 and covered by Semafor, estimates that Europe controls about 5% of global AI compute, compared with roughly 80% for the US. The EU's flagship InvestAI initiative, announced at the Paris AI Action Summit in February 2025, pledged €200 billion for artificial intelligence development across the bloc, including €20 billion for large AI gigafactories. That sounds large until it is placed next to the American buildout. US technology companies have committed hundreds of billions of dollars to AI infrastructure, with hyperscalers racing to secure data centers, chips, grid access, and long-term power contracts.
The problem is not that Europe has done nothing. It is that the money is too thinly spread, too slow to arrive, and too often presented as a political signal rather than an operating plan. Some of the EU's headline figure includes repackaged or mobilized funding over several years, not fresh capital that can be immediately converted into chip clusters and training capacity. For founders trying to build frontier models, that distinction is not academic. It decides whether they can train, deploy, and improve quickly enough to stay in the race.
The fictional narrative in Europe 2031 runs from January 2025 through March 2031, and it uses Mistral as a case study in what looks like progress but may still fall short. Mistral's €1.7 billion funding round, led in large part by ASML, was a major European technology moment. It made the Dutch chipmaking equipment giant Mistral's largest shareholder and gave Europe a cleaner story to tell about strategic AI capacity. Yet the comparison with US rivals remains severe. OpenAI and Anthropic have raised capital at a scale that makes even Europe's strongest AI startup look underpowered.
That is the heart of the report's argument. The structural problem for European AI companies is not talent or ambition; it is capital density. A frontier model company burning through compute, salaries, and inference costs at scale cannot be built through public-private initiatives stretched across years and 27 member states. The gap between what Mistral can raise in Paris and what OpenAI can raise in San Francisco is not just a fundraising problem. It is an infrastructure problem showing up as one.
Europe 2031 traces part of the original misread to DeepSeek. When the Chinese lab released its R1 model in January 2025 and showed competitive performance at a fraction of the reported training cost, many European policymakers treated it as evidence that efficiency could offset the compute gap. The report argues the opposite. DeepSeek did not prove that compute is optional. It showed that a non-US lab could compete when it had already built enough technical depth, engineering discipline, and infrastructure to make efficiency count.
The window that is open right now
The scenario closes with an epilogue set in 2034, looking back at choices made in June 2026. The authors argue that summer 2026 is the last practical window before the feedback loops between compute, model capability, customer revenue, and reinvestment become too powerful for Europe to catch from behind. That may sound dramatic, but the mechanism is familiar. Better infrastructure produces stronger models. Stronger models win more customers. More revenue funds more infrastructure. Once that loop is running at American scale, catching up becomes much harder.
For European founders, the implications are immediate. Building a foundation model company in Europe now means competing for compute at prices shaped by American hyperscalers, hiring engineers who can command US compensation packages, and convincing investors who increasingly treat European AI as strategically interesting but commercially secondary. The report is not an argument for founders to leave. It is an argument for European institutions to stop treating AI investment as a diplomatic gesture and start treating it more like defense procurement.
The broader market is already moving in that direction. Anthropic's valuation reached $965 billion in May, while OpenAI confidentially filed for an IPO in June. Those figures are not just signs of investor enthusiasm. They reflect a belief that the companies with the largest compute bases and deepest capital pools will have durable advantages in models, applications, and enterprise distribution.
Europe can still change the trajectory, but only if the response becomes more concentrated and more operational. Gigafactories matter only if they arrive fast enough, at the right scale, with dependable power and customers ready to use them. Funding announcements matter only if they turn into GPUs, data center capacity, and companies that can sell globally. The next few months will show whether Europe wants AI sovereignty as a slogan or as an industrial strategy.
Also read: MetaX seeks a Hong Kong listing after shocking Shanghai with a 700 percent debut • Apple bets its anti-flattery Siri AI will outperform ChatGPT's engagement play • Mistral AI at €20 billion is a bet on European sovereign AI coming into its own
