Why data center growth forecasts are essential to mitigating their impact on the grid

Much of the concern surrounding artificial intelligence is about power: the technology’s economic power to reshape work, the political power resulting from concentrated wealth, and the military power enabled by new forms of weapons. However, the rise of AI also impacts a more literal form of power: electricity, from the data centers that consume it.

Since ChatGPT was launched in November 2022, nationwide electricity usage has increased by more than 4 percent, greater than the previous 15 years’ worth of growth, causing many to question whether the benefits of the technology and its associated infrastructure will outweigh its costs.

Consumers are worried about rising energy costs and grid reliability, and environmentalists are decrying data center carbon emissions. One planned facility, the Meta Hyperion project in Louisiana, could consume as much energy as the entire state of Connecticut.

While experts can try to mitigate potential impacts, addressing these concerns comes with its own challenges: They will need estimates of how many data centers will be built, where they will be located, and how much electricity each will demand. Such information will allow officials to accurately forecast data center growth and will be key to moving from a reactive to a proactive strategy.

There is no doubt those estimates are difficult to come by, but getting them wrong will lead to consequences beyond the data centers themselves. Over-forecasting demand can lead to costly overinvestment and additional burdens on ratepayers, while under-forecasting can increase the risk of blackouts.

Multiple sources of uncertainty—gaps in public information, a rapidly changing market environment, and unknown policy futures—complicate forecasters’ attempts to reduce these risks and contribute to forecasting errors. Despite these challenges, experts and policymakers can still take steps to protect ratepayers, promote information transparency, and plan for the wide range of impacts data centers may have on the grid. Upfront deposits and safeguards like “take-or-pay” provisions can align the financial risks of data centers with ratepayers. Experts could also shift from pursuing a single “most likely” data center scenario to analyzing a wide scope of possible futures to capture least regrets solutions and mitigate the most impactful consequences.

A cautionary tale. More than four decades ago, an inaccurate electricity demand forecast led to the largest public bond failure in US history. In July 1983, the Washington Public Power System (WPPS) announced its inability to pay $2.25 billion in outstanding bonds. Bondholders, including ordinary citizens seeking a haven for retirement funds, saw the value of their investments drop nearly 90 percent immediately afterwards. The utility company did not finalize settlements until a decade later, and the aftershocks shadow conversations on utility planning to this day.

The bond disaster was, in part, due to the utility company’s inability to predict the future. The electric power industry relies on demand, or load, forecasts to guide investments. Load forecasts predict the amount of electricity required five, 10, or even 25 or 40 years in the future. They are important because power infrastructure requires years to plan, finance, and construct: about four years for large natural gas power plants, more than 10 years for the average overhead transmission line, and nearly two decades for the most recent nuclear power plant expansion. For example, to meet the energy needs of 2040, utilities need to begin planning today. Load forecasts are a critical part of those processes.

The Washington Public Power System bonds were intended to avert an energy disaster.  Forecasts published in the late 1960s and early 1970s predicted the Pacific Northwest would experience serious electricity shortfalls by the 1980s. Their reasoning was not unsound: The region’s electricity demand, fed by abundant hydropower, had grown seven percent annually for the previous 20 years. Forecasters assumed this growth trend would continue and rapidly exceed the available electricity supply.

In response to their forecasted shortfall, the Washington Public Power System began constructing five nuclear power plants to ensure it would have sufficient electricity supply to meet their demand. They were financed in part by the municipal bonds.  Unfortunately, the load forecasts driving these investments were inaccurate.

Using historic growth trends, even aggressive trends, can be reasonable. Under-forecasting load can lead to infrastructure under-investment, degraded system reliability, and potentially dangerous blackouts if actual demand exceeds the grid’s ability to deliver electricity. Officials saw these effects in the summer of 2020. California experienced electricity demand above their planning targets due to unprecedented heat waves across the Western United States. This unforeseen demand, in combination with other factors, prompted grid operators to order utilities to cut off consumers from power on August 14 and 15 of that year. Nearly half a million people lost power during these periods.

Over-forecasting demand carries its own hazards, however. It prompts utilities to undertake costly infrastructure investments with little return. Without sufficient demand, power plants may be abandoned mid-project, power lines to deliver the electricity may go underutilized, and debt financed against future electricity sales may sour if those sales fail to materialize. Ratepayers can be caught paying for infrastructure that provides little benefit to the public, and investors could receive pennies on the dollar.

Load forecasts aim to guide industry planners between these twin hazards, avoiding either over-investment or shortfall, and enabling utilities to deliver electricity to consumers safely, cheaply, and reliably.

For the Washington Public Power System, the expected demand never materialized. The energy shocks of the mid-1970s broke the historic trendline, and actual electricity demand fell far short of earlier forecasts. Only one of the five nuclear plants was completed. To this day, part of the monthly electric bill for some Pacific Northwest ratepayers is allocated to pay for reactors that never generated a single watt of electricity. This was, in part, due to an incorrect electric load forecast.

Today, electricity forecasts predict unprecedented surges in demand driven by energy-hungry data centers. Forecasters in Texas, a prime data center destination, estimate their annual energy demand will more than double between 2025 and 2035; this is equivalent to adding over two Californias of electricity use in the next 10 years. US utility five-year spending plans have increased by 20 percent since last year alone. After almost two decades of nearly flat electricity demand, upwards-pointing load forecasts are driving massive infrastructure investments.

