AI Should Pull Its Own (Electric) Weight, and a Little Bit of Yours

Whether a data center increases or decreases electricity bills for everyone else depends on whether the data center pays more than what the utility spends to serve it. A data center’s electric bill, like every other customer’s, consists of marginal costs—supplying electricity, expanding the grid to accommodate the new loadand a contribution toward the fixed and shared costs of maintaining and running the grid, e.g., building and maintaining common transmission lines, energy efficiency programs, etc.

The answer to the first question (what the additional costs of serving a data center are) isn’t exactly straightforward. The answer to the second question (how much a data center ought to pay for supporting common grid infrastructure) is a policy question, not a math problem with a precise answer.

So it’s no surprise that there is a lot of confusion about how data centers impact rates for all other customers. Claims that data centers are causing household bills to go up are met with counter claims that data centers are doing their bit to keep electric rates in check. Legislators, consumer advocates, and data center lobbyists are all putting forth competing proposals, some better than others, on what a fair data center electricity rate looks like.

A recent study shows that states with rising electricity sales tend to have decreasing average rates. Doesn’t this suggest that data center load growth will reduce rates for everyone else?

Although this finding is consistent with the economics of the electrical grid, it doesn’t necessarily apply to data centers. Typically, increasing sales reduces average rates because when fixed grid costs are spread across more kilowatt-hour sales, the average rate per kilowatt-hour decreases. This is why heat pumps and electric cars put downward pressure on rates. However, this doesn’t justify the claim that data centers will reduce rates because they increase sales.

First, correlation isn’t causation. States where electricity is already cheap are also more likely to attract industries that consume a lot of it. Causality may run the other way—low prices induce more consumption, rather than high sales causing low prices. The study does not test for that.

Second, even if the relationship is causal, data centers are not a typical load growth. The amount of electricity they consume and the speed with which they are coming online can change the nature of energy and capacity markets. Their sheer size also means that they could require costly upgrades to grid infrastructure, and the uncertainty regarding their future opens the possibility that some of these new investments may become superfluous.

Most important, even in instances where data centers put downward pressure on electricity rates, other consumers won’t see lower bills if utilities and regulators give data centers a free pass.

In the short run, data centers increase utility spending in a few ways. They increase electricity demand, which raises prices for wholesale energy, capacity, and renewable energy credits for everyone. These higher prices are a signal for new supply to come online, which takes time. When that happens, which depends on grid operators getting their act together, prices could return to prior levels or lower. Utilities with long-term contracts are protected from these increases in market prices to some extent; most utilities still depend on short-term markets for energy, capacity, and renewable energy credits.

Data centers often require the utility to expand the grid directly, such as new transmission lines and infrastructure to connect these hyperscalers. Utilities may also need to expand transmission elsewhere to relieve congestion. But this category of spending is harder to attribute to data centers because utilities often expand transmission for many reasons—integrating renewables, meeting electrification load—and can plan upgrades to serve multiple goals at once.

Data centers, like all other loads, depend on existing grid infrastructure. A significant portion of every customer’s monthly bill covers the fixed costs of this common grid: tree trimming, transmission infrastructure repayment, climate policy compliance, and so on. These costs don’t change when the data center comes online or goes off. How they get divided among data centers and existing customers is a policy question, not a technical one. This means that utilities, state, and federal regulators can decide how much data centers should pay for these costs to make bills more affordable for all other customers.

In some parts of the grid and to some extent, yes, but not universally. Many factors, including general inflation, have driven rate increases nationally, and exact causes are specific to each utility’s grid, customer base, policies, and market conditions. Determining the extent to which data centers are responsible requires a case-by-case analysis.

That said, at PJM—the large mid-Atlantic and Midwest grid where data center concentration is highest—load growth from data centers has put upward pressure on both energy and capacity markets. Although capacity market spikes have drawn the most attention, evidence suggests that rising energy prices have had a bigger impact. Because energy markets run 24 hours a day and 365 days a year, even modest increases in energy prices can have a large impact on customers.

Not necessarily. New generation takes a notoriously long time to site, permit, and build, like with new buildings in a major city. Any new contracts that data centers sign today are likely with resources that would have come online anyway. Just like gentrifiers can outbid existing residents for nicer and newer homes and raise rents on existing residents when they move to the neighborhood, data centers are already outbidding utilities for new energy contracts and putting upward pressure on energy markets. 

Unless data centers bring additional generation online—generation that would not have been built but for the data center’s contract—energy and capacity prices will increase relative to a scenario in which the data centers didn’t exist.

Requiring data centers to bring additional new resources online rather than contracting with resources that would have come online anyway can help keep grids reliable and keep energy and capacity market prices in check. Requiring those new resources to be clean avoids locking in new fossil fuel plants to serve data center load.

Demand flexibility is critical to ensuring that data centers can be accommodated on the grid without making expensive investments. Regulators, utilities, and grid operators should capture as much demand flexibility as they can. But it’s no silver bullet. It would alleviate pressure on capacity prices, and it would also enable data centers to come online quicker. But it won’t solve the larger problem of rising energy prices or, for that matter, renewable energy credit prices. And utilities will still have to spend to interconnect these data centers.

At a minimum, data centers should pay all directly attributable costs plus some share of fixed grid costs. Directly attributable costs include the cost of buying energy and capacity to serve their load and whatever the utility spends to expand the grid to accommodate them. This, however, cannot account for the fact that data center demand can increase energy and capacity prices for everyone, but asking data centers to bring their own capacity helps.

A reasonable floor to determine shared grid and policy costs is to make data centers contribute in the same manner as other large commercial and industrial customers do. Usually, when studies say that data centers are not shifting costs on other customers, they mean that data centers are paying more than the directly attributable costs, which are just one portion of this minimum expected amount. No one gets a gold star for doing the least they possibly could.

The most they are willing to pay, which may depend on where they are located and why. 

Not all data center loads are geographically flexible. Data centers that locate in high-cost states like California or Massachusetts are often there because they need to be close to their end users (this is often the case with data centers that require lower latency, which is another way of saying real-time use, as opposed to other facilities that have more flexibility in terms of proximity to users, such as for AI training or cloud storage). These facilities are relatively more inelastic to electricity prices, meaning that regulators in those states have more room to require them to contribute a larger share of fixed system costs—and thereby putting as much downward pressure on rates for all other customers as possible.

The picture is more complicated in places like northern Virginia’s “Data Center Alley,” where data centers locate primarily because costs are low and there is extensive existing fiber-optic network infrastructure already in place. There, the industry argues that higher charges will push valuable economic investment to neighboring states. The way to solve that is to set a minimum standard across states for how much data centers must contribute, either via interstate agreements or federal intervention. Without such a floor, locationally elastic data centers will simply site wherever costs are lowest, and those least-cost locations may also be the most carbon-intensive parts of the grid.

Without adequate protections in place, other customers end up holding the bill for stranded investments that utilities made on behalf of data centers. Utilities and regulators need to do three things. First, create a separate rate category, or customer class, for new data centers to make it easier to track utility investments for data centers and the extent to which utilities recover these costs from data centers via monthly bills. 

Second, institute a meaningful, nonrefundable entry fee or require data centers to pay at least a portion of the investments the utility makes on its behalf. And finally, setting an exit fee may help ensure that all remaining costs get recouped if the data center shuts down prematurely. These contractual obligations dissuade speculative data center load and reduce the planning uncertainty that has made electricity demand forecasting so difficult in recent years. American Electric Power, an Ohio utility, has a model tariff that includes these elements. And as expected, it helped get rid of some speculative data center demand.