Three doors are open, but they’re likely closing soon
Beijing. On July 7, Reuters reported that China’s Ministry of Commerce has been meeting with Alibaba, ByteDance, and Z.ai about restricting overseas access to their most advanced models — closed and open versions both, possibly starting with future releases, with leaks treated as a national security offence. Six of OpenRouter’s top ten by token volume this month are Chinese, DeepSeek V4 Flash leading outright at 20.9 trillion tokens, and most ship open weights that European teams fine-tune and self-host for free. The models Beijing is discussing locking up are what the world’s inference already runs on.
Washington. In June, the US ordered that foreign nationals not have access to Anthropic’s most advanced models. Anthropic couldn’t verify nationality in real time, so it shut them off globally until controls were lifted. Access to American frontier models is now a policy decision, revocable on national security grounds, with Europe holding no vote.
The labs themselves. No ban needed. Meta ended the Llama line with closed-weight Muse Spark in April 2026, and Alibaba shipped Qwen 3.7 Max in May as API-only — the first Qwen flagship without downloadable weights. Europe was locked out of Llama 4 before the pivot anyway: its license denies usage rights to EU-domiciled companies, citing “regulatory uncertainties” around the AI Act. No government sanctioned Europe. A legal department decided the compliance risk wasn’t worth it.
That’s the mechanism to worry about. Europe doesn’t need to be targeted to be locked out. It gets locked out as a side effect of other people’s security policies and legal departments.
While forward-thinking European inference providers like Inceptron have been focused on openness and governance, that hasn’t been the wider EU strategy.
Why Europe can’t just build its way out
The standard answer is: invest. And Europe is investing, with real numbers attached. InvestAI, announced in February 2025: €200bn mobilized, €20bn earmarked for four to five “AI gigafactories” of 100,000+ chips each. The Commission received 77 expressions of interest for gigafactory sites across 16 member states — demand isn’t the problem. The EIF launched a €15bn fund of funds in March to unlock up to €80bn for European scale-ups. ASML put €1.3bn into Mistral. Bpifrance, the French state bank, sits in the lending consortium for Mistral’s datacenters. On paper, Europe looks like a continent that got the memo.
What investment can’t buy is deployment time.
Look at what a committed national response looks like. On July 6, South Korea announced at least $880bn for chip manufacturing, AI datacenters, and robotics. President Lee Jae Myung called it a matter of survival and announced it standing next to the leaders of Samsung and SK Hynix. One president, two chipmakers, one room, one decision.
Now the EU execution scoreboard, seventeen months after the InvestAI announcement: the January 2026 regulation that created the legal basis commits €4.12bn of actual EU money, capped at 17% of each facility’s capital expenditure. The formal call for proposals has been delayed twice and is now expected in summer 2026. Operational target: 2027–2028. From €200bn headline to €4.12bn legal commitment to a tender that still isn’t published — that’s the announcement-to-silicon gap, and it’s the whole story.
Meanwhile, the largest AI facilities actually under construction on European soil aren’t European. Microsoft and Nscale are scaling the Sines campus in Portugal from 12,600 Blackwell Ultra GPUs in early 2026 to more than 66,000 Rubin GPUs on a site permitted for 1.2GW. Europe’s fastest datacenters are American, serving American model developers.
The timeline math doesn’t improve from here. Compute is pre-bought years out — hardware prices are climbing fast enough that Apple and Microsoft raised device prices on component costs, and labs are designing their own chips to escape the queue. Grid connection in Europe adds years on top. A sovereign fab supply chain, even with ASML sitting in Veldhoven, is a seven-to-ten-year project. China, cut off from top Nvidia silicon, could fall back on a forced domestic bet: DeepSeek designed V4 top to bottom for domestic chips — inference runs on Huawei’s Ascend, and parts of training have moved there too. China had a fab industry to bet on. Europe doesn’t.
There’s a subtler lock too. Frontier models aren’t trained from scratch anymore; they’re distilled from bigger internal teachers. Google has said its Gemini Flash models are distilled from the larger Pro tier. Alibaba now runs the same split in the open: the Qwen flagship is API-only while the open-weight releases land a tier below it, and Meta’s Behemoth teacher never shipped at all. The teacher never ships. If you have neither the compute to train a teacher nor access to anyone else’s, you can’t bootstrap a competitive generalist. That’s how “a few years behind” hardens into “structurally out.”
