It is no longer just a technology play. When OpenAI knocks on the doors of Congress or the White House with a proposal to hand over a 5% stake, the message is clear: generative artificial intelligence is becoming, in every respect, a matter of state. The news, bouncing out of Washington in recent hours, marks the start of a phase in which LLMs are not just corporate tools but genuine geopolitical assets.

The decision by OpenAI to put a shareholding – direct or indirect, depending on the hypothesis – on the table raises one unavoidable question: what exactly does the U.S. government want? And, more importantly, what guarantees does it demand in return? It is not the first time a tech company has courted the administration, but the difference lies in the asset being exchanged. This is not an operating system or a social network, but models capable of processing sensitive information, drafting strategic documents, and assisting military or intelligence decisions.

In the backstage of tech politics, the term “alignment” now carries a double meaning: ethical and operational. On one hand there is the desire to govern risks (bias, disinformation, autonomous weapons); on the other, the need for privileged access to a technology that looks set to be as critical as the jet engine was to the 20th century. A 5% stake might appear symbolic, but in the corporate world it is often enough to secure a board seat or, at a minimum, stringent observation rights. And for those building data centers with hundreds of thousands of GPUs, the stakes are enormous.

The real issue, for AI-RADAR readers, is different. If artificial intelligence becomes an ever more powerful tool of national power, companies operating in regulated sectors – finance, healthcare, defense, critical infrastructure – must ask what margins of control they can retain. The option of on-premise deployment, already the subject of deep analysis on this publication, ceases to be merely a TCO or latency calculation and turns into a condition of operational independence. A self-hosted LLM, running on proprietary hardware under granular access policies, can offer guarantees that a cloud-based model, however transparent, cannot provide when a foreign government – or, under certain conditions, one’s own – is calling the shots.

It is no coincidence that many organizations have accelerated experimentation with serving frameworks such as vLLM, TGI, or Ollama, combined with quantization techniques to run models of tens of billions of parameters on machines with constrained VRAM. Data sovereignty is not an ideological flag; it is a contractual and, increasingly, regulatory requirement. GDPR in Europe has already blazed a trail; the AI Act will tighten the constraint. Against this backdrop, a move like OpenAI’s raises the alert level, because no one can rule out that the next compliance requirement might hinge on the ownership stake of a provider.

Whether OpenAI’s proposal will materialize in a formal agreement or a more nuanced understanding remains to be seen. But one point is already clear: the development of Large Language Models can no longer ignore the table in Washington. And for anyone designing tomorrow’s AI infrastructure, ignoring this convergence of technology and politics would be the greatest risk of all.