In 2023, data centers consumed approximately 176 TWh of electricity across the United States, roughly 4 percent of the nation's total consumption. By 2028, that figure is projected to reach somewhere between 325 and 580 TWh, a range that itself tells a story about how uncertain and how large this transformation is. Grid Strategies (2025) reported forecasts of roughly 166 GW of new peak demand over five years, with about 90 GW of that attributed to data centers, though independent analysts caution these utility-reported figures may be overstated by up to approximately 40 percent and should be treated as a planning range rather than a point forecast. Even at the lower bound of that range, this is the most consequential load shift the North American grid has faced in a generation. Understanding this compact, the mutual dependency between the AI economy and the electricity system, is not background reading for a working energy professional in 2026. It is the operating environment.

What the Compact Means: Two Systems That Need Each Other

The word "compact" is chosen deliberately. A compact is a mutual agreement with obligations flowing in both directions. The AI economy needs power: enormous, reliable, always-on power for training runs, inference serving, and the cooling infrastructure that prevents those GPU racks from becoming expensive space heaters. A single large AI training cluster can consume 50 to 100 MW continuously for weeks. A hyperscale data campus serving inference at scale can consume hundreds of megawatts around the clock, every day, at load factors that most utility planners have never planned for in a commercial customer.

The grid, in turn, needs AI. Not as a future capability or a strategic investment hedge. Right now, in 2026, the load curve has broken in ways that traditional forecasting methods cannot handle. The ARIMA model trained on fifteen years of smooth residential and commercial growth has never seen a 400 MW step load arriving overnight when a data-center commissioning milestone is reached. The topology optimization system that was calibrated for a specific network congestion pattern has never managed the power flows associated with a 1 GW campus cluster. The interconnection queue study tools that were designed for one or two hundred requests per year are now processing thousands, with the queue backlog at over 2,060 GW at end-2025. The grid needs AI tools to forecast, optimize, and process the workload that the AI economy creates.

This is not a vendor pitch. It is a structural dependency that neither side can escape or ignore. A hyperscaler building a 1 GW campus in a power-constrained market needs a utility that can deliver that power reliably, on a schedule compatible with a data-center build timeline that moves in eighteen-month windows, and with a tariff structure that both parties can defend to their respective stakeholders. The utility serving that campus needs forecasting tools sophisticated enough to model a load that ramps from zero to 800 MW over eighteen months and then runs at 92 percent load factor essentially forever. The compact is not aspirational. It is the reality that both industries are navigating in real time.

The compact is also defined by tension, and that tension is important to name explicitly. The AI industry needs cheap, abundant, reliable electricity. The electricity industry needs cost recovery for the massive infrastructure investment that serving AI-scale loads requires, amortization of capital over regulated rate-base timelines, and rates that are just and reasonable to all customers, including the residential ratepayer whose bill should not be inflated to subsidize the computing bills of trillion-dollar technology companies. Navigating that tension is the most important policy and regulatory work of the decade in the energy sector, and it is work that will require the skills developed throughout this entire program.

The Scale of the Load Surge: Numbers You Must Understand

Treat the following numbers as orders of magnitude to verify against current sources, not as precise forecasts to cite without context. The underlying direction is clear and consistent across a wide range of analyst and research sources, even where the specific figures vary. What matters for a planning professional is the order-of-magnitude implication, not the precise decimal.

U.S. data-center electricity consumption was approximately 176 TWh in 2023. This represented roughly 4 percent of total U.S. electricity consumption in that year. The 325 to 580 TWh range for 2028 would represent somewhere between 7 and 12 percent of projected total consumption, depending on which growth assumptions you apply to the overall grid. That upper bound is roughly comparable to the total electricity consumption of several large states, added to the national demand picture in five years.

The peak demand framing is operationally more relevant than the energy framing for most planning and operations professionals. The utility-reported forecast of approximately 90 GW of data-center-attributed peak demand growth in five years, part of the broader 166 GW total growth figure from Grid Strategies (2025), is what drives transmission infrastructure sizing, reserve margin requirements, and the interconnection queue backlog. Analysts who have reviewed the underlying survey methodology caution that these self-reported utility projections may embed double-counting and optimistic assumptions, potentially overstating actual growth by up to 40 percent, so planners should use a range rather than a single point. Even so, adding tens of gigawatts of a specific type of peak demand, demand with a load factor of 85 to 95 percent rather than the 35 to 45 percent of a typical residential mix, changes the fundamental economics of capacity planning in ways that the current IRP models were not designed to capture well.