What buying access actually costs
If Europe doesn’t build, it buys. The price of bought compute now includes foreign policy. The template exists: when Microsoft put $1.5bn into the UAE’s G42 in 2024, the deal required G42 to drop Huawei and ZTE and align with US-governed compute. Geopolitical alignment as a contract term. That’s the shape of the deal a compute-poor Europe negotiates toward by default: multi-year licensing on someone else’s terms, with someone else’s foreign policy attached.
Three moves that fit inside the window
None of this argues for surrender. It argues for sequencing by speed.
Fork the open weights now. Weights already downloaded can’t be revoked — Meta hasn’t pulled Llama 4 Scout and Maverick, and Beijing’s restrictions, if they come, will likely apply to future models. Every quarter of delay shrinks the set of frontier-class weights Europe can legally hold, fine-tune, and host on the compute it already has. This is the only move measured in months. The candidates pick themselves from the benchmark data: on Kilo bench, the only open models in the top ten are Kimi K2.7 Code, Kimi K2.6, and GLM 5.2 — near-SOTA performance, weights downloadable today under MIT-style licenses. Add DeepSeek V4 and MiniMax M3 from the volume leaders and you have the shortlist. Localize them, host them in EU datacenters, build on them. It’s not glamorous. It works today.
Fund Mistral like infrastructure, not like a startup. Mistral is the one European lab in the frontier conversation, and it’s already acting like an infrastructure company: €722M in debt financing (its first) for Nvidia-powered datacenters, a Paris facility live around June 2026, a €1.2bn Swedish site targeting 2027, 200MW planned. ASML led its €1.7bn Series C with a €1.3bn investment and a seat on Mistral’s strategic committee — Europe’s two most strategic AI assets are already on the same cap table. ARR crossed $400M in early 2026. Arthur Mensch told the French National Assembly in May that Europe has a short window to avoid deeper dependence on American infrastructure. The honest caveat: Mistral still distributes through Azure, Google Cloud, and AWS. Even the sovereignty champion runs on the hyperscalers. That’s an argument for scaling it faster, not for pretending the dependency isn’t there.
Own the harness, since you can’t own the models. There are three layers to this problem. The model layer, Europe doesn’t control — one frontier lab, no teacher models. The infrastructure layer, Europe can’t control fast — 2028 at the earliest, on the current tender schedule. The orchestration layer is the one Europe can control completely, today, at near-zero cost.
Concretely: build on harnesses that treat every model as a swappable component. On a model-agnostic stack, the same news is a config change. Model access is now a policy variable that flips quarterly. The rational architecture is one where no flip is a rebuild — open-source orchestration, model-agnostic evals, inference layers that swap Mistral for a forked GLM for whatever ships next without touching the application.
This also rescues the specialization strategy. Europe probably can’t train a frontier generalist, but a strong harness routing between smaller specialized models is a different competition — one where the WEF argued in January that Europe can still win, pointing at physical and industrial AI: €2.5 trillion in manufacturing value added, 219 robots per 10,000 workers, among the highest automation densities in the world. Specialization plus orchestration is a strategy. Specialization alone is a consolation prize.
Europe was locked out of Llama 4 by a legal department — no sanction, no vote, no warning. Beijing’s restrictions, if they land, will work the same way: a policy meeting in one capital, a pulled download link everywhere else. The gigafactories can wait until 2028. The forks can’t.
Good piece.
I agree that open weights alone are not sovereignty. But I wonder whether Europe’s realistic path is to treat them as a learning substrate rather than a final destination.
This could be a form of learning by building, an industrial extension of learning by doing. Europe would not acquire AI capability mainly by observing the frontier or importing its outputs, but by building with what is already available.
By creating derived models through distillation, continued pretraining, domain adaptation, evaluation, and deployment, European teams would accumulate the capabilities that matter: data curation, training operations, inference, safety work, and production feedback.
The point would not be to claim independence merely because a foreign model runs locally. It would be to turn every deployment into reusable technical capacity, including better datasets, benchmarks, tooling, and teams able to move progressively closer to training their own models.
Perhaps Europe’s strategy could therefore be framed as learning by building, using open weights as a bridge toward capability rather than accepting them as a permanent substitute for it.
I loved seeing this, was beginning to think the selections I made opening this app were just performative...
Interesting piece, who would have thought a time would come where there is a possibility of country specific monopolies on frontier AI
But this isn't really an EU-specific problem, it's structural for any country without its own frontier lab. The US move on Anthropic access and the reported Beijing restrictions are two data points in the same pattern, the second a country builds a genuine frontier model, access to it becomes a policy lever, not a market good.