The Stargate UK project, a partnership between OpenAI, Nvidia and the British government, was launched as the country’s flagship AI infrastructure. But according to The Guardian last Friday, OpenAI apparently never set foot on the key site designated for the data centre before the programme was announced.

This is more than a quirky news item. In a sector where press releases often outpace due diligence, the lack of a physical site visit for a facility supposed to host high-density AI workloads – likely GPUs for LLM training and inference – raises serious questions. Choosing a data centre location is never just a real estate decision: it involves power supply, latency to end users, network resilience, and, for on-premise or hybrid deployments, direct data control. Relying on announcements without physical verification looks like the opposite of the rigour any organisation should demand when investing in critical infrastructure.

For those in the self-hosted world, the parallel is immediate. When a company decides to deploy servers equipped with NVIDIA H100s or the newer B200s to run LLMs locally, a site survey is not optional. Power capacity, cooling, physical security and connectivity must be assessed before signing contracts. The Stargate UK episode is a reminder that even big names can skip essential steps, risking foundations that may not hold.

Beyond the embarrassment, sovereignty remains the key issue. A state-backed project should embody maximum rigour in ensuring data and workloads remain under national jurisdiction, especially with AI. If the site selection phase was this fragile, how robust will daily operations be? The question applies not just to the UK, but to any government attaching its name to such initiatives without proper technical verification.

On the cost side, failing to evaluate a physical site can lead to TCO surprises. A seemingly cheap site can turn out unsuitable for power or cooling, forcing expensive retrofits or a change of direction after public money has been spent. AI-RADAR has previously covered how physical location directly affects the total cost of on-premise inference workloads – a point the Stargate UK actors appear to have overlooked.

For now, the story remains a news item. But it is a warning for anyone planning large-scale AI deployments. Site surveys rarely make headlines, but they are often the only thing standing between a solid infrastructure and a looming failure.