The load factor difference deserves a worked example. Consider a 400 MW data center connected to a single 345-kV substation, running at 90 percent load factor. That substation sees 360 MW of average load, 24 hours a day, 365 days a year. Contrast this with 400 MW of residential peak demand, which at 35 percent load factor represents about 140 MW of average load concentrated in a few hundred peak hours per year. The data center requires 2.6 times more energy per megawatt of peak demand than the residential customer. The generation capacity needed to maintain planning reserve margins scales with peak demand, but the fuel and operating costs scale with energy consumption. A grid serving a data-center-heavy load mix needs more baseload generation with high capacity factors, not more peaker units that run a few hundred hours per year. This is a structural shift in generation mix economics that ripples through every IRP and resource procurement decision.

The FERC and NERC Response to an Unprecedented Load Shift

FERC's 2026 rulemaking on loads over 20 MW is the federal regulatory system's response to a grid that was not designed for this scale of individual load connection. The historic interconnection policy framework was built almost entirely around generators: how do you safely study, impact-assess, and interconnect a new power plant into the transmission system without compromising the reliability of existing users? The load side of the bus was assumed to be diffuse, predictable, and roughly proportional to the number of customers served. A 400 MW data center connecting to a single 345-kV substation in a market that had never seen a commercial customer of this scale is neither diffuse nor predictable by traditional methods. The FERC rulemaking is essentially the first national reset of load interconnection policy, and it was driven by a specific technology sector choosing specific locations for specific infrastructure investment reasons.

NERC's parallel response, the Computational Load Entity category committed in the March 2026 FERC filing with standards delivery committed for December 31, 2026, goes further in its implications. The CLE category makes large compute loads registered grid actors with reliability obligations. This is not cosmetic change in taxonomy. A registered entity has legal obligations to provide accurate and timely load data, to participate in reliability studies when their facility's behavior is relevant, to implement specific operating procedures, and potentially to implement demand response capabilities that the bulk electric system can call on during emergency conditions. The accountability framework that has governed generators and transmission operators for decades is being extended, for the first time, to the load side of the balance.

NERC's Level 3 Alert issued in May 2026 around the data-center load growth was the clearest public signal that NERC views this load surge as an active reliability risk requiring immediate attention from registered entities. A Level 3 Alert is not a planning notice or a future-dated concern. It is NERC's highest-urgency notification mechanism, used when reliability threats are immediate and widespread enough to require prompt action. The fact that an AI-economy load phenomenon triggered a Level 3 Alert in 2026 confirms that the compact's grid consequences are a live operational issue, not a theoretical future planning scenario.

The Forecasting Breakdown: Where Your Models Are Failing Right Now

The load forecaster sitting at a planning workstation in 2026 in a territory experiencing significant data-center load growth faces a problem that would have seemed hypothetical in 2020 but is operationally urgent today. The step-load problem. A 400 MW data center connecting to your transmission system does not smooth in over several years the way a growing commercial district does. It connects under a defined commissioning schedule, typically ramping from zero to 50 percent of design capacity in the first six months and reaching full design capacity within twelve to eighteen months. After that ramp period, it runs at near-constant high load essentially continuously. Your ARIMA model, your regression-based seasonal decomposition, every statistical model trained on historical data that predates this connection, has never seen this pattern and cannot extrapolate to it reliably. The model's confidence intervals during the ramp period may look normal while the actual forecast error is far outside its historical calibration range.

The AI forecasting advantage in this environment is real and documented: approximately 1 to 2 percent MAPE day-ahead for well-implemented AI models versus 3 to 5 percent for traditional ARIMA and regression methods, on historical data representing typical operating conditions. But this advantage depends critically on the model having been trained on data that includes the new load regime. In the transition period, when the data center is commissioning and ramping but the model has not yet seen a full year of the new regime at this facility, even AI-based forecasting is operating with limited relevant training data for this specific input condition. The transition period is the period of maximum forecasting risk, and it is exactly the period when the forecasts being produced are driving procurement decisions worth hundreds of millions of dollars.

The right operational response is not to wait passively for the training data to accumulate. It is to incorporate data-center load ramp curves directly as exogenous inputs to your forecast model, treating the commissioning schedule as a known driver rather than an unknown pattern that the model must infer from the demand signal alone. This requires proactive coordination with large-load customers: obtaining their commissioning timeline milestones, their anticipated ramp rate, and their target operating load level under a formal data-sharing agreement. Most data-center developers are willing to provide this data to their serving utility because a well-informed utility provides better and faster service. The coordination is commercially rational for both parties.

