The New Large-Load Compact

Why AI Data Centers Can No Longer Be Treated as Passive CustomersFor decades, even very large electric customers fit inside a familiar planning model. A factory, refinery, industrial campus, hospital complex, hospital system, or manufacturing facility requested service. The utility studied the load, identified required facilities, assigned costs, and folded the demand into its forecast and capital plan. That model treated the customer primarily as load: something to be forecast, connected, and served.Artificial intelligence (AI) data centers are testing that assumption. Their power requirements are large, their commercial timelines are compressed, and their reliability expectations are unusually high. A developer may seek service in one or two years. A utility may need several years to complete studies, secure transformers, build substations, expand transmission, arrange generation, evaluate fuel requirements, and obtain permits or regulatory approvals.The result is a different planning problem. At sufficient scale, a large load affects the system around it. Its timing influences transmission investment. Its location affects deliverability. Its operating profile affects resource adequacy. Its flexibility affects emergency operations. Its onsite generation affects fuel, emissions, standby service, and grid coordination. Its cost responsibility affects affordability and public acceptance.Recent work from Berkeley Lab organizes the challenge across five functional areas: load forecasting, interconnection, resource planning and procurement, markets and operations, and cost allocation and ratemaking. That framing is useful because large-load connection bottlenecks do not sit in one process. They cut across the full machinery of electric-system planning.The Federal Energy Regulatory Commission's (FERC's) June 2026 large-load tariff actions point in the same direction. By directing the six jurisdictional regional grid operators to justify or reform rules for large energy users, including data centers and advanced manufacturing facilities, the Commission effectively recognized that current procedures may not be adequate for the scale, speed, and complexity of AI-era demand.The appropriate response is neither to treat data centers as ordinary load nor to treat them as inherent threats. A better framing is a reciprocal compact. Large customers need clearer and faster pathways to service. Utilities and grid operators need better information, firmer commitments, enforceable flexibility, and clearer cost responsibility. Regulators need assurance that reliability is protected and that existing customers are not left paying for speculative or poorly structured projects.The grid can no longer plan merely around large loads. It has to plan with them.

"Time to power" has become one of the defining constraints in data-center development.For utilities, power delivery is a planning, engineering, procurement, permitting, and regulatory sequence. Load must be studied. Facilities must be designed. Equipment must be ordered. Transmission impacts must be evaluated. Generation or market supply must be available. Costs must be allocated. Regulators may need to approve investments, tariffs, or special contracts.For AI infrastructure developers, power availability is a commercial gating item. Compute capacity has strategic value only if it can be deployed when the market needs it. In that environment, a multiyear utility-service timeline can function like a denial even when the utility is not saying no.That difference in clocks is reshaping site selection. Earlier data-center development cycles often emphasized proximity to fiber, customers, tax incentives, land, and major metro areas. Those considerations still matter. In constrained regions, however, electric capacity, substation availability, transmission deliverability, water availability, and utility-process maturity can determine whether a project is feasible.Berkeley Lab's Speed to Power report identifies 41 potential solutions for accelerating large-load connections and highlights several recurring challenges: load forecast uncertainty, process coordination, interconnection uncertainty, capacity adequacy, operational impacts, and cost-shifting or stranded-cost risk. Those categories mirror the issues now confronting utilities, grid operators, regulators, and large customers.Scale intensifies the difficulty. Lawrence Berkeley National Laboratory (LBNL) estimated that U.S. data centers consumed 176 TWh of electricity in 2023, about 4.4% of total U.S. electricity consumption. Depending on demand growth, efficiency, and broader economic conditions, LBNL estimated that data-center electricity use could reach 325 TWh to 580 TWh by 2028, or roughly 6.7% to 12% of projected U.S. electricity consumption that year (Table 1).[caption id="attachment_264575" align="aligncenter" width="1280"]

