What is behind-the-meter power?

The meter marks an accounting and electrical boundary. Electricity delivered by a utility crosses it; electricity made and consumed on the customer side generally does not. The U.S. Energy Information Administration calls on-site, behind-the-meter generation distributed generation and includes solar, storage, and combined heat and power in that category. The term describes where power is produced and consumed, not whether the fuel is clean, whether the plant is permanent, or who owns the generator.

Imagine a campus with a utility service and a gas turbine connected to its own electrical bus. When the turbine serves the buildings, its output is behind the meter. When the campus imports electricity during turbine maintenance, that imported energy crosses the meter. If the same plant exports power, the point of interconnection and metering arrangement become part of the project design. A single campus can therefore have behind-the-meter generation, grid imports, and exports at different times.

This distinction is easy to lose when people use “behind the meter” as a synonym for off-grid. DOE’s microgrid explanation describes both grid-connected and islanded operation. A site can be behind the meter and connected to the grid; it can also be designed to separate safely and operate as an island. Those two operating states require different protection, controls, and reserve decisions.

Start with the one-line diagram, not a slogan. One arrangement keeps the grid as the normal supply and uses on-site generation during a disturbance or peak. Another runs on-site generation as the normal supply and retains a grid tie for scheduled imports or backup. A third is a true island with no usable utility tie. DOE describes microgrids as local generation and loads that can operate with the larger grid or independently; the ability to separate requires coordinated controls and protection.

Consider a data hall drawing 60 MW at its facility bus. If it has a 60 MW grid connection and a 60 MW on-site plant, the operating agreement still has to say whether simultaneous generation can export, whether the grid may supply the full load when the plant trips, and what happens during a transfer. A breaker opening is not a complete operating plan. The UPS may carry the IT load for a short interval, while generators and pumps must start or remain online.

A site that has never connected to the grid avoids a grid-service process for that absent connection. A site that keeps or later adds a tie must study the tie and its operating modes. FERC’s 2026 large-load action specifically addresses co-located generation and load, including how they take transmission service. That is why a project team should label each mode on the one-line: normal generation, import, export, island, restart, and maintenance.

A data center needs a firm electrical path at the date it intends to place servers in service. If a planned utility connection requires line or substation work, that delivery date may be later than the building schedule. On-site generation gives the developer another sequence to evaluate: acquire equipment, secure fuel and permits, build the plant, and energize the local bus. That sequence can be attractive, but it is a separate construction program with its own critical path.

The pressure is visible in electricity forecasts, although those forecasts do not prove any one site will be delayed. The IEA Energy and AI report projected about 945 TWh of global data-center electricity use in 2030 in its 2025 base case. The December 2024 Berkeley Lab U.S. report modeled U.S. demand scenarios through 2028. These are system-wide annual energy estimates; a developer still needs a local peak-demand and transmission study.

The queue number often quoted in this debate belongs to generators. Berkeley Lab’s Queued Up: 2026 Edition counted 1,312 GW of generation and about 749 GW of storage seeking transmission interconnection at the end of 2025. It explicitly excludes load requests and behind-the-meter projects. The reported median of more than five years from request to commercial operation for projects built in 2025 is a generator-project statistic, not a promised data-center wait. Our interconnection queue guide explains what the figure can and cannot tell you.

The first number is the facility peak, which includes IT equipment and the power needed to cool and distribute electricity. The second is the load profile: how often the site runs at peak, how quickly a synchronized AI workload steps up, and which loads can pause. The third is the acceptable outage or transfer interval. These numbers decide whether a generator nameplate is enough after ambient derating, maintenance, and a failed unit. The IEA separates servers, storage, networking, and facility infrastructure; it does not treat a GPU’s rating as the whole campus draw.

For example, a campus with 50 MW of IT demand and an assumed design PUE of 1.2 has a 60 MW facility requirement. That arithmetic is a planning scenario, not a site forecast. If the cooling pumps, switchgear, and UPS losses were omitted, a nominal 50 MW plant would be undersized before a generator goes out of service. Our AI data-center power guide develops the rack-to-facility calculation.

