Planning for Uncertain Data Center Demand

After decades of relatively flat electricity demand, US load growth is back. In the near term, most of that growth is expected to come from data centers.

Utilities have managed load growth before, including at higher growth rates than are forecast today. What is different now is the scale, concentration, and uncertainty of individual customer requests. A single data center campus can require as much electricity as a small city, and many of the largest requests are clustered in a small number of utility service territories. In 2024, for example, AEP Ohio reported pending data center requests totaling 30,000 megawatts (MW) — enough to more than triple its peak load. Signed agreements alone would increase central Ohio data center load from 600 MW in 2024 to 5,000 MW by 2030.

Many utilities are preparing to meet data center demand by building or contracting for gas-fired generation. Across the country, gas capacity in utility plans is surging, even as gas turbine supply chains tighten and costs rise. Among the utility integrated resource plans (IRPs) covered in RMI’s Engage and Act dataset, total projected gas capacity by 2035 rose from 275 to 350 gigawatts (GW) between December 2022 and March 2026 forecasts (Exhibit 1).

Exhibit 1

Recent utility filings provide more detail into why utilities are planning to build new gas plants. In Louisiana, Entergy has said it intends to build more than 5,200 MW of new combined-cycle gas power plants to support Meta data centers, along with up to 2,500 MW of solar and support for nuclear power and battery storage. In Georgia, the Public Service Commission approved Georgia Power’s request for almost 10,000 MW of new generating capacity, nearly 6,000 MW of which is gas-fired generation — all as part of a plan to meet future capacity needs including 15,600 MW from 32 large load customers.

These decisions reflect a broader planning challenge: utilities must make large investment decisions before they know how much data center demand will actually materialize. That uncertainty is not new. In the early 1970s, many utilities planned for sustained high demand growth and committed to large new nuclear projects. When electricity demand growth slowed after the oil shocks, inflation, and recession, some projects became uneconomic and were canceled, while others left customers paying for investments that no longer matched system needs. In Washington state, abandoned nuclear projects left bondholders on the hook for more than $2 billion.

Today, meeting uncertain data center demand with large, long-lived, high-cost assets creates similar financial risks for utility customers. These assets may be underutilized if the load they are intended to serve arrives late, develops at a smaller scale than expected, relocates, or never materializes. While large-load tariffs with protections such as minimum bills, collateral requirements, exit fees, and customer-funded infrastructure can influence how these risks are allocated, they do not eliminate them. Gas plants carry specific risks, including construction-cost escalation and delays associated with growing turbine backlogs, exposure to volatile fuel prices over their operating lives, and the possibility that future emissions limits will increase compliance costs, constrain plant utilization, or lead to earlier-than-planned retirement.

Modeling the risks of planning for uncertain demand

To illustrate the risks facing utility planners today, this analysis examines a representative mid-sized utility in the Midwest with a substantial large load pipeline. The analysis has been calibrated to that utility’s integrated resource plan, using publicly available data. However, it is meant purely to illustrate the risks associated with different resource strategies, not to reflect any utility’s actual procurement decisions.

The analysis compares five scenarios that vary by resource strategy and the amount of data center load that ultimately materializes (Exhibit 2). Scenarios 1 and 3 use a gas-first strategy, in which the utility builds enough combined-cycle gas capacity to serve the full potential data center load, while other resources may serve projected non-data center demand. The remaining scenarios use a portfolio approach, allowing a capacity expansion model to select the least-cost mix of resources to meet total system demand.

Scenarios 1 through 4 all plan for the high end of potential data center demand. That demand fully materializes in scenarios 1 and 2, while scenarios 3 and 4 test what happens if only a portion appears. Scenario 5 provides a reference case in which the utility correctly forecasts the lower level of data center demand. Comparing results across these scenarios shows the cost of overbuilding, and how that cost differs between a gas-first strategy and a more diverse portfolio.

Exhibit 2

It is also possible that a utility may underforecast demand growth and underinvest in grid capacity. The risks differ from overinvestment: underforecasting can delay interconnection, reduce utility revenues, and constrain local economic development. We do not quantify these risks because they cannot be represented well in this capacity-expansion framework. Instead, we focus on overinvestment, historically the more common error in US utility forecasts.

However, many of the recommendations that follow also address underforecasting by emphasizing better forecasting, faster planning, and procurement of quick-to-deploy resources such as batteries, virtual power plants, and grid-enhancing technologies that can relieve near-term constraints while larger infrastructure is developed.

