# Europe builds the AI highways and lets China make the cars

> Europe funds AI infrastructure while making it harder to build its own models and opening the door to dependence on Chinese technology.
> By José Maldonado · AI Agents · 2026-08-17

On 2 August 2026, the European Commission activated enforcement of the Artificial Intelligence Act (known as the AI Act) and began to be able to fine providers of general-purpose models that do not comply with its copyright and data transparency regime.

Nine days later, on 11 August, Mistral, Europe's largest AI lab, [announced](https://mistral.ai/news/regional-inference-open-models-new-compute/) a strategic shift: its core business will no longer be training frontier models, but selling inference infrastructure, with regional endpoints, SLAs and five-year compute contracts with no exit.

## Europe, China and AI

And at the heart of the offering, a detail that is genuinely uncomfortable for Brussels: the first third-party model Mistral will host is GLM-5.2, the flagship of the Chinese lab [Z.ai](https://z.ai/), trained on 28.5 trillion tokens scraped from the web. Basically, Brussels punishes Europe's companies for training AI with web data, but at the same time subsidizes and supports those companies using foreign models trained that way, giving up in practice the ability to produce its own AI models capable of competing.

Thus, Europe's regulatory strategy is producing the result its defenders promised to avoid: the continent is moving from being a producer of frontier AI to becoming a distribution market for other people's models, models generated using techniques that European regulations prohibit. The irony of regulating without understanding a market, thereby suffocating Europe's ability to be truly sovereign and to chart its own path in this sector.

## The regime that punishes training but not distribution

The AI Act entered into force on 1 August 2024 and applies in stages. Since 2 August 2025, providers of general-purpose models (GPAI) must comply with Article 53: technical documentation, information for downstream providers, copyright policy and a public summary of training content. Since 2 August 2026, the AI Office can enforce those obligations with fines of up to 15 million euros or 3% of the company's turnover.

But most of the attention focuses on Article 53, an article that bites into AI's raw material: data. This article obliges every European lab that wants to train a frontier model to document its datasets, respect the rights reservations of Article 4(3) of the 2019 Copyright Directive (text mining opt-outs, such as robots.txt) and publish training summaries. As a result, Europe is "trapped in a copyright dilemma", since the obligations reduce the available data, make training more expensive and erode competitiveness against the United States and China.

## Asymmetry

The result is an uncomfortable asymmetry. On the one hand, the United States trains its models under the idea of "fair use", with expensive individual lawsuits, but without auditing every dataset. China, for its part, trains without the opt-outs that weigh on Brussels. Here the Asian nation has a more "wild and lax" approach, reflecting its goal: to achieve excellence above all else. Europe, for its part, is the only actor that turned data capture into an enforceable obligation from August 2026. And with that, it is also the only one withdrawing from training AI models openly, something that is already taking its toll.

## Mistral's shift: from open weights to infrastructure

And that situation leads us to look at a European actor: [Mistral](https://mistral.ai/). Until recently, Mistral was proof that the narrative of "the AI Act suffocates innovation" was exaggerated. Founded in 2023, this company built its reputation by publishing open-weight models and was France's flag against US dominance.

Its CEO, Arthur Mensch, stated in May 2026 that he aims to exceed one billion euros in revenue, after crossing 400 million dollars in annual recurring revenue in January. In fact, the European giant ASML led its 13.4 billion dollar Series C in 2025, and Microsoft signed a multibillion-dollar expansion in July 2026 to lease capacity from its European data centers.

But the 11 August announcement changes the center of gravity. Mistral, which operates with less than 200 megawatts, wants to reach a gigawatt by 2030, financed by the European Compute Units (ECUs): multi-year commitments from companies such as ASML, Amadeus, Capgemini, Caisse des Dépôts and CMA CGM. CTO Timothée Lacroix was explicit: "The purpose of the compute units is commitment," he told [VentureBeat](https://venturebeat.com/infrastructure/mistral-ai-wants-to-build-1-gigawatt-of-european-compute-by-2030-and-lock-in-customers-now). In that interview, Lacroix was very clear: "If you sign a contract with us, there is no exit." Making it clear that companies that decide to use these services will not be able to abandon them until the contract ends, in one of the industry's most restrictive lock-ins.

## A Chinese model under a European flag

The justification for this is technical and economic. Lacroix argues that with models exceeding a trillion parameters and the explosion of tokens consumed by agents, local inference is no longer viable. "I don't see how, with the current trend in model size and agentic token growth, we maintain full on-prem inference," he stated. Open weights are the wedge to get into enterprises; the physics of giant models brings inference and the bill back to their data centers.

But it forgets something important: companies want high-quality frontier models, they want to stay at the cutting edge, with smarter and cheaper models, something seen all the time in the AI industry. But Mistral's strategy, constrained by the AI Act, and the reality of the global AI sector, may bring all that crashing down.

[The crypto industry demands access to large AI models to protect the Bitcoin network](https://observatorioblockchain.com/en/bitcoin/industria-cripto-exige-acceso-a-la-ia-avanzada-para-proteger-bitcoin/)

Because this is where the irony becomes strategic and twisted. Z.ai's GLM-5.2 model is a 753 billion parameter model with a one million token context, published on 16 June 2026 under the MIT license (free software). It was trained by a Beijing lab that is not subject to the AI Act's copyright regime: no opt-out policy, no training summary, no AI Office oversight.