Table 1. U.S. data-center electricity use, historical and projected. Source: Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report[/caption]These figures require careful use. A projection is not an obligation to build. A request for service is not the same as energized load. A queue entry is not the same as a financeable project. Some announced data-center demand may be delayed, downsized, relocated, served behind the meter, or never built.Uncertainty, however, sharpens the planning problem. If utilities and grid operators assume too much load will materialize, they risk overbuilding infrastructure and shifting costs to existing customers. If they assume too little, they risk underbuilding and facing reliability, congestion, or economic-development consequences. The planning challenge is to distinguish among speculative, proposed, mature, contracted, staged, flexible, self-supplied, and energized load.A 500-MW request from a customer with site control, financing, a phased energization plan, credit support, and a defined supply strategy is different from a 500-MW exploratory inquiry. A fully firm data-center load differs from one that can curtail during emergency conditions. A facility with observable, fuel-secure onsite generation differs from one with emergency generators permitted only for limited backup operation.The old interconnection model asked whether the grid could serve the load. The new planning model must ask under what conditions the load can connect quickly without weakening reliability, affordability, or system visibility.

FERC's June 2026 large-load action should be understood as a planning intervention, not only a tariff proceeding. The Commission ordered the six regional grid operators under its jurisdiction to justify or reform tariffs for data centers and other large energy users. The stated objective was speed to power, but the action also emphasized ratepayer protection and the need for clear, consistent tariff provisions for large-load customers.That matters because tariffs and interconnection procedures convert planning assumptions into enforceable obligations. They define who can request service, what information must be provided, how studies are conducted, how costs are assigned, and what rights and obligations accompany service.FERC supplies the regulatory anchor for this discussion, but the planning reforms required to operationalize that action extend beyond the tariff order itself. Drawing from FERC's tariff concerns, Berkeley Lab's speed-to-power framework, the Energy Systems Integration Group's (ESIG's) interconnection recommendations, and emerging large-load practice, five reform areas emerge: project maturity, cost responsibility, coordination, study design, and flexible service.The first reform area is project maturity. Utilities and grid operators need practical filters for distinguishing serious projects from speculative requests. That requires information on site control, creditworthiness, financing, phased energization schedules, equipment procurement, commercial commitments, and intended service characteristics. Berkeley Lab identifies milestone requirements for inclusion in forecasts and processes for monitoring project duplication as potential solutions, reflecting the same underlying need: forecasted large load should be disciplined by evidence of readiness.The second reform area is cost responsibility. Existing customers should not be asked to absorb costs caused primarily by a new large customer unless there is a clear system benefit. Some upgrades triggered by large-load growth may provide broader network value; others may be customer-specific. The rules need to separate those categories in a transparent and defensible way.The third reform area is coordination. The ESIG Large Loads Task Force report reinforces this point by finding that large-load interconnection practices vary across utilities and regions, and that many procedures were not designed for facilities of today's scale. It also highlights inadequate coordination between utilities and independent system operators/regional transmission organizations (ISOs/RTOs), duplicated effort, and inconsistent visibility into cumulative system impacts.That ESIG finding supports the practical importance of FERC's action. A utility may study the local interconnection facilities, while an ISO or RTO may need visibility into cumulative transmission, resource adequacy, or operational impacts. Poor coordination can create delay, duplicated analysis, and planning blind spots. ESIG's recommendations for formal information-sharing protocols and joint study procedures provide a useful implementation path.The fourth reform area is study design. ESIG identifies cluster study approaches as an emerging option as large-load request volume grows. A cluster study for large loads must account for more than transmission impacts. It should consider service firmness, ramp schedules, onsite generation, flexibility, affected systems, and the sequencing of network upgrades.The fifth reform area is flexible service. ESIG calls attention to non-firm, surplus, and provisional service options that could accelerate large-load interconnection where firm service is not immediately available. These options deserve serious evaluation, but only with disciplined performance requirements. Non-firm service should not become vague reliability optimism. It needs clear operating rules, curtailment triggers, telemetry, communications, compensation, and consequences for nonperformance.Speed without discipline can create reliability and affordability risk. Discipline without speed can push customers toward workarounds that are less visible to planners. The better objective is structured acceleration: faster pathways for mature, transparent, financeable, operationally clear projects that accept appropriate obligations.

