Equal kWh, Different Peak Demand: Read the AI Facility Electricity Quantity

Energy adds across time; interval demand selects a maximum

EIA distinguishes power in watts or kilowatts from energy accumulated in watt-hours or kilowatt-hours. An average power of 60 kW over one hour corresponds to 60 kWh. That energy total does not reveal whether the power was steady or concentrated within part of the hour.

For an interval-average measurement, divide the energy accumulated during the interval by its duration in hours. If a fifteen-minute interval contains 15 kWh, its average is 15 kWh divided by 0.25 hours, or 60 kW. The result remains a power quantity; it is neither 60 kWh nor 60 kW per hour.

Laurens Electric Cooperative describes charging applicable customers for the highest fifteen-minute average recorded on the demand meter. Its explanation also names fifteen- or thirty-minute demand periods. That is a provider example of the mechanism, rather than a universal interval or a tariff assigned to an AI facility.

References: [E1][U1]

Four fixed windows produce equal energy and different maxima

Consider two wholly fictional alternatives at the same meter over the same hour. Use four consecutive, fixed fifteen-minute windows with identical boundaries. Every entry below is an interval-average kW value, rather than an instantaneous sample or a GPU setting.

Profile A averages 60 kW in each window: 60, 60, 60 and 60 kW. Each window contributes 60 kW × 0.25 hours = 15 kWh. The hour therefore accumulates 15 + 15 + 15 + 15 = 60 kWh, and its largest interval average is 60 kW.

Profile B averages 120 kW in the first two windows and 0 kW in the last two: 120, 120, 0 and 0 kW. Its window energies are 30, 30, 0 and 0 kWh, giving the same 60 kWh total. Its largest interval average is 120 kW. Both profiles also have the same whole-hour average of 60 kW; averaging the entire hour hides their different fifteen-minute maxima.

For the billing comparison that follows, stipulate that all other billing-period intervals are identical between the alternatives and average no more than 60 kW. The remaining energy is therefore equal, and no other interval exceeds either alternative's stated maximum. The billing-period maxima are consequently 60 kW for A and 120 kW for B.

These invented profiles establish neither equal completed AI work nor actual GPU or facility behavior. The zero entries serve the arithmetic; they do not describe a real hall idling at zero.

Apply the two quantities to a deliberately simple charge structure

Assume the same flat energy rate, r_e, expressed as an amount per kWh, and the same positive demand rate, r_d, expressed as an amount per kW for the billing period. Assume this demand term applies directly to the maximum interval-average quantities just established. These are hypothetical terms without market prices.

Let E be the common billing-period energy quantity: the example hour's 60 kWh plus the identical energy from all remaining intervals. The energy component is E × r_e for either profile. The demand component is (60 kW) × r_d for A and (120 kW) × r_d for B.

The difference between these two-component amounts is therefore (60 kW) × r_d, which is positive. Equal energy quantities produce equal energy components under the stipulated flat rate, while the selected demand quantities differ. B's demand component is twice A's; that does not make its entire bill twice as large.

This calculation isolates energy and demand components. It supplies neither a complete bill nor evidence that changing a real operating profile preserves workload output, availability or any other project requirement.

A higher interval elsewhere can make this hour irrelevant to the maximum

Change the example's assumption about the rest of the billing period. Suppose an unchanged interval elsewhere has average demand H, where H exceeds 120 kW. Under the same maximum-based rule, both alternatives now have billing-period maximum H. Replacing B's hour with A's hour leaves that demand quantity unchanged.

A maximum depends on the highest qualifying interval across the relevant period, rather than on each hour independently. Lowering a local peak that does not determine that maximum cannot reduce this particular demand term. The four-window arithmetic alone therefore cannot establish a billing saving.

The billing rule defines which peak matters

DOE FEMP distinguishes several demand bases. Non-coincident demand uses the customer's highest monthly demand regardless of time of day. Time-of-use demand considers specified hours and may have separate windows. A look-back ratchet also considers earlier months' peaks. A system-peak basis concerns the customer's contribution during the utility's peak, which need not coincide with the customer's own maximum.

Those rules can change which intervals or historical values determine a charge. Different energy rates by time or season can also break the example's equality of energy components. Fixed charges and minimum requirements add further conditions. Fifteen minutes is an explicit assumption here; the actual interval, eligible windows, billing period and any floor must come from the applicable provider terms.

References: [F1]

A GPU transient is a different observation from a facility billing peak

NVIDIA's nvidia-smi documentation read on October 20, 2026 defines GPU Average Power Draw as a board-level average over the preceding second on supported devices. Its Instantaneous Power Draw is the last measured board reading. Those fields describe a different boundary and time basis from a facility's interval-average demand.

A short excursion contributes energy during the time it persists. Its highest sample does not reveal the energy accumulated throughout a complete billing interval. Calling that sample a peak therefore does not establish the applicable interval average, still less the maximum across the billing period.

The boundary matters just as much as the averaging. If cooling equipment, conversion losses, network equipment or other tenants fall within the facility meter's boundary, their contributions belong in that boundary's total too. A GPU report leaves those contributions unidentified. Separate component peaks may also occur at different times; adding their individual maxima does not reconstruct a common facility interval.

A proposed cap describes a control intention, while a demand record describes an observation under a billing rule. The companion power-cap article addresses control authority and server-input boundaries. Even a supported system limit would not, by itself, establish this meter's billed demand or the customer's charge.

References: [N1]

Read kWh, reserved kW and metered demand as different purchase quantities

The unit is necessary but insufficient. A line priced per kWh concerns an energy quantity. A line expressed in kW might instead concern a reservation or a named measured-demand basis. The word power cannot decide which one the offer uses.

If a hypothetical offer prices reserved capacity R in kW, R is a different input from the measured maximum D used above. Reducing D does not change R by arithmetic alone. Conversely, knowing R does not reconstruct measured kWh. Keep the actual offer's definition attached to each quantity before comparing amounts.

For colocation, distinguish the facility's utility billing boundary from the quantity attributed to the customer. A whole-facility maximum does not assign a tenant's contribution or explain its charge without the provider's applicable allocation method. Establish whether energy or demand is already included before adding a separate component.

Actual terms and customer attribution belong with the provider and the appointed energy and commercial team. The useful decision is which stated quantity the buyer is paying for, and whether the evidence actually describes that quantity.

Use confirmed facility inputs to scope the physical installation

Once the customer and facility have resolved the offered capacity and responsibilities, the physical work can be defined from the selected equipment, rack positions, customer-side cable endpoints and access dependencies. The related capacity and metering guides address those adjacent planning questions.

Leviathan's colocation infrastructure installation service can discuss agreed customer-side equipment mounting, physical IT cabling, labeling and handoff records, subject to scope, site requirements and availability. This article's billing education adds no utility-control or billing service to that offer.

Use the related project inquiry for the physical installation: share the nonconfidential facility location, equipment population, requested tasks, available rack and cable plans, and intended work window. Keep utility-account and confidential commercial records with the authorized review team.

References: [L1]