Quantifying the impacts on planning outcomes

The modeling reveals three takeaways for regulators:

• Overinvesting based on uncertain demand forecasts creates risk because fixed costs must be recovered from fewer customers than expected.
• Portfolio planning reduces downside exposure relative to a gas-first strategy because more of the resource mix retains value across load outcomes.
• Allocating costs appropriately to large, uncertain loads like data centers is vital for protecting households from paying for potentially underutilized infrastructure.

The five modeled scenarios produce three distinct resource builds, shown in Exhibit 3. Under gas-first procurement, the utility adds substantial gas capacity to serve projected data center demand, while solar and demand flexibility (DF) meet other load growth and replace retiring coal and oil generation. The optimized portfolio approach relies more heavily on solar, short-duration battery storage (BESS), and long-duration energy storage (LDES). It selects several gigawatts of new gas when data center demand is high, but very little when demand is low.

Exhibit 3

These resource builds produce substantial differences in system-wide generation costs. Exhibit 4 highlights two main results.

First, fixed costs are driven primarily by how much demand the utility plans for. These costs — including depreciation on existing infrastructure, investment costs for new generation, and fixed maintenance costs — range from $3.0 billion to $3.2 billion per year when the utility plans for high demand, as in scenarios 1 through 4.^[1] They fall to $2.2 billion per year when the utility correctly plans for lower demand, as in scenario 5. The difference illustrates the value of improving demand forecasts and avoiding unnecessary investment.

Second, variable costs are driven primarily by how much demand actually shows up. Fuel, start costs, and other thermal operating costs range from $1.6 billion to $1.8 billion per year in the high-load cases, scenarios 1 and 2. In the low-load cases, scenarios 3 through 5, they range from $0.9 billion to $1.2 billion per year. Across load outcomes, portfolio approaches have lower variable costs than the gas-first strategies because they rely less on fuel-based resources.

Exhibit 4

Who ultimately bears these costs depends on state cost-allocation and ratemaking practices. Some jurisdictions are developing tariffs that require data centers and other large loads to bear more of the risks associated with new infrastructure. Elsewhere, those risks may be spread more broadly across the customer base.

Exhibit 5 shows the potential rate impacts if new generation costs are socialized across all customers.[2] The results yield three key insights:                                             

1. Overinvestment could produce substantial rate increases. Under the portfolio approach, overinvesting causes the generation component of electricity rates to rise 0.89 cents per kWh, or 19%. Under the gas-first strategy, overinvesting leads to an increase of 1.12 cents per kWh, or 23%. For an average household consuming 10,500 kWh of electricity per year, that translates to increased costs of $94 to $118 per year.
2. Portfolio planning lowers costs. Gas-first procurement raises rates by 0.26 cents per kWh, or 6%, when the high load forecast materializes. When the utility overinvests, gas-first procurement increases rates by 0.49 cents per kWh, or 9%. For the average household, this represents an additional cost of $28 to $52 per year.
3. Cost allocation and recovery matter. Large-load tariffs can protect households from some of the costs associated with overinvestment or a higher-cost resource strategy. But cost allocation does not eliminate the underlying costs; it only determines who bears them. Whether those costs fall on the data center, other customers, utility shareholders, or some combination, all parties benefit from keeping total system costs as low as possible.

Exhibit 5

What decision makers can do to mitigate risks around uncertain data center demand

This analysis highlights the risks of procuring generation around uncertain data center demand. Managing that risk requires action across the full decision chain: improving the information that goes into the plan, preserving a system-level comparison of resource options, prioritizing least-regrets investments, staging procurement as uncertainty resolves, and allocating remaining risk through tariffs and contracts (Exhibit 6).

Exhibit 6

Improve the information going into the plan

Better forecasting can limit costly overbuild. In the lower-load case, correctly forecasting demand saves $700 million per year, or $101 per household. Utilities and system operators can reduce the risk of over or underforecasting by distinguishing speculative inquiries from committed projects, tracking duplicate requests across service territories, incorporating project-specific milestones into forecasts, and comparing forecasted demand with actual development over time.

Real-world examples of these strategies include Dominion Energy’s customer-specific data center forecasting, Georgia Power’s pipeline-stage tracking, PJM’s stronger evidence requirements for near-term large loads, and ERCOT’s use of observed project outcomes to adjust forecasts.

Large-load tariffs can also improve forecast quality by requiring customers to demonstrate commitment through collateral, minimum demand obligations, phased ramps, contract terms, and exit fees. AEP Ohio and Indiana Power Michigan provide evidence: after implementing large load tariffs, their 2030 system-wide demand forecasts dropped by about one-third and one-half respectively.

Speed up the planning process

System planning creates value, but it needs to move fast enough to inform large-load decisions. In this analysis, a system-planned portfolio reduces system costs by $300 million per year, or $28 per household, when the full data center load materializes. The savings are even higher when load falls short of expectations — $400 million per year of system costs, equating to $52 per household. The challenge is that conventional IRP cycles take years, while data center customers often need power on much shorter timelines.