The large-load issue becomes more serious when viewed through resource adequacy. The question is not only whether enough capacity exists on paper. It is whether dependable energy can be delivered to the right location at the right time under stressed conditions.The North American Electric Reliability Corporation's (NERC's) 2025 Long-Term Reliability Assessment (LTRA) forecast that summer peak demand would grow by 224 GW, a more than 69% increase over the prior LTRA forecast, with new data centers for AI and the digital economy identified as a major contributor (Table 2). The implication is not that every region faces the same risk or that every data-center announcement will become actual load. The implication is that load growth is now large enough, uncertain enough, and regionally concentrated enough to require more disciplined planning categories.[caption id="attachment_264576" align="aligncenter" width="1280"]

Table 2. The North American Electric Reliability Corporation (NERC) 2025 Long-Term Reliability Assessment (LTRA) demand-growth signal. Source: NERC[/caption]Reserve margins remain important, but they do not capture everything planners must manage. A region may show adequate capacity and still face risk if resources cannot run during extreme weather, if fuel is constrained, if transmission cannot deliver power into a load pocket, if imports are unavailable, or if storage duration is insufficient for the event.For large-load integration, that distinction is decisive. A 500-MW data-center load in a constrained transmission area is not equivalent to a 500-MW load near deliverable generation. A fully firm 500-MW load is different from a staged or interruptible 500-MW load. A customer with observable, fuel-secure onsite generation should not be treated the same as a customer with emergency-only backup that cannot operate for extended periods.Fuel is a particular concern. If new large loads are served by gas-fired generation, either on the grid or onsite, the adequacy question becomes partly a gas-deliverability question. Firm pipeline transportation, fuel contracting, winter deliverability, dual-fuel capability, emissions limits, and local infrastructure all become relevant. If diesel backup or bridge generation is used more frequently, fuel logistics, air permits, refueling, and local emissions constraints become more important. If water-cooled thermal resources serve new load, water availability and cooling constraints can affect reliability and siting.Flexibility can help, but only if it is operationally real. A data center may be able to shift workloads, curtail noncritical computing, stage load ramps, or rely on onsite generation during emergencies. Those capabilities could be valuable. They need telemetry, dispatch protocols, communication rules, performance standards, compensation structures, and consequences for nonperformance.Resource adequacy is therefore also a customer-commitment question. The system needs to know how large loads will behave when reliability is at stake.

Data centers have long relied on layered power-continuity systems. In the traditional model, the grid served the facility under normal conditions. Uninterruptible power supply (UPS) systems bridged momentary disturbances. Backup generators carried the load during longer outages. The arrangement assumed that the grid was the primary source of power and that customer-side equipment existed to protect the facility when grid service failed.That model remains relevant, but its role is changing. As utility power becomes constrained in key markets, onsite power is moving from insurance to strategy. Some data-center developers are evaluating onsite generation as bridge power until grid service is available. Others are considering it as supplemental power during constrained periods. Some may pursue adjacent generation or co-located power arrangements to accelerate service. In the most aggressive cases, onsite generation can become a primary power strategy rather than an emergency measure.This shift has practical advantages. Onsite generation can shorten time to power (Figure 1), reduce dependence on congested transmission paths, improve resilience, and allow staged energization. Properly designed, it may also provide grid services such as voltage support, spinning reserve, synchronous support, or emergency response.[caption id="attachment_264580" align="aligncenter" width="800"]

