The Sovereign AI Balance Report: Energy Edition

For decades an energy company's customer relationship was a meter reading and a bill. Liberalised markets ended the quiet: customers compare, switch, and churn, and the suppliers growing today are the ones turning meter and account data into an experience, personalised tariffs, consumption insight, EV and solar propositions, service that resolves in one message. That data is not neutral telemetry. A half-hourly meter feed reveals when a household wakes, works, travels, and sleeps, which makes it personal data of real intimacy, and the entire basis of the growth strategy.

Underneath the retail contest, the operator remains critical national infrastructure: grid operations, trading, and the operational-technology layer carry NIS2 and resilience duties that have nothing to do with marketing. Energy's AI question is therefore double, an experience contest on personal data above, a resilience mandate below, and both point to the same answer on placement.

At the sovereign end sit two families. The customer stack: churn and retention models, tariff personalisation, consumption insight, and next-best-offer over meter and account data, personal, revealing, and the whole basis of growth. And the operational core: grid and asset optimisation, predictive maintenance, energy-trading models, and security analytics, where resilience rather than privacy is the driver. Together they land the split around sixty-five per cent sovereign.

At the frontier edge: campaign content and copy, public-data and market research, and internal drafting, where capability pays and neither a household nor the grid is exposed. Claude where the data is non-personal; sovereign inference the moment a customer or a substation is in the loop.

The customer case first: a supplier that sends meter-level life patterns to a third-country API has exported the most intimate dataset it holds, to power the very personalisation that was supposed to differentiate it. A relationship grown on rented infrastructure is a relationship someone else can observe, reprice, or interrupt. The operational case is the sector's own discipline: a critical workload mediated by an external model in another jurisdiction is a resilience risk dressed as a convenience, and NIS2 and the Critical Entities Resilience directive make controlling that dependency an obligation, not a preference.

At operator volumes, self-hosted inference is also, fully costed, the cheaper option, including the energy line an operator is uniquely placed to understand. The figures here are Rindogatan models; the resilience logic is the directive's, not ours.

Classify customer and meter-data workloads as sovereign-by-default alongside the operational-technology and trading core, and map both families against GDPR and NIS2/CER obligations. Stand up sovereign inference behind the retail platform and inside the operational perimeter before scaling any pilot.

Prove the customer case first, a churn or tariff-personalisation pilot shows up in the numbers within a quarter, keep the frontier to the non-personal edge, and govern placement so neither a household's data nor a critical dependency is quietly externalised.