You Can’t Just Plug in a Data Center

To meet booming AI-driven energy demand, utilities and regulators must adopt better rules, better forecasts, and better commitments

Source: https://www.aboutamazon.com/news/aws/aws-data-center-inside

Most of us experience electricity as a subscription service. We move in, call the utility, set up an account, and expect the electrons to arrive. The power system is one of the great coordination achievements of modern life precisely because, for most of us, it disappears into the background. When it works, we do not notice it. When it fails, we notice nothing else.

That mental model does not scale. A data center requesting hundreds of megawatts is not another customer signing up for service. At that scale, electricity service becomes an infrastructure negotiation—and the problem is not just how much power data centers consume. The problem is how a large, fast-moving, uncertain new customer becomes a permanent enough fixture around which to build power systems.

This distinction becomes more important as artificial intelligence drives a surge in data center development and an expectation of ensuing demand growth. Public debate has focused on quantities: how much electricity data centers will use, how much new generation we will need, how much bills might rise. Those questions matter, but their focus on the shortage of electricity infrastructure misses a deeper problem: a shortage of institutional infrastructure.

Beyond Plug-and-Play

Large new electricity demand does not simply appear on the grid. It has to be forecast, studied, planned, financed, physically connected, and then delivered reliably hour after hour. These steps need much more than engineering; they require credible commitments.

A data center developer may want the utility to reserve hundreds of megawatts of future capability, while the utility may have to build substations, transmission upgrades, and transformers before the data center is fully operating. Regulators must then decide whether those investments are prudent, who should pay for them, and what happens if the project is delayed, uses less electricity than expected, or never materializes at all. The grid-owning utility has to make durable investments under uncertainty (and so does the hyperscaler). Better forecasting can reduce that uncertainty but not eliminate it. The harder institutional question is how to write rules and contracts that make investment possible without asking ordinary customers to become involuntary insurers of speculative demand.

Historic trends. New uncertainty.

Data centers are not the first large customers utilities have ever served. The electricity system has a long history with energy-intensive industrial loads—steel mills, aluminum smelters, chemical plants, auto factories—that shaped the 20^th-century economy and power grid. These customers received special contracts, industrial rates, and sometimes interruptible service in exchange for paying lower prices. But many historical large loads were anchored to physical production processes, local resource bases, and long-lived industrial assets, arriving on timescales that aligned reasonably well with utility construction cycles.

Data centers scramble that model by combining features that collectively strain the inherited utility planning process. They arrive in large concentrated increments, with a single campus demanding hundreds of megawatts and a handful of projects transforming a utility’s load (demand) forecast. They are fast, with developers caring intensely about “speed to power” while transmission upgrades, substations, and generation procurement run on slower clocks. They are mobile, shopping across utility territories and grid regions, with each site inquiry showing up as potential future load before a developer commits to building anywhere. They are uncertain—some backed by major companies with real demand, others contingent on financing, permits, chips, or favorable tax treatment. And they may be flexible, though not automatically, with some computing workloads more amenable to shifting across time than others.

A new paper on large-load tariffs and load forecasting, by Angela Navarro and Molly Knoll for the National Association of Regulatory Utility Commissioners, captures the key shift: Historical load forecast errors were driven by population growth, marginal changes in end-use consumption, and macroeconomic factors that were diffuse and incremental. Today’s large, discrete loads are localized, infrastructure-intensive, and often seeking service faster than utility planning cycles can accommodate. The old forecasting world was about gradual demand trends. The new forecasting world is about project-level shocks. Poor load forecasting, welcome to your main-character era.

When the forecast becomes the problem.

For much of the 20^th century, load forecasting was necessary but quiet, most noticeable when something went wrong, because demand growth was generally steady. The data center boom changes that. In regions with substantial data center interest, the load forecast becomes contested terrain.

The underlying difficulty is the phantom load problem. If developers request service in multiple locations before deciding where to build, utilities and regulators may see more proposed demand than will ever materialize. PJM, the grid operator in the mid-Atlantic region, has responded by requiring firmer commitments for near-term forecast years while treating longer-term projects as less certain. ERCOT in Texas adjusted its large-load forecasting assumptions after observing an average project delay of 180 days and data centers’ consumption of less than half the capacity originally requested for study.

A forecast is easy to inflate when someone else bears the cost of relying on it. Forecasting is both technical and strategic, with developers wanting to reserve capacity, utilities potentially expanding their rate base, economic development officials courting investment, and consumer advocates all having stakes in the outcome. In such an environment, a load forecast is a claim on future infrastructure, and when forecasts are wrong, the consequences are expensive. Transformers, substations, and transmission upgrades cannot be returned with a receipt and a sheepish apology. Bad forecasts become bills.

The real problem is credible commitment.

This observation brings us to the central institutional economics question: Who should bear the risk created by uncertain large-load forecasts?

The easy answer says that big technology companies should pay more because they are large and profitable. This claim has political force but is not quite the right analytical frame. The better question is what commitments a large customer should make before everyone else is asked to commit resources based on the forecast.

When a utility constructs infrastructure to serve a large new customer, it makes investments that are durable, capital-intensive, and difficult to repurpose. The relevant concept from transaction cost economics is asset specificity: A transformer, substation, or transmission upgrade built to serve a particular large load in a particular location may have little value if that load never materializes. This characteristic transforms a forecasting problem into a contracting problem.