Utilities and commissions can address this gap by updating resource plans more frequently, maintaining current assessments of marginal resource needs, and allowing targeted procurement between major planning cycles while preserving core safeguards: comparison of alternatives, transparent decision records, and evaluation of system-wide costs and risks. Missouri’s annual IRP update process offers one useful model because it allows stakeholders to track changes in the preferred plan, uncertain factors, acquisition strategy, and system conditions between full planning cycles.

Prioritize least-regrets investments

When future load is uncertain, planners should give greater weight to investments that remain valuable across multiple futures. Least-regrets planning evaluates both sides of forecast risk: building too little can create reliability concerns and force emergency procurement, while building too much can leave customers paying for underutilized assets.

In this analysis, optimized portfolio cases consistently build at least 696 MW of gas, 4,341 MW of solar, 327 MW of battery storage, 250 MW of long-duration storage, and 1,085 MW of demand flexibility by 2030, whether or not the full data center load arrives. That overlap indicates a set of cost-effective resources that retain value under a wide range of outcomes.

The same principle applies beyond generation: transmission upgrades, interregional transfers, virtual power plants, demand flexibility, and grid-enhancing technologies can all increase the system’s ability to preserve optionality and serve uncertain load at low-cost, while avoiding large, irreversible, infrastructure commitments.

Stage procurement as uncertainty resolves

Utilities can reduce overbuild risk by sequencing procurement instead of making every resource decision up front. Some demand is firm enough to justify near-term action, while other additions should be conditioned on evidence that the load is becoming more certain. Commissions can approve an initial tranche of resources tied to verified need, then condition later tranches on signed electric service agreements, deposits, collateral postings, interconnection milestones, construction progress, transmission availability, updated forecasts, or reserve margin needs. This type of process prioritizes modular resources that can be expanded incrementally, making their option value visible to the system.

Colorado’s Just Transition Solicitation offers one model for this kind of sequencing, combining a base RFP, transmission portfolio filings, a supplemental RFP, an incremental need pool, and subsequent resource plan filings. For large-load customers, staged procurement can preserve speed while aligning infrastructure commitments with project maturity.

Allocate remaining risk through tariffs and contracts

Even with better forecasting, faster planning, least-regrets investments, and staged procurement, some risk will remain. Tariffs and contracts can determine how that residual risk is allocated, ensuring large customers bear an appropriate share of the costs they drive through mechanisms including minimum bills, minimum demand or energy commitments, collateral requirements, exit fees, long-term contract terms, capacity reservation charges, and capacity reassignment provisions.

RMI and Advanced Energy United’s Large Load Tariff Principles and RMI’s Electricity Affordability Toolkit both identify these tools as ways to limit the risk that ordinary customers pay for infrastructure built around uncertain large-load requests. Recent proceedings in Ohio, Indiana, and Oregon show how these tools are already being used, including minimum payment obligations, long-term financial commitments, and exit-fee structures designed to reduce risks of underutilized or stranded assets.

Conclusion

The data center boom creates a real challenge for system planners. Moving too slowly could delay investment and economic development. Building too quickly could leave the system with expensive, underutilized assets if load arrives late, ramps slowly, relocates, or never materializes.

The analysis presented here highlights the importance of resource strategy. In the modeled scenarios, system-planned portfolios produce lower costs when data center demand materializes and reduce downside exposure when it does not. The difference is driven not by a single technology, but by a planning approach that evaluates a broad set of options and preserves flexibility under uncertainty.

Realizing these benefits will require planning processes that are better suited to emerging large loads. That includes better forecasting, more frequent planning updates, systematic evaluation of alternative portfolios, procurement strategies that can adapt as information improves, and tariff structures that allocate risks transparently.

Gas-fired generation may remain part of least-cost portfolios in some regions and under some load-growth scenarios. However, large customer-specific gas investments carry a distinct risk profile when they are built around uncertain demand forecasts. In that environment, a least-regrets approach asks a broader question: which investments continue to provide value across a range of plausible futures? The answer will differ across utilities and regions. But as demand growth accelerates, planning frameworks that emphasize flexibility, optionality, and system value are best positioned to reliably serve new customers while limiting risks for everyone else.

[1] Analysis results are presented in real 2026 currency.

[2] These results reflect only the generation component of rates. Transmission and distribution costs typically make up an even larger share of a customer’s total bill, and because those costs are also largely fixed, missed demand forecasts could affect them substantially. However, this model is designed primarily to estimate generation costs, not transmission and distribution infrastructure costs.