Siemens webinar puts AI racks at 230 kilowatts, recommending scenario-based planning

NVIDIA's remarks at a Siemens Grid Software webinar, recapped in sponsored content published by Utility Dive, put the change in per-rack terms. Traditional data center racks typically draw 15 to 25 kilowatts; current AI infrastructure can require about 230 kilowatts per rack; future systems could approach one megawatt apiece. Against the midpoint of the older band, today's AI rack is more than eleven times the draw, and the projected future rack fifty times it. The recap does not treat that arithmetic as incremental growth.

The unit matters because it travels. A rack figure converts to a site figure only after a developer commits to a rack count, and that commitment may not be settled when a utility is asked to plan for it: the same 230 kilowatts can describe a substation-scale problem or a manageable one. Large loads, the recap notes, may cluster geographically and seek service much faster than generation, transmission and distribution infrastructure can be planned, permitted and built. That demand is arriving amid transmission constraints, lengthy interconnection timelines and competition for grid capacity. The recap's own framing is that the next generation of data centers changes more than the quantity of electricity a utility must supply; it changes the speed, concentration and uncertainty of load growth.

The distinction the session draws between a data center and an AI factory explains where the density comes from. Traditional data centers hosted enterprise applications and delivered computing and storage services, while AI factories create value through model training and inference, so as the technology advances toward reasoning models and physical AI systems, computing and electricity requirements rise together. NVIDIA's observation, as the recap records it, is that energy has become part of the AI technology stack, which makes power availability a strategic factor in where AI infrastructure gets developed at all.

AI also appears in the session as the instrument rather than the load, automating studies, analyzing data and helping utilities plan faster. That dual role framed a discussion whose speakers came from Siemens, NVIDIA and Dominion Energy and whose title ran from interconnection queues to grid readiness. The recap offers no measured accuracy for the planning tools it describes, and no word on what any of it costs to deploy.

Deterministic forecasting weakens when the planner faces many possible futures, and the recap enumerates them: an AI facility may be delayed, resized, relocated or phased; its load profile may change as the technology evolves; renewable output, electrification, distributed resources and transmission constraints stack more variables on top. The recommendation that follows is to complement deterministic studies with scenario-based planning and a second method the available text never names, because the passage ends mid-word.

Coordination is where the arithmetic turns into a scheduling problem. Utilities, regulators, system operators, developers, technology providers and customers are asked to align earlier on project certainty, ramp schedules, location, reliability and potential load flexibility. Each of those items sits with a different party, and none of them controls the others' calendars, which is why the recap's emphasis lands on alignment rather than on a single fix. More infrastructure is necessary, the argument runs, but so is planning that is faster and more collaborative.

Dominion Energy supplies the only utility-side experience in the material, and it arrives compressed: its Northern Virginia work reinforced that integrating large loads is not solely a utility challenge, and that stakeholders must align timelines and assumptions early. The recap carries no Dominion load figures, no detail on the service requests its territory is handling and no account of what aligning early has cost or required. A single sentence from the utility behind the session's only concrete example leaves the operational questions open — how early is early, which assumptions have to match, who concedes when they do not.

Read the numbers with the positions behind them in view. Siemens Grid Software hosted the webinar, NVIDIA supplied the per-rack figures that define the load, Dominion is the utility that would serve it, and the write-up sits in sponsored content rather than on a news desk. The figures are NVIDIA's webinar remarks and no more. The recap does not say how they were measured, and no queue position, cost estimate or tariff accompanies them.

For anyone underwriting power-hungry assets, the useful part of the session is the set of variables it asks planners to hold open: project certainty, ramp schedules, location, reliability and potential load flexibility. Those read as diligence questions as much as forecasting ones, and each can move a site's megawatt figure in either direction, set by parties who do not answer to one another. A plan built on one rack count and one ramp rate has answered those questions by assumption, and the assumptions sit outside the planner's control.

The material leaves the money open. It says nothing about who pays for the generation and transmission a cluster of AI factories implies, how those costs would be shared across projects arriving on different schedules, or what a utility does with a facility whose density climbs after its service terms are set. The nearer test is whether scenario-based planning turns up in a utility's own filings rather than in a vendor's webinar, and whether the next set of per-rack figures arrives with a method attached.