Integrated resource planning is, specifically, the process through which regulated utilities determine how much generation capacity they need, when they need it, and how it should be sourced. Utilities designed the process around decades of relatively stable industrial load growth, seasonal demand patterns, and incremental changes in the customer mix. A utility could, with reasonable confidence, project its peak demand five to ten years forward and plan its generation and transmission investments accordingly. That confidence is, however, gone.
AI data centers are creating load growth patterns that are incompatible with the planning assumptions embedded in conventional integrated resource planning methodologies. The growth is too fast, too large, and too concentrated. Planning frameworks designed for a different era cannot absorb it. A single AI campus bringing 300 megawatts of load into a balancing area in 12 months represents a demand event that would previously have taken a decade of organic growth to produce. When ten such projects appear in the queue simultaneously, the planning frameworks simply break down. Utilities across the United States are, consequently, in the middle of a fundamental methodological revision. Most of the energy policy community has not yet fully absorbed it.
Understanding how and why conventional utility load planning fails when confronted with AI-driven demand is essential context for anyone operating at the intersection of AI infrastructure and energy policy. What utilities are doing to rebuild their frameworks, and what that means for data center developers and grid reliability, follows directly from that understanding.
What Conventional Integrated Resource Planning Was Designed to Do
Conventional integrated resource planning operates on a ten-to-twenty-year planning horizon. Utilities project load growth based on historical patterns, economic forecasts, and the expected evolution of their customer mix. They model the generation and transmission investments required to meet that projected load with a defined reliability standard, typically expressed as the probability of loss of load over a planning period. They submit those plans to their state public utility commission for approval and execute the approved investments over the planning period.
The methodology rests, however, on several foundational assumptions that AI data center load growth has invalidated. The first is that load growth is gradual and relatively predictable. Industrial load in the United States grew at roughly 1% to 2% annually over most of the past two decades, with deviations driven by broad economic cycles that were themselves relatively foreseeable. The second is that new large loads appear infrequently, and the standard interconnection study process, designed for loads in the tens of megawatts rather than hundreds, can accommodate them. The third is that the geographic distribution of load growth is diffuse, spread across a utility’s service territory in ways that allow existing transmission infrastructure to serve most new loads without major upgrades.
Why AI Load Growth Breaks the Planning Assumptions
AI data center load growth violates, notably, all three assumptions simultaneously. Growth is not gradual. Dominion Energy Virginia, the utility serving the highest concentration of data center capacity in the world, has seen its load forecasts revised upward by gigawatts in the space of months. Hyperscaler and AI infrastructure operators disclosing expansion plans are, specifically, the driver. The pace of revision is incompatible with the planning cycle. By the time a utility has completed an integrated resource planning update that incorporates AI load growth, the underlying data has changed significantly.
CenterPoint Energy disclosed in its April 2026 earnings that it has 12.2 gigawatts of firmly committed industrial load in its Houston Electric service territory. That figure represents, indeed, a planned 50% increase in peak demand by 2029. No conventional integrated resource planning model had, indeed, a scenario for that magnitude of demand growth in that short a period. The geographic concentration is equally challenging. AI data centers cluster in specific markets where power, fibre, and land converge. Rather than diffuse load growth across a service territory, utilities are seeing concentrated load additions in specific transmission corridors that create localised congestion without obvious conventional solutions.
The Forecasting Problem Is More Fundamental Than It Appears
The forecasting breakdown that AI data center growth has created is not, however, simply a matter of using better data or updating models more frequently. It reflects a deeper structural problem with how utilities receive information about future load growth. Under conventional planning practice, a utility learns about new large industrial loads through the interconnection application process. An industrial customer applies for interconnection. The utility conducts an interconnection study. The load appears in the planning model if and when the project proceeds to construction. Conventional practice models an approved interconnection as certain load.
AI data center developers have, specifically, disrupted this information flow in two ways. First, the lag between when a developer announces a project and when it submits an interconnection application can be months or years. Second, the conversion rate from interconnection application to actual load has been historically low. As Lawrence Berkeley National Laboratory’s Queued Up analysis found, only 13% of capacity entering interconnection queues between 2000 and 2018 was ever actually built. Utilities that plan against their full interconnection queue are overbuilding. Those that discount queue applications heavily are underbuilding. Neither approach is adequate.
