The pattern is hard to miss. The most capable AI models come from American labs, trained on American-built chips, funded by American money at a scale no European effort approaches. Europe shows up in the story mainly as the place that writes the rules — the AI Act, the GDPR, the Digital Markets Act. This is not an argument that Europe's approach is wrong. It is an accounting of why the gap exists, using the numbers, because the reasons are more structural than most commentary admits.
The money gap is not close
Stanford's AI Index Report puts US private AI investment at $109.1 billion for 2024 — nearly twelve times China's $9.3 billion and twenty-four times the UK's $4.5 billion. In generative AI specifically, US investment exceeded the combined total of China, the EU and the UK by $25.4 billion, a gap that widened from the year before.
Zoom out and it gets starker: from 2013 to 2024, US private AI investment totaled around $471 billion. Germany managed $13 billion and France $11 billion over the same period — combined, roughly a twentieth of the US figure.
The response is real but slow. The EU launched InvestAI in February 2025 to mobilize €200 billion, including a €20 billion fund for AI gigafactories, and France announced its own €109 billion national plan. Against that, the US Stargate project alone is a $500 billion commitment — and it is private money moving now, not public money moving through a tender process. The first European gigafactory is not expected to start construction until 2027.
Regulation that pushes products out
The AI Act entered into force on 1 August 2024. Its phase-in has been aggressive where it touches model builders: the banned-practices rules applied from February 2025, obligations for general-purpose AI models from August 2025, with fines up to €35 million or 7 percent of global turnover.
The measurable consequence came fast. In July 2024, Meta announced it would not release its multimodal Llama model in the EU at all, citing "the unpredictable nature of the European regulatory environment" — after regulators forced it to pause training on public Facebook and Instagram content from European users. Apple delayed Apple Intelligence in the EU the same season, citing the Digital Markets Act. Whatever one thinks of these complaints, the effect is concrete: European users got the worse version, and European developers building on those models had less to build with.
Even Brussels has started to blink. The Digital Omnibus, adopted in June 2026, pushed the AI Act's high-risk obligations back to December 2027 and August 2028. When the legislator defers its own deadlines, that is close to an admission that the original pace was a problem.
The chips problem under everything
Two decades ago Europe produced about a quarter of the world's semiconductor manufacturing capacity, and 19 percent of leading-edge capacity. Today those figures are roughly 8 percent and zero, according to a Kearney analysis cited by the EU Council — and a 2026 Commission staff document states plainly that the EU has no foundry capacity for leading-edge chips and is fully dependent on designs originating in the US.
The cruel irony is that Europe makes the most important machine in the supply chain: Dutch company ASML is the sole producer of the EUV lithography equipment every leading-edge fab needs. Europe sells the world the means of chip production and then buys the chips back from Taiwan.
Training frontier AI models runs on those leading-edge chips. No domestic supply, and every European lab queues behind everyone else — subject to export rules written in Washington.
Compute: arriving late, in small numbers
Europe's first exascale supercomputer, JUPITER at Forschungszentrum Jülich, was inaugurated in September 2025: about 24,000 Nvidia GH200 superchips, a €500 million investment, ranked fourth in the world. It is a genuine achievement and a useful machine.
It is also a useful measure of the gap. One European exascale system, publicly funded, took years. US companies stand up larger GPU clusters privately, repeatedly, in months, backed by that $109 billion a year — and Stargate-scale commitments promise several orders of magnitude more. When training the next frontier model is a compute auction, Europe is bringing a good computer to a bidding war.
What Europe has going for it
Honesty requires the other column. Regulation is also a brand: GDPR-shaped privacy expectations shaped products worldwide, and European firms like Proton turned compliance into a product advantage. ASML's monopoly means Europe cannot be bypassed entirely. Mistral AI in Paris keeps a frontier-class European lab alive. And European supercomputers lead the world in energy efficiency — a metric that will matter as power becomes the binding constraint.
What it would actually take
Closing the gap is not about writing better rules or worse ones. It is money and speed: venture capital at US scale, energy priced for data centers, gigafactories built in years not decades, and procurement that gives European labs customers before they need American ones. The €200 billion InvestAI pledge is the right size — the question the timeline answers is whether it arrives in time.
No comments yet