In territories where multiple data-center interconnections are occurring simultaneously, the most effective forecasting architecture separates the mass-market forecast from the large-load pipeline forecast. The AI model trained on historical mass-market data continues to forecast the temperature-sensitive residential and commercial load that it handles well. A separate large-load pipeline model handles each large load above a defined threshold as an exogenous step-function driver, with a commissioning schedule, a load ramp curve from industry benchmarks, and an uncertainty range. The total forecast combines the two components. This two-component design prevents the mass-market model from being distorted by step-load events it was never designed to handle, while ensuring that the large-load additions are incorporated accurately.

Tariff and Cost Allocation Challenges: The Decade's Most Contested Regulatory Question

The most contentious political-economic dimension of the compact is the question of cost allocation: who pays for the grid infrastructure needed to serve large data-center loads, and in what proportions? When a 400 MW data-center campus requires a new 345-kV transmission line, a new substation, reactive power support equipment, and transmission protection upgrades, the total capital investment can reach several hundred million dollars. The data-center developer and its utility serving neighbors have competing legitimate interests in how those costs are assigned.

The data-center customer's argument is that any infrastructure that also benefits other customers, by improving contingency coverage in the area, reducing locational congestion costs, or providing a new pathway that the long-term transmission plan would have built eventually anyway, should be socialized across the rate base rather than assigned entirely to the causer. There is genuine merit in this argument for the subset of infrastructure that provides real system-wide reliability benefit. The counterargument from existing customers is that infrastructure built primarily to serve a single commercial customer generating substantial profits from cloud computing should not be subsidized by residential ratepayers who will see higher bills to fund a grid upgrade that primarily serves a technology company's bottom line. This argument also has merit when the infrastructure truly has minimal value to existing customers.

State utility commissions in data-center-dense states, particularly Virginia, Texas, and parts of the Pacific Northwest, are actively working through this tension in dozens of open dockets in 2026. The AI tools most directly relevant to this regulatory workload are not grid operations tools. They are regulatory-analytics tools: document summarization to compress lengthy commission orders and intervenor filings into action-relevant summaries; pattern extraction to identify how comparable jurisdictions are resolving analogous cost-allocation questions; and data-request response tools to manage the enormous discovery burden that major rate cases involving large-load customers generate. A regulatory affairs team tracking forty simultaneous state proceedings on related questions needs AI analytical support; human reading capacity cannot keep up with the document volume.

There is a second, more constructive dimension of the tariff question: the design of flexible-load tariff provisions that compensate data centers for participating in demand response and operating as reliability assets rather than treating them purely as passive loads. The Emerald AI and National Grid flexible-load partnership is the production demonstration that large compute loads, with the right coordination technology and tariff incentives, can modulate their consumption in response to grid conditions and provide flexibility services comparable to a demand-response aggregator or a peaker unit with a fast ramp rate. A data center that can reduce load by 20 to 30 percent for two to four hours in response to a grid emergency is providing a reliability service with real avoided-cost value. Designing a tariff that compensates this flexibility at a rate that is commercially attractive to the data-center operator, transparent about the value calculation, defensible before a commission as just and reasonable, and aligned with NERC reliability obligations, is one of the most technically and commercially sophisticated tasks in the current utility regulatory environment.

The Reliability Arithmetic: What 90 GW of New Load Means for Grid Infrastructure

Even the lower-bound scenario for data-center load growth requires a substantial expansion of the full infrastructure stack: transmission capacity to deliver generation to where the load is located; planning reserve margins maintained above the new, higher peak; reactive power support to maintain voltage stability in the new load zones; and the real-time dispatch and control infrastructure to manage it all. A standard industry planning approach suggests that maintaining a 15 percent planning reserve margin above a 90 GW demand addition requires approximately 103 GW of new installed generation capacity. Given typical transmission losses and regional reserve-sharing requirements, the actual generation procurement need could reach 135 to 180 GW of new installed capacity to maintain existing reliability standards on that scenario. All of that generation must pass through an interconnection queue that already had over 2,060 GW of requests at end-2025, with a median request-to-COD time that had doubled to over four years.

The interconnection queue backlog is the primary physical constraint on how quickly the compact can be operationalized. Data centers need power from new, often clean, generation. That generation must interconnect through a queue that is measured in years per project. The technology companies developing these data centers are signing power purchase agreements with clean-generation developers who then enter the queue. The queue's four-year median timeline is longer than most data-center build-and-commissioning cycles. This creates a gap period during which new data-center load is served from existing market resources, which in many regions means natural gas peakers, not the clean generation the technology company announced in its sustainability report. Shortening the interconnection queue is therefore not just an administrative efficiency goal. It is a reliability-critical and decarbonization-critical infrastructure problem at national scale.