Should these forecasts be trusted? Data centers are expected to be the primary driver of US demand growth over the next five to10 years. Modeling from the Electric Power Research Institute, an independent organization that conducts research on electricity, suggests data center electricity demand growth could outpace all other factors and, in highly impacted markets like Virginia, consume nearly 60 percent of all electricity by 2030. However, the range of possible demand futures is wide; the same analysis suggests the upper bound of data center growth potential is more than twice as large as the lower bound.

While demand forecasting is inherently uncertain, forecasting data center growth is especially challenging. Forecasting begins with understanding the current level of demand, but no complete national inventory of data centers and their electricity use currently exists. Many utility companies do not publicly disclose granular load data and until recently did not track data centers as a specific category of customer. Forecasters are forced to start with fragmented and incomplete data. Better forecasts will require better data, and that data must come from somewhere.

Estimating future growth challenges. Utilities typically use an interconnection queue to process and track applications of those who wish to connect to their grids. When a data center submits an application to the queue, the utility studies its expected impacts on the grid. Utilities also use the queue to inform their demand forecast; if a data center applies to connect 100 megawatts of demand in 2030, then that can become a data point in the utility load forecast.

Unfortunately, not all the applications into the queue will result in actual realized demand.  Applicants could withdraw from the queue because costs become prohibitive or the project financing falls through. Some developers will “queue shop” by applying for interconnection in multiple utility territories and withdrawing applications that move too slowly or become too costly. Others could encounter construction issues or might not use the full amount of electricity they originally planned.

The utility AEP Ohio is an illustrative example. After implementing additional requirements on data center interconnection applicants in 2025, it saw queue volume drop from 30 gigawatts to 13 gigawatts. For reference, the withdrawn demand is about equal to the generation capacity of all of Oregon’s power plants combined.

Policy shifts can also significantly impact data center growth projections. Many data centers site themselves in specific states to take advantage of data center-specific tax incentives. For example, Georgia offers 100 percent sales and use tax exemptions for qualified data center equipment. Should states repeal their incentives, as at least nine have considered, this may shift demand to other regions. New York state is even considering a full moratorium on large data center construction for a one-year period.  These shifts can render years of forecasting effort obsolete with a single stroke of a governor’s pen.

Last, at the broadest level, it is understood that much of the current forecasted data center energy demand is speculative. Facilities are being built in anticipation of future artificial intelligence computing needs. The true pace, and ultimate adoption level, of AI is unknown. Continued technology breakthroughs could drive electricity demand higher than today’s most aggressive forecasts, while a market collapse could precipitate an artificial intelligence “winter,” which has happened twice before. Or, society may be at the top of the Gartner Hype Cycle, an adoption model for some emerging technologies, waiting to plunge off the Peak of Inflated Expectations into the Trough of Disillusionment before climbing the much gentler Slope of Enlightenment. As in the pacific northwest 40 years ago, people could see abandoned nuclear reactors silhouetted against the West Texas skies.

Moving forward into an uncertain future. Data centers stand to be the most important driver of nationwide electricity demand, and inaccurate forecasts could lead to disruptive over- or under-investments in electricity infrastructure. How then do experts move forward, given the deep uncertainties that plague these forecasting efforts?

First, they should move past thinking in terms of a single “most likely” forecast. The sources of forecasting uncertainty are too strong to identify the “most likely” future with confidence. Instead, experts should consider a wide range of possible growth futures: some lower, some higher, and some in the middle. Analyzing the consequences of these futures can highlight common risks or opportunities that inform “least regrets” solutions.  For example, if seven out of nine scenarios point to a need for a new power plant, then it might be a safe bet. If the remaining two futures warn of dire consequences should the expected demand fail to materialize, then the utility can take measures to mitigate those specific risks. There may still be some hard tradeoffs, but officials will be equipped to make more sound decisions after weighing all reasonable options.

Second, experts should align the financial risks borne by data centers with those borne by the utilities that serve them. Data centers can be built or modified far more quickly than power plants or transmission lines. If a utility builds infrastructure to meet a data center’s projected demand, only for the data center to shrink its footprint or cancel the project altogether, the utility and its ratepayers could be stuck paying for that underutilized or abandoned infrastructure while the data center carries none of the investment risk.  Requiring upfront deposits from data centers applying for interconnection can screen out speculative applicants. Safeguards like “take-or-pay” provisions, where an applicant is liable to compensate a utility for power it had promised to consume but did not, can help guarantee revenue. These types of provisions shift the risk burden back towards the data centers, protect ordinary ratepayers, and provide greater certainty for infrastructure investment.

Lastly, officials should encourage data transparency. Utilities can build their institutional knowledge and credibility by sharing their forecasting results, with as much detail and fidelity as permitted by law and regulation, with the broader community. Data center developers and operators can share details about historical development timelines, key challenges, and facility characteristics. Google already releases quarterly efficiency reports for much of its data center fleet; this is invaluable to understanding how data center energy needs change throughout the year.

AI poses a foundational challenge to our electric grid. By understanding the when, where, and how much, people can chart a path towards the best future for themselves, their society, and the world.  Humans are not powerless.