1. Reciprocating engine–based onsite generation is increasingly relevant to data-center power strategy as developers seek faster access to reliable supply in constrained grid regions. Wärtsilä announced in 2025 that it would supply 282 MW of flexible engine capacity, using 15 18V50SG natural gas engines, for a new Ohio data-center project. Courtesy: Wärtsilä[/caption]It also creates obligations and limits. A generator fleet designed for rare emergency use is different from one used for routine or prime service. Fuel logistics become more important. So do air permits, emissions compliance, maintenance intervals, noise, local opposition, and gas deliverability. If the equipment is intended to support the grid, operator visibility and performance obligations matter.A data center may appear to reduce grid burden by adding onsite power, but the grid may still be expected to provide backup service during equipment outages, fuel interruptions, maintenance windows, or emergency conditions. If the system does not understand those dependencies, planners may understate risk.The right question is what function onsite generation performs: emergency backup, bridge supply, prime power, co-located supply, grid-supporting resource, market participant, or customer-side insurance. Each answer has different implications for planning, tariffs, emissions, fuel, and cost allocation.Onsite power and flexible load are best understood as planning variables. They may help solve the time-to-power problem, but only if their capabilities and limits are visible to the system.

The large-load compact cannot be designed only for hyperscalers, RTOs, ISOs, and transmission planners. It also has to work for existing customers and local communities.That point is especially important for public-power utilities, municipal systems, cooperatives, and smaller utilities. A very large load can be an economic-development opportunity, but it can also be a financial and operational risk. Smaller systems may not have the staff, balance sheet, or negotiating experience of a large investor-owned utility or RTO. Yet they may face sophisticated data-center developers seeking rapid service, special terms, or major infrastructure commitments.The opportunity is real. A data center can expand the local tax base, support economic development, improve utilization of existing infrastructure, and justify upgrades that may benefit other customers. The risks are also real. If the project does not materialize, existing customers could be left with study costs, planning costs, or stranded infrastructure. If the project materializes without adequate cost assignment, customers may face rate pressure. If the load arrives before sufficient power supply is available, reliability can suffer.Local utility questions belong in the national data-center power debate. Does the community support studying the project? What are the expected local benefits? What are the local impacts? What contractual protections exist? Has the customer provided credit support? Are upfront payments required? What rate design applies? Who pays if the project is delayed, downsized, or cancelled? How are power-supply obligations handled?These are not anti-growth questions. They are the minimum questions needed to distinguish economic development from unmanaged risk.

Large-load integration begins with electricity, but it rarely ends there. A data-center project that appears to be an electric-service question can quickly become a cross-infrastructure planning problem.Gas infrastructure is one example. If data centers rely on gas-fired onsite generation, or if utilities add gas-fired resources to serve new demand, the reliability question becomes partly a fuel-deliverability question. Pipeline capacity, firm transportation, fuel contracting, storage, winter deliverability, dual-fuel capability, emissions limits, and local infrastructure may all affect whether generation is dependable during stressed conditions.Diesel markets present a different concern. Emergency generators may be manageable when they run rarely. If diesel equipment is used more frequently as bridge or prime power, fuel logistics, refueling during emergencies, local emissions, permitting, and supply availability become more material.Water is another constraint. Data centers may require water for cooling, depending on design and location. Thermal generation serving those facilities may also require cooling water. In water-constrained regions, data-center siting, generation technology, electric reliability, and local water utility capacity can become linked.Equipment supply chains may be the most immediate limitation. Transformers (Figure 2), switchgear, breakers, cables, engines, turbines, power electronics, protection systems, and substation equipment all affect the real timeline to serve new load. A project may have a signed agreement and still be constrained by equipment availability.[caption id="attachment_264579" align="aligncenter" width="1280"]

2. Large-load integration depends not only on generation but also on transmission and substation equipment such as large power transformers. Siemens Energy announced a $150 million expansion in Charlotte, North Carolina, to address U.S. power-transformer shortages. Courtesy: Siemens Energy[/caption]Permitting connects these systems. A generation solution may be technically available but delayed by air permits. A transmission solution may be justified but delayed by siting. A water solution may be feasible but locally contested. A gas solution may require infrastructure that has its own approval process.Electricity is the center of the issue. It is not the only system being tested. Large-load review should therefore account for the full infrastructure stack needed to support a project reliably, affordably, and lawfully.