The structure is a classic setting for what are known as contractual hazards. One party must make a durable investment before the future is fully known; the other party has better insight into its own plans and commercial prospects, but it, too, has durable investments on the line. Once the investment is sunk, bargaining positions change. MIT professor Paul Joskow’s work on coal contracts showed that contract duration tends to increase when relationship-specific investments matter most; when one party must invest in durable assets to serve another, longer contractual commitments reduce the risk of being left exposed. Nobel laureate Oliver Williamson described the need for “hostages” to ensure credible commitment: A deposit, collateral requirement, minimum bill, or termination fee gives the customer something at stake, making its forecast more than cheap talk.

The relevant question is practical: What kind of contingent contract can make investment possible when the future is uncertain? When parties cannot know the future, they can still agree in advance on what happens under different future conditions. If the data center arrives on schedule, here is how costs will be recovered. If it builds more slowly, here are the minimum charges. If it leaves early, here is the exit fee. If it offers flexibility during grid stress, here is the compensation.

That is not punishment. It is governance.

Turning forecasts into commitments.

Properly understood, a large-load tariff is a rulebook for joining the grid when a customer is large enough to alter system planning, infrastructure needs, and cost allocation. These tariffs may require long-term contracts, minimum billing commitments, collateral, exit fees, phased ramp schedules, and notice requirements. The basic institutional function is consistent: Large-load tariffs convert forecasts into commitments.

The new National Association of Regulatory Utility Commissioners paper adds an important insight. Tariffs are more than customer-protection mechanisms—they can also improve forecasting. Minimum demand obligations, collateral requirements, phased load ramps, eligibility thresholds, contract lengths, and exit fees all contain information about the timing, magnitude, persistence, and credibility of large-load requests. Navarro and Knoll propose that regulatory commissions use tariff terms directly in load forecasting, applying them to baseline planning assumptions and sensitivity analyses, and updating forecasting assumptions as actual queue and realization data accumulate. This approach reveals that a large-load tariff is more than a price, it is a forecasting institution.

The components of a well-designed tariff are not random regulatory appendages. Long-term contracts address duration. Minimum bills address downside demand risk. Deposits and collateral bond the customer’s promise. Exit fees protect against costly early departure. Ramp and build schedules make timing uncertainty contractible. Curtailment provisions that lay out the conditions under which the large load customer will reduce its demand define what flexibility means operationally. The goal is not a complete contract specifying every possible future state; no tariff can perfectly anticipate AI demand trajectories, chip constraints, or capital markets. The more modest and useful goal is contingent contracting: specifying what happens under the most important foreseeable deviations from the forecast.

The experience of AEP Ohio illustrates this well. After regulators approved and the utility adopted revised data center tariff terms in July 2025—including 25-megawatt eligibility requirements, an 85 percent minimum billing obligation, 12-year contract terms, and termination fees equal to minimum charges for 36 months—the interconnection queue fell from 30 gigawatts to 13 gigawatts. The missing 17 gigawatts was not necessarily “phantom” in any simple sense. But once capacity requests became costly to maintain, the queue began to contain better information, filtering speculative entries before utilities built around them. Better load forecasting can tell us which futures are more likely. Large-load tariffs help decide who is responsible when the likely future turns out to be wrong.

Can the cloud learn to bend?

Requiring data centers to pay for the infrastructure they require is the necessary first principle. Existing customers should not subsidize speculative load growth. But stopping there misses the question of whether data centers might reduce the infrastructure they require by becoming more responsive to grid conditions.

Not all megawatts are created equal. A customer requiring firm, round-the-clock service imposes different costs from one able to reduce demand during scarcity, shift computing tasks across time, or coordinate operations with grid needs. Some workloads—AI training, batch processing, cooling systems—may offer flexibility that latency-sensitive inference cannot. This flexibility is not automatically valuable, however. It has to be defined, contractualized, and verified. A data center claiming flexibility should have to specify what that means operationally: how much load can be reduced, how quickly, for how long, and at what compensation. Flexibility becomes useful to the grid only when it is reliable enough to plan around, making this too a credible-commitment problem.

Large-load tariffs could become tools for institutional discovery, revealing which customers are firm, which are flexible, and which are speculative—making their characteristics visible and actionable rather than forcing public utility regulators to guess.

The larger lesson.

The data center boom did not create weaknesses in electricity regulation. It removed the luxury of ignoring them.

Load forecasts are no longer technical appendices to utility plans; they are central to economic development, reliability, affordability, and decarbonization. When a forecast becomes the basis for billions of dollars of infrastructure on both sides of a transaction, the assumptions behind it deserve scrutiny: which projects are included; how credible they are; what commitments customers have actually made; and how much projected demand is firm, flexible, speculative, or duplicative.

But forecasting is only part of the solution. Forecasts reduce but do not eliminate uncertainty. Once we better understand the likely timing, magnitude, and persistence of large new loads, we still need rules assigning responsibility when expectations prove wrong. Data centers are neither villains nor ordinary customers. They are large, sophisticated, fast-moving actors whose demand can reshape the electricity system, and good institutions should be able to tell the difference between those creating real value and those claiming capacity without credible commitment.

Better forecasting tells us which futures are more likely. Better contracting tells us what happens when the actual future is different. The first reduces uncertainty; the second governs the uncertainty that remains.

You cannot just plug in a data center. But with better rules, better forecasts, and better commitments, perhaps we can learn how to connect one.

First published at Dispatch Energy at https://thedispatch.com/newsletter/dispatch-energy/data-center-electricity-use-regulators-utilities/