How Utilities Are Revising Their Planning Frameworks
Several utilities have begun implementing changes to their integrated resource planning methodologies that reflect the specific challenges of AI-driven load growth. The changes vary across utilities and regulatory jurisdictions, but several common themes are emerging that represent the leading edge of a broader methodological evolution.
The first and most fundamental change is the introduction of scenario-based planning that explicitly models AI data center load growth as a separate demand category with its own assumptions, uncertainty ranges, and planning implications.
Rather than incorporating AI load into the general industrial growth forecast, utilities are developing AI-specific load scenarios. These span from conservative scenarios, in which a significant fraction of announced projects do not materialise on schedule, to aggressive ones, in which announced timelines are largely met. Planning against a range of scenarios rather than a single point forecast allows utilities to identify investments that are robust across scenarios and to stage capital commitments in ways that reduce the cost of being wrong. As we have covered in our analysis of how AI infrastructure is permanently restructuring the US power sector, the implications of AI load growth for utility investment planning extend far beyond near-term capacity additions to the fundamental economics of the regulated utility model.
The Challenge of Modelling Conversion Rates
The scenario-based planning approach requires utilities to make explicit assumptions about the conversion rate from interconnection applications to actual load, a parameter that planning models have historically treated as a binary variable. Under conventional practice, utilities model an approved interconnection as certain load. Evidence from the AI infrastructure buildout shows that this assumption is wrong.
The conversion rate from AI data center interconnection application to actual energised load is meaningfully lower than 100%, and utilities that model it as such are overbuilding. The appropriate conversion rate varies by market, by developer type, and by project stage, and determining it requires qualitative assessment of project readiness that most utility planning teams do not currently have the capacity or data to conduct systematically. Several utilities are, consequently, developing frameworks for tiering interconnection queue applications by project readiness, assigning different probability weights to loads at different stages of development, and updating those weights as projects progress or stall.
The development of those tiering frameworks is, in turn, creating a new category of information exchange between utilities and data center operators. Operators who provide utilities with detailed project readiness information, including power purchase agreements, equipment procurement status, and construction financing terms, are supporting utility planning in ways that benefit both parties.
Engaging Data Center Operators Directly
The second significant change is the development of direct engagement programmes between utilities and large data center operators that go beyond the standard interconnection application process. Several utilities, including Duke Energy in the Carolinas and Georgia Power in Georgia, have established dedicated data center teams that work directly with operators to understand their load growth plans, phasing, and operational characteristics before formal interconnection applications are submitted.
This direct engagement serves multiple purposes. It allows utilities to incorporate qualitative intelligence about operator plans into their load forecasts before those plans appear in interconnection queues.
At the same time, operators gain visibility into transmission constraints and grid conditions that affect their site selection and phasing decisions before they commit to specific locations. This process also creates a channel through which utilities can communicate capacity constraints and investment timelines to operators who might otherwise discover those limits only after submitting a formal interconnection application. The engagement model is, consequently, shifting the relationship between utilities and data center operators from a purely transactional one to a strategic partnership that serves both parties’ planning needs. As we have covered in our analysis of why US utilities are becoming the most powerful players in AI infrastructure, the utilities that have built the deepest developer relationships are attracting disproportionate AI infrastructure investment.
How FERC Interconnection Reform Is Reshaping Planning
The Federal Energy Regulatory Commission’s interconnection queue reform process, which produced Order 2023 in 2023, is reshaping the information flow that underlies utility load planning in ways that have received insufficient attention. Order 2023 introduced cluster studies, deposit requirements, and readiness milestones designed to reduce the number of speculative projects in interconnection queues and improve the signal quality of the queue as a leading indicator of actual load.
The intent is to make interconnection queue data more useful for planning purposes by filtering out applications that lack the financial and operational commitment to proceed.
Early evidence from markets that have implemented the cluster study methodology suggests that it is reducing queue volumes, though the magnitude of the effect varies significantly across regional transmission organisations. For utility load planners, a smaller and higher-quality interconnection queue is, consequently, more useful as a planning input than the historically inflated queues that included large numbers of speculative applications. The reform process is, however, still being implemented inconsistently, and its full impact on planning data quality will not be visible for several years.