A practical large-load compact begins with a simple premise: AI data centers and other major loads are neither ordinary customers nor inherent threats. They are large, fast-moving, operationally consequential customers whose decisions now affect the system around them.That requires a reciprocal planning relationship (Table 3). Large customers should receive clearer and faster pathways to service. In return, utilities, grid operators, regulators, existing customers, and communities should receive better information, firmer commitments, enforceable flexibility, fair cost responsibility, and operational visibility.[caption id="attachment_264577" align="aligncenter" width="1280"]

Table 3. A practical large-load compact. Source: Emeka N. Obikwelu, PhD, PE, PMP[/caption]Project maturity is the first requirement. A large-load request should not be treated as actionable demand merely because it is large.Cost responsibility is the second. Customers that drive customer-specific facilities should expect to bear those costs, while regulators should explain when broader network benefits justify shared costs.Flexibility is the third requirement. Some data-center loads may be able to curtail, stage service, shift workloads, or use onsite resources during stressed periods. That flexibility must be verified, measured, and enforceable.Onsite-power clarity is the fourth requirement. Backup power is not prime power. Bridge supply is not market-facing generation. A co-located resource is not automatically disconnected from the grid's obligations.Community and ratepayer protection is the fifth requirement. Existing customers should not become the default backstop for speculative, poorly structured, or inadequately secured projects. Communities should understand the local power, water, land, fuel, environmental, and emergency-service implications before major commitments are made.Regional discipline is the final requirement. The Electric Reliability Council of Texas's (ERCOT's) Batch Zero process is one example of a region attempting to impose structure on a large-load request pipeline. According to ERCOT, it is tracking more than 438,000 MW of large-load requests, nearly 89% from data centers, and Batch Zero is intended to group qualified large-user applicants into a new study process. That request volume should not be read as committed demand, but it shows why screening, maturity requirements, onsite-generation treatment, and curtailment pathways are becoming central to planning.

The next phase of large-load integration will be shaped less by broad statements about AI demand and more by the details of implementation.The first issue to watch is how utilities, RTOs, ISOs, and regulators define project readiness. Large-load processes will need clearer distinctions among speculative inquiries, mature projects, contracted load, staged load, flexible load, and energized demand. Queue volume alone will not be enough.The second issue is how flexible service develops. Non-firm, surplus, provisional, interruptible, or staged service could help some projects connect sooner, but only if operating limits are clear. Flexibility must be observable, measurable, enforceable, and reflected in planning models.The third issue is how onsite and co-located generation are treated. Customer-side power may reduce grid burden in some cases, but it may also create standby, fuel, emissions, and emergency-service obligations. The planning value of onsite generation should depend on function, visibility, fuel assurance, permits, and performance requirements.The fourth issue is whether large-load review expands beyond electricity alone. Gas deliverability, diesel logistics, water availability, equipment supply chains, permitting, and local infrastructure may determine whether a project that looks feasible on paper can actually be served.

The power sector has served large industrial customers before. What is different now is the combination of scale, speed, reliability expectations, regional clustering, infrastructure constraints, technology competition, and uncertainty about how much AI demand will ultimately materialize.If every large-load request is treated as fully real, the system may overbuild and shift costs to existing customers. If large-load growth is discounted too aggressively, the system may underbuild and face reliability risk. If onsite generation is ignored, planners may miss a potential resource. If onsite generation is overcredited, planners may assume reliability value that does not exist. If flexibility is recognized but not enforceable, the system may rely on a tool that fails when needed.The answer is to define the conditions under which large loads can connect quickly, pay fairly, operate transparently, and support rather than destabilize the grid. Large loads should no longer be treated merely as passive customers at the end of the delivery system. At their current scale, they are becoming planning actors whose decisions affect everyone else connected to the grid.The next grid-planning question is not simply how fast large loads can be connected. It is what obligations should come with connection at that scale.—Emeka Obikwelu, PhD, PE, PMP is director of Grid Systems at the U.S. Department of Energy's Office of Electricity. The views expressed in this article are his own and do not represent the position of the Department of Energy or any federal government agency.