The White House is preparing an event, expected within weeks, at which electric utilities, the companies that build and run data centres for Big Tech, and governors of the states hosting the biggest buildouts will be asked to promise that none of this will end up on household electricity bills. Three people familiar with the preparations confirmed the plan to The Next Web.

The move comes as the energy appetite of data centres built to train and serve Large Language Models grows at a pace that is straining power grids across entire regions. It is not simply a matter of installed capacity: the geographical concentration of clusters with thousands of GPUs—from NVIDIA H100 modules to the new Blackwell generation—is creating demand spikes that transmission and distribution infrastructure cannot absorb without massive investment.

The pledge the White House aims to extract has immediate political weight, but it exposes a fundamental conflict. Utilities, regulated by state commissions, will need to finance grid upgrades and new generation capacity: if costs cannot be passed through to households via general rate increases, they will have to be loaded onto large industrial customers—the data centre operators themselves—or absorbed through public incentives. Either way, someone pays.

For enterprises weighing whether to keep LLM deployment on-premise or move to the cloud, the question is far from academic. Energy is a primary TCO driver for continuous inference on self-hosted hardware: servers packing hundreds of gigabytes of VRAM and running around the clock can rack up annual electricity bills that, in many geographies, exceed the hardware amortisation. If regulatory pressure makes data centre electricity more expensive—because utilities offload upgrade costs onto industrial contracts—the economic comparison between cloud and on-premise will shift, not always in the cloud’s favour. Conversely, a company with on-premise infrastructure in a region with direct access to stably priced renewables could gain a competitive edge on heavy inference workloads precisely because it is insulated from the socialisation of network costs.

The presence of governors at the table is a telling detail. States such as Virginia, Ohio and Texas are seeing an explosion of interconnection requests for mega‑campuses from cloud operators, often in exchange for tax breaks and discounted land. The federal government is now asking those same states to shield residents from rate impacts, which could trigger a renegotiation of local agreements and, in turn, influence the siting decisions for future data centres. The geography of AI, already sensitive to latency and water availability for cooling, may add “regulated energy cost” as a decisive variable.

For those tracking on-premise deployment dynamics, AI-RADAR provides analytical frameworks to model TCO by incorporating exposure to energy costs and local tariff policies, a factor that is gaining weight relative to traditional metrics like memory bandwidth or available teraflops.

Structurally, the White House initiative signals that AI’s era of energy innocence is over. When an industry’s electricity consumption becomes the subject of a political compact between governments and corporations, efficiency ceases to be a mere technical goal and becomes a prerequisite for a social licence to operate. For anyone designing on-premise inference infrastructure, that means every watt saved counts not just on the balance sheet but also in the viability of the operational model before increasingly attentive regulators.