“Don’t wait for artificial intelligence’s Hiroshima moment before writing the rules.” The analogy chosen by UK Foreign Secretary Yvette Cooper could not be starker. In an essay published for the Chatham House think tank and reported by Bloomberg, Cooper warns that AI could become “the greatest security challenge of the next decade.”

Her statement lands at a time when the race for generative AI – and ever more capable models – is raising not just ethical but also geopolitical questions. Without shared rules, the minister argues, we risk repeating the history of nuclear weapons: a disruptive innovation left ungoverned until it produces catastrophe. The call is for international coordination that is currently missing, as many governments prioritize national competitiveness over common standards.

Behind the diplomatic message lies an issue that directly affects those designing or managing AI infrastructure in enterprise environments. The fear of malicious use – from large-scale disinformation to autonomous attacks on critical networks – is pushing many organizations to rethink where and how they run their models. If the risk is global, control must be local.

It is no coincidence that the sovereignty debate has moved beyond privacy circles and into national security territory. Running a Large Language Model on a foreign cloud means exposing prompts, weights, and sensitive data to different legal jurisdictions. This is driving growing interest in on-premise or self-hosted architectures, where ownership of the entire pipeline – from fine-tuning to inference – stays in-house. Frameworks like vLLM or TensorRT-LLM, combined with hardware offering enough VRAM for advanced quantization (INT8 or FP16), now allow significant workloads to be handled while meeting compliance requirements.

Of course, the trade-off between control and cost remains central. Managing an on-site GPU cluster carries a TCO that not every organization can sustain, and the skills required to maintain an on-premise deployment are non-trivial. But when the alternative is dependency on cloud providers that can change terms or be subjected to regulatory blocks, investing in technical autonomy takes on strategic value. Cooper herself, without delving into technicalities, notes how the absence of common rules increases uncertainty for businesses.

The warning from London fits into a broader picture. The European Union has already launched the AI Act, and other jurisdictions are considering restrictions on AI chip exports. The direction is clear: AI will never be just another business. And those planning scalable architectures today are already factoring in the possibility of isolating them, quite literally, in air-gapped environments.