AI queue-study automation is directly relevant here. The program has covered queue-study automation in multiple lessons across the L1 to L5 curriculum, and the case for it at the level of this lesson is strategic rather than operational. If median interconnection study time can be compressed from four years to one or two years through AI-assisted completeness checking, systematic study-report drafting, and cluster analysis that processes related projects simultaneously rather than sequentially, the generation needed to serve data-center load connects two to three years faster. That acceleration has direct implications for reserve margin adequacy, for the clean-energy-transition timeline, and for the competitiveness of utility service territories as data-center investment destinations.

The Energy Irony: AI Training the Grid That Powers AI

There is a recursive quality to the compact that deserves explicit acknowledgment, because ignoring it leads to planning assumptions that will be outdated within a planning cycle. The same AI training runs that are driving data-center electricity consumption growth are also developing the AI systems that utilities are deploying to manage the resulting grid stress. Training a large foundation model requires megawatt-hours, sometimes tens of thousands of megawatt-hours per training run for the largest models. The model that emerges from those training runs is then licensed to a utility or developed by a utility to forecast the load that the next generation of training runs will create. The AI that manages the grid is powered by the very grid it is managing.

The practical planning implication of this recursive relationship is that the pace of AI capability development creates a continuously moving target for grid planners. A load growth assumption that was calibrated carefully in January 2026 may be materially outdated by September 2026 if a major model architecture breakthrough triggers a new wave of training infrastructure investment. The hyperscalers respond to capability milestones, not to utility IRP cycles. The IRP cycle that updates every three to five years is structurally mismatched with the technology investment cycle that updates continuously.

Planning organizations that integrate real-time data-center pipeline tracking into their forecasting workflows, treating hyperscaler construction permit filings, power purchase agreement announcements, and interconnection request patterns as leading indicators of demand, will be better positioned than organizations relying solely on annual load-growth surveys and econometric models calibrated to pre-AI-era data. This does not require abandoning the integrated resource planning framework. It requires supplementing that framework with a large-load pipeline intelligence function that monitors and processes signals from the technology investment landscape on a near-real-time basis.

The grid does not power the AI economy as a passive service provider. It is the physical substrate of the digital age. The reliability professionals who maintain it are among the most consequential infrastructure stewards of this century, and the decisions they make in the next five years about how to serve, price, and plan for AI-scale loads will shape both the grid and the economy for decades.

Key Takeaways

• U.S. data-center electricity consumption is projected to grow from approximately 176 TWh in 2023 to 325 to 580 TWh by 2028; Grid Strategies (2025) utility-reported forecasts put data-center-attributed load at roughly 90 GW of a 166 GW five-year peak demand growth figure, though analysts caution these numbers may be overstated by up to 40 percent and should be treated as a planning range with real uncertainty, not a precise commitment.
• The compact is structurally mutual: the AI economy needs reliable and affordable power at unprecedented scale; the grid needs AI forecasting, optimization, interconnection-queue automation, and regulatory-analytics tools to manage the load surge that the AI economy creates. Neither side can operate effectively without the other.
• Data-center load shape, 85 to 95 percent load factor, 24-hour-a-day near-constant base load, step-function commissioning ramps, is qualitatively different from any load type that ARIMA and regression-based forecasting models were trained on. The step-load problem is the highest-urgency forecasting challenge in most high-growth utility territories right now.
• FERC's 2026 large-load rulemaking and NERC's Computational Load Entity category are the first systematic regulatory responses to load-side scale: loads over 20 MW now face national interconnection policy scrutiny; large compute loads are becoming registered grid actors with reliability obligations by December 31, 2026.
• The cost-allocation question, who pays for transmission and substation upgrades driven by large-load customers, is the decade's most contested utility-regulatory issue; AI regulatory analytics tools, document summarization and pattern extraction across multi-jurisdictional dockets, are essential for managing the workload of this proceeding landscape.
• Flexible-load tariff design, compensating data centers commercially for providing demand-response flexibility as a reliability asset, is one of the highest-value regulatory-strategy capabilities for energy professionals in 2026. The Emerald AI and National Grid partnership is the production proof of concept.
• The interconnection queue backlog, over 2,060 GW with median over four years to COD, creates a feedback loop: data centers need new generation, new generation needs queue processing, the queue is already overloaded by the wave of prior requests. AI queue-study automation is a reliability-critical path shortener, not simply an administrative efficiency gain.