Next ask which demand must remain energized after one generator trips. A site may shed test clusters while preserving storage and network services, or it may require the entire training run to continue. Those are different plant sizes. Controls also matter: NVIDIA documents coordinated AI workload swings and rack-level power smoothing in GB300 NVL72. A plant that has enough steady-state megawatts can still have a transient problem when many racks change state together.

No technology solves every part of the load shape. Gas turbines and engines can be dispatched when the data center needs them, but their fuel supply and air-emissions conditions must work at the site. Geothermal can deliver continuous output where a resource is developed and proven, but drilling and reservoir performance make it a different development project. Solar contributes when irradiance is available; storage shifts energy and responds quickly, but its duration is finite. DOE’s large-load microgrid discussion explicitly treats generation, storage, cooling, and water as one design problem.

Take a hypothetical 80 MW campus. A solar array may reduce gas consumption on sunny hours, while batteries handle fast power changes or a short transfer. Neither establishes overnight firm supply without enough energy or another generator. An EGS project could eventually cover a stable base load, but it must first prove wells and complete the power plant. The DOE EGS explanation distinguishes an engineered subsurface reservoir from a closed-loop geothermal design.

That leads to a useful procurement question: which source is firm at the facility bus under the proposed operating conditions? A nameplate on a brochure cannot answer it. Ask for the output at local temperature and elevation, fuel conditions, water needs, maintenance intervals, and how a unit failure changes delivered power. If a proposal combines sources, request a dispatch model for a hot afternoon, a cloudy evening, a plant outage, and a grid outage.

It can avoid a generator export request only if there is no export and the applicable local rules allow that arrangement. It does not erase the study and service questions for an existing or proposed utility tie. If a site wants the grid to cover its full load when its plant is down, the provider must evaluate that service, any needed network work, metering, and operating rules. If it wants to sell surplus power, export rights create another set of requirements.

As of November 2026, large-load treatment is changing. DOE’s October 2025 direction asked FERC to consider rules for large loads. FERC’s June 2026 show-cause orders asked six regional grid operators to justify or reform their tariffs; they did not create a universal finished path for every data center. An operator should therefore check the transmission provider’s current tariff and the serving utility’s process for its specific connection.

For contrast, FERC Order No. 2023 concerns generator interconnection and first-ready, first-served cluster studies. Applying its generator queue duration to a load-only request gives a false schedule. A co-located campus can touch both sets of rules, depending on whether it seeks grid imports, exports, or both. The practical document to request early is a written description of permitted operating modes, not a verbal assurance that “behind the meter skips the queue.”

A turbine selected for output still needs permission to operate and a reliable way to receive fuel. EPA’s data-center air resources identify federal standards for stationary combustion turbines and engines and explain that state and local agencies issue most air permits. The allowable emissions, hours of operation, and controls for a proposed unit have to be checked with the permitting authority. A standby engine authorized for emergencies is not automatically a legal primary generator.

Suppose a project plans to run a diesel-backed campus continuously because the equipment is already near the site. The plant may be physically capable of producing the megawatts, yet the approved operating conditions may allow only emergency use. Likewise, a gas turbine with a manufacturing slot is of little use if the gas interconnect and pipeline capacity cannot support its daily fuel requirement. These are design gates, not details to settle after the turbine order.

Water also belongs on the list. The heat from a data center must be rejected even when the utility is unavailable. A cooling concept that relies on grid-fed pumps or a water supply that is constrained during heat waves can undermine the power plan. DOE’s microgrid discussion calls out cooling loads and water as part of integrated large-load design. Ask the generation and mechanical teams to show the same outage scenario on their drawings.

