On 23 June 2016, 52% of British voters decided to leave the European Union. A decade later, the bill has arrived: a succession of seven prime ministers, a structurally weakened pound, and an economy that a study published Monday estimates is 6–8% smaller than it would have been had the country remained. These figures are more than a political story — they offer a powerful metaphor for anyone weighing the real cost of a sovereignty decision, including the shift from cloud architectures to on-premise stacks for AI workloads.

The illusion of total control

The Leave campaign promised command of borders, laws and trade. Reality delivered a permanent renegotiation of agreements, loss of privileged access to the single market, and compliance costs that eroded the industrial base. In the same vein, migrating inference and training infrastructure for Large Language Models from a hyperscale cloud to a private data centre is often framed as a net gain in control and confidentiality. But it hides parallel trade-offs: direct management of GPUs, networking, storage and updates introduces operational complexity that the cloud had absorbed. The British lesson suggests that the price of sovereignty is never just the hardware sticker.

Visible and invisible costs of the do-it-yourself route

The study puts the GDP gap at 6–8 percentage points: foregone foreign investment, lower productivity, trade barriers. Translated into on-premise terms, this mirrors the Total Cost of Ownership of an inference cluster. It is not enough to tally the cost of NVIDIA H100 or A100-80GB GPUs, servers and bandwidth. You must account for 24/7 maintenance, the need for in-house expertise to optimise quantization and serving pipelines, downtime from failures, and the rapid obsolescence of hardware that the cloud lets you swap with an API call. On top, the weakened pound reminds how an autonomy choice can hurt purchasing power toward global suppliers — anyone buying accelerators in dollars faces similar dynamics when exchange rates turn adverse.

Data sovereignty: a value, not a dogma

For banking, healthcare and defence, keeping data within national borders — and the models that process them — may be a regulatory obligation or a GDPR guarantee. Here sovereignty is not ideology but requirement. Yet the Brexit case shows that distance from an integrated ecosystem creates daily efficiency losses. A self-hosted infrastructure demands constant guard on security updates, compliance auditing and external integrations. AI-RADAR explores these junctions, offering frameworks to assess the full cost of on-premise deployment without falling into simplifications.

The next ten-year horizon

The UK is rewriting its trade and technology agreements, striving to regain ground with the AI Safety Summit and bilateral partnerships. Similarly, the evolution of specialised chips, the efficiency of serving frameworks like vLLM and distributed fine-tuning techniques are narrowing the performance gap between cloud and on-premise. But the underlying question remains: how much is an organisation willing to pay for technological independence? Brexit, with its missing 6–8% of GDP, provides a metric — not an answer — for those planning their next LLM deployment.