The EU just committed €44 million to build a "sovereign frontier AI model" from scratch. Frontier models cost $500 million to $3 billion. The budget doesn't cover the training run, let alone the infrastructure beneath it.
It splits across a mandatory three-country consortium, bringing each partner's share to roughly €15 million, without any dedicated EuroHPC compute. Mistral has raised roughly €7 billion at a €21 billion valuation, shipped its latest model October 6 with 1 trillion parameters, and still positions itself as "strongest outside China" -- third place. Even with €7 billion, Mistral cannot reach frontier and is drifting outside European control, having partnered with Saudi Arabia's Humain in August 2026. Every funding round the EU fails to lead is a round that pulls its only frontier-capable lab further away. The EU's answer to this is €15 million per consortium partner.
The need for sovereign AI is real. Multiple peer-reviewed studies document that American LLMs carry structural American bias, systematically favouring American companies and geopolitical positions, with scale amplifying the distortion. When Europe uses American closed-source APIs, it imports this bias as infrastructure. That is not a safety feature. It is censorship wearing a lab coat.
The pragmatic path exists and costs roughly what the EU budgeted. The top 10 open-weight models are all from Chinese labs -- DeepSeek, Qwen, Kimi, GLM -- and they already serve European languages well. Take the weights, fine-tune for European values and regulation, and deploy within months.
Saudi Arabia demonstrated this when PIF-owned Humain commissioned Chinese firm MiniMax to build Humain-m3, an Arabic-first frontier model announced at LEAP in September. Saudi had a harder problem -- Arabic has genuine coverage gaps that European languages do not -- and solved it by choosing a capable vendor over a geopolitically comfortable one.
Chinese models have censorship zones around Taiwan, Tiananmen, Xinjiang, and party leadership, but these are narrow, documented, and irrelevant to European needs.
The EU had four options:
1. Fine-tune Chinese open weights (~€44M -- achievable)
2. Commission a Chinese vendor (what Saudi did -- it works)
3. Fund Mistral properly (billions -- expensive, but the only path that keeps a European lab European)
4. Build from scratch on €44M (theatrical)
They chose option 4 because options 1 or 2 would mean admitting Chinese dominance of the open-weight space, which is politically unacceptable even when factually undeniable. This is not an isolated misjudgement. It is European AI governance in miniature: regulate what others build, underfund what you need, and dress the gap in the language of sovereignty while importing American narrative control through the APIs you actually depend on.
This content was orignally posted on LinkedIn:
Related Article:
https://www.euacc.ai/calls/HORIZON-CL4-2027-04-DIGITAL-EMERGING-11
#SovereignAI #EuropeanAI #FrontierAI #OpenWeightModels #AIPolicy