The Load Profile Problem and Its Planning Implications
A second planning problem compounds the load growth forecasting challenge and has received less attention: AI data center load behaves very differently from the industrial loads utilities have historically planned around. These differences create reliability and grid stability challenges that conventional planning frameworks are not equipped to address.
Conventional large industrial loads consume power at relatively stable rates that change slowly and predictably. A steel mill ramping production affects its power consumption over hours. A conventional data center modulating between peak and off-peak operation changes its load over minutes. AI training and inference workloads create demand profile characteristics that neither of those historical analogies captures. As we have covered in our analysis of how agentic AI is creating a power demand profile that nobody designed data centers for, the shift from conventional to agentic AI workloads is making the load profile problem more acute, not less.
Ramp Rate and Frequency Regulation Implications
The ramp rate characteristics of AI training workloads create specific challenges for grid frequency regulation. A large AI campus initiating a training run can ramp from near-idle to full load in minutes, withdrawing hundreds of megawatts from the grid at a rate that significantly exceeds the ramp capability of conventional generation resources. Grid operators managing frequency in real time use a combination of automatic generation control, spinning reserve, and demand response to absorb load ramps. The reserve margins adequate for conventional industrial load profiles are not, in many cases, adequate for AI training ramp events at the scale and speed that modern campuses produce them.
Utilities and regional transmission organisations are, consequently, revising their reserve margin requirements and their methodology for counting demand response resources in reliability calculations. The revisions are not yet complete in most jurisdictions. The NERC guidance issued in April 2026 on data center load loss risks reflects the reliability organisation’s own assessment that current planning frameworks have not yet caught up with the operational reality of AI load at scale.
Where Transmission Congestion Is Already Visible
Northern Virginia, greater Phoenix, and the suburban corridors around major Texas cities are all experiencing transmission congestion. The current grid configuration was not planned for this level of AI data center concentration. Congestion creates real-time energy price premiums in affected areas. It increases the risk of transmission constraints that limit low-cost generation dispatch. In extreme cases, it creates reliability risks that require generation redispatch or demand curtailment to manage.
The Transmission Investment Mismatch in Practice
Transmission infrastructure takes, moreover, five to ten years to plan, permit, and construct. A utility planning transmission investments today is planning against a load forecast that may be substantially wrong by the time those investments are in service. Correcting that congestion through conventional transmission construction would take years and cost billions.
Accelerated Transmission Planning: What Utilities Are Trying
Several utilities are, consequently, experimenting with accelerated planning and permitting processes for transmission investments directly associated with large new industrial loads. These processes compress the conventional planning timeline by beginning environmental review and landowner negotiations earlier in the planning cycle and by identifying investment options that use existing rights-of-way wherever possible. Re-conductoring and substation upgrades on existing infrastructure, which do not require new rights-of-way or full environmental impact assessments, are, consequently, receiving increased attention. As we have covered in our analysis of re-conductoring as the fastest grid upgrade nobody is moving fast enough on, upgrading existing transmission corridors rather than building new ones is one of the few near-term tools available for addressing AI-driven transmission congestion on a timeline that is actually useful.
What Needs to Change in Utility Regulation to Enable Better Planning
Utilities are undertaking planning framework revisions in regulatory structures that were not designed to accommodate the pace and scale of change that AI data center growth demands. In most US states, integrated resource planning proceedings operate on two-to-four-year cycles, which are fundamentally incompatible with the twelve-to-eighteen-month horizon that drives AI data center development decisions.
Regulatory reform that would enable utilities to plan and invest more responsively to AI-driven load growth requires changes at multiple levels of the regulatory structure. At the state public utility commission level, it requires shortening the interval between integrated resource planning updates. It also requires expedited approval processes for capital investments directly associated with large new industrial loads, and clearer frameworks for cost allocation between data center customers and other ratepayer classes. At the federal level, it requires FERC engagement with the interconnection study reform process and potentially new guidance on how regional transmission organisations should incorporate AI load uncertainty into their transmission planning processes.
Annual Planning Updates and the Case for Shorter Cycles
First, several commissions are considering whether to require utilities to file annual load growth updates specifically addressing large industrial customer interconnection requests, rather than waiting for the two-to-four-year integrated resource planning cycle to incorporate that information. Annual updates would significantly reduce the information lag between when data center development activity is occurring and when it appears in formal utility planning documents.
The