Capacity is only one requirement. The plant has to regulate voltage and frequency, ride through load steps, coordinate protective devices, and restart after a complete shutdown. A grid-connected site can normally draw stability from the larger network; an island must provide it locally. DOE’s microgrid basics warns that independent operation needs controls and protection that keep equipment and workers safe.

Consider two turbines serving one bus. If one trips while both are near maximum output, the second unit cannot instantly replace the lost power. The operating plan may use batteries, UPS, reserved headroom, or automatic load shedding. The choice must specify which racks stop first and whether cooling continues. A test that removes a unit under a representative AI workload says more than a static statement that the site has “N+1” machines.

Maintenance is a second reliability test. A turbine taken down for inspection changes the available capacity for days or weeks. The site should model an unplanned trip during that scheduled outage. Fuel systems, emissions equipment, heat rejection, and switchgear also need maintenance and spare-part plans. Redundancy in generation cannot compensate for a single switchboard or gas connection shared by every unit.

A staged build can align IT deployment with real delivered power. For example, a campus might commission one data hall and its electrical plant first, then add another hall only after a second generation block, cooling train, and fuel path pass acceptance tests. This reduces the risk of placing servers in a building whose electrical bus cannot support them. It also lets operating data replace assumptions before the next phase.

A common plan is to start with temporary or modular generation and later add a grid tie or another firm source. The word “temporary” should describe a documented exit condition. Which equipment is removed, which assets remain as standby, and can the later connection be made without a long outage? GE Vernova describes its TM2500 as a mobile unit for baseload bridge service, but that product description does not establish a delivery date for a particular campus.

The alternative is to design permanent on-site supply from the start. That makes ownership or long-term operating contracts, staffing, parts, fuel, and environmental compliance central to the campus business case. A real schedule should put all critical paths on one page: permits, fuel, generation equipment, substation and switchgear, cooling, controls, construction, testing, and any future grid agreement. Our turbine backlog explainer covers why a “fast” equipment class may still have a long sourcing timeline.

Ask for a site-specific single-line diagram and a stated power boundary. Does the quoted output mean gross generator terminals, net plant bus, or usable capacity after cooling and auxiliary loads? Ask for an hourly dispatch profile and a peak-load test case. Put local weather, fuel conditions, emissions limits, planned maintenance, and a largest-unit trip into the assumptions. If the proposal uses storage, specify its usable energy and what happens after discharge.

A 100 MW proposal could mean 100 MW of turbine nameplate at standard conditions or 100 MW delivered to the data center after every plant auxiliary. Those are materially different commitments. EIA distinguishes generating capacity from actual electricity produced; that same distinction matters when a developer reads a plant offer. The contract should identify the measurement point and performance conditions.

For a grid-tied proposal, request a service diagram showing maximum imports and exports and the party responsible for each study and upgrade. For an islanded proposal, request the black-start sequence, spinning reserve logic, and equipment that holds the load between a disturbance and stable generation. For any proposal, request an acceptance-test plan that uses a representative stepped load. These documents expose whether the generation concept and the data-center concept actually describe the same facility.

If the power concept calls for turbines, generators, or electrical equipment, send the required output, fuel, voltage, condition preference, location constraints, and target date with a request for a quote. We will confirm what our network can source for that configuration. If you have equipment to redeploy, tell us its model, condition, and documentation so we can see whether our network can place it. The parties assess the equipment and transaction directly; a sourcing introduction is not a performance guarantee.

The most damaging mistake is calling every behind-the-meter site “off-grid.” A campus that needs a utility tie has a grid-service and protection question even when its normal power comes from on-site generation. Another is using a generator-interconnection statistic as the expected wait for a data-center load; the Berkeley Lab queue dataset explicitly excludes load requests.

A third mistake is treating nameplate capacity as usable facility power. Ambient conditions, plant auxiliaries, scheduled outages, and the largest-unit trip all change the answer. Finally, do not assume that generation equipment itself grants permission to run continuously. Verify the permit, fuel path, interconnection mode, and cooling system against the same operating scenario before scheduling server delivery.