Brussels' turning point

After years of delegating to American big tech, Europe is fed up. The idea of developing a “made in EU” artificial intelligence is no longer just a wish: it has become a political priority. The news is circulating among experts: Brussels is looking with interest at building a European LLM. This is not an isolated project, but a response to two converging pressures: dependence on foreign cloud infrastructure and the growing need to keep data on the continent, in line with GDPR.

Admitting that Europe could produce a top-tier model is a stretch. The computational resources needed to train a state-of-the-art LLM are enormous: we are talking about thousands of GPUs with hundreds of gigabytes of VRAM, staggering electricity consumption, and expertise that only a few global labs possess. But counterbalancing the enormity of the challenge, according to analysts, is an unexpected factor: Donald Trump.

The Trump factor and technological sovereignty

The former American president, with his unpredictable policies on trade, research, and technology controls, is pushing Europe to accelerate. Regulatory uncertainty in the United States, combined with geopolitical tensions, makes any supply chain based on cloud servers and AI services controlled by a handful of overseas companies fragile. The message is clear: better to invest today in a local stack than to find oneself tomorrow with access blocked or under onerous conditions. Trump, inadvertently, is acting as a catalyst for European digital autonomy.

In this scenario, technological sovereignty is not an ideological banner but a calculation of TCO and operational resilience. Those planning on-premise deployment know well that initial costs (CapEx) are high: purchasing servers equipped with high-performance GPUs, storage for datasets and models, infrastructure for fine-tuning and inference. However, in the long run, direct control over data can reduce the risk of lock-in and operational costs tied to public APIs.

The hardware hurdle: VRAM, consumption, and supply chain

Building a competitive European LLM means setting up a training run on a scale never attempted on the continent. The need for GPUs – typically models like NVIDIA H100s or future generations – collides with a tight supply chain and export restrictions. Moreover, video memory (VRAM) is the real bottleneck: to accelerate training, multi-GPU configurations with fast interconnects (NVLink or equivalent) are required. Even for on-premise inference of an already-trained model, VRAM sizing becomes critical: dozens of gigabytes are needed just to load an LLM in FP16, and pushing quantization (INT8, FP8) can bring benefits, but at the cost of quality.

Companies and research centers evaluating self-hosting must grapple with a trilemma: compute power, latency, and energy budget. There are no magic formulas; every choice affects TCO. Those currently relying on cloud services may avoid CapEx, but they pay in growing OpEx and less control over data residency. For public administrations and enterprises handling sensitive data, the trade-off is increasingly tilting toward on-premise.

The European path: between ambition and realism

Even if Europe fails to build an LLM that beats American models in the short term, the effort would have systemic ripples. It would trigger investments in local datacenters, enable open-source frameworks for model serving, and nourish an ecosystem of specialized startups. Moreover, political pressure on algorithm transparency and privacy could push companies to choose on-premise or hybrid solutions, where data never leaves the corporate perimeter.

AI-RADAR is observing this movement closely. For those evaluating on-premise deployment, known trade-offs exist: the initial hardware cost, managing training and inference pipelines, and the need for in-house expertise. But the stakes are data sovereignty and operational continuity, elements that in the new geopolitical landscape take on strategic value. Perhaps Europe will not build the next GPT, but it is laying the foundations for an AI infrastructure less exposed to the whims of the White House.