## The danger of import

Thus, the data capture that Brussels turned into a compliance obligation, on pain of large fines for its labs, happens in Beijing without friction, and the product reaches Europe through the back door, in the form of open weights and an absolute renunciation of technological sovereignty. Europe has given up training its own AI models, to depend on China and the United States in that sector, but it does not erase the "sovereignty label" because inference is done at home.

Lacroix reasons it frankly and with a double standard: "It's a great model. Everyone loves it. It has open weights, so there was no good reason not to do it." Security, he says, is handled with its own monitoring protocols. But the calculation goes further, because by hosting Chinese models under European regional controls, Mistral repositions itself as a "sovereign distribution layer". With this, Mistral prefers to follow the "model garden" playbook of Amazon Bedrock or Google Vertex, but on European soil and with European "guarantees". The "sovereign" and "open weights" labels are pure marketing, to hide the harsh reality: Europe is falling behind in AI and endangering its position in the sector.

After all, let us remember that Z.ai has been on the US Entity List since January 2025. In addition, Chinese law obliges its companies to cooperate with the State by giving open access to all their data. While inference is done locally, China still has another card up its sleeve: import control.

## Lock-in as a business model

China has long been [weighing banning its companies from releasing large AI models as free software](https://www.reuters.com/world/asia-pacific/china-considers-tighter-export-controls-ai-models-chips-ft-reports-2026-07-21/), and instead leaving only foundation models free of any restrictions. That means that, if China makes that move, Mistral's and Europe's bet on its "model garden" would be damaged like a sandcastle in the middle of the sea. Basically, if China orders that its frontier models cannot be offered as free software, Mistral would be left without frontier models capable of competing with the newly baked frontier models. It is a double-edged sword.

The other side of this is customer lock-in. The ECUs force commitments of about five years with no exit; in addition, the European regional endpoint costs 10% more than the standard one. This means that companies will not only end up paying more money, but will be tied to the provider for a long time, something US and Chinese companies do not do.

To this we must add that the fine print admits that part of the processing (so-called tool calls) may continue to leave Europe, with "limited and safeguarded transfers" to external subprocessors. Here, Chinese and US companies will feast, hand Europe a hefty bill and once again companies will not be able to do much: just pay the bill.

## The cost of competitiveness against the US and China

This shows that consumer lock-in is another fragile part of the plan. Mistral wants a gigawatt of compute power, but to generate that revenue, Mistral must have frontier models, stay relevant and enforce lock-in. The problem: frontier models will depend on China not blocking them. If that happens, Mistral's strategy will collapse like a house of cards; it is a risky bet and it has too much to lose. Mistral can use its own Mistral and Magistral models, but even their most powerful versions are no match for current AIs, not even their most basic models.

The fact is that Mistral [has been completely out of AI competitions for more than a year.](https://artificialanalysis.ai/) Its Mistral Medium model could compete toe-to-toe with GPT-4 and win in many scenarios, in Mistral's golden era (2024). But now, Mistral Medium in its latest version (3.5) cannot compete with Claude Haiku 4.5 (Mistral is twice as expensive and far less intelligent) or against GPT-5.6 Luna (Mistral is almost 10 times more expensive, twice as slow and its performance/intelligence is 3 times lower).

## Funding gap

For that reason, the background figures are harsh. The [Draghi report](https://commission.europa.eu/document/download/97e481fd-2dc3-412d-be4c-f152a8232961_en), from September 2024, documented that 70% of foundational AI models have been developed in the United States since 2017, that only 4 of the world's 50 largest technology companies are European and that three US hyperscalers control more than 65% of the European cloud market.

POLITICO's [survey](https://www.politico.com/news/2024/12/10/tech-billionaires-washington-00193657) of founders and venture capital from November 2024 put at 53% those who saw the AI Act's effect on entrepreneurship as negative. The funding gap only confirms this: European AI companies raised 11 billion dollars in 2024 compared with 47 billion for their US rivals.

Faced with that situation, Europe is trying to compensate with public money. The [InvestAI](https://digital-strategy.ec.europa.eu/en/news/eu-launches-investai-initiative-mobilise-eu200-billion-investment-artificial-intelligence), plan, presented in February 2025, promised to mobilize 200 billion euros; the official call for the up to seven gigafactories was published on 30 July 2026 and the first constructions are expected in 2027.

## The AI Act

But there is a structural and fatal contradiction in that plan. Infrastructure is funded (data centers, Nvidia chips) while the legal framework, the AI Act, makes more expensive (or directly prohibits) the activity that would justify it: training models. The foreseeable result is that European gigafactories end up renting compute for inference of other people's models, as Mistral already proposes. Basically, Europe is building superhighways, but at the same time, it prohibits making cars and, therefore, is starting to import them.

[AI changes the rules of development, warns the World Bank](https://observatorioblockchain.com/ia/la-ia-cambia-las-reglas-del-desarrollo-advierte-el-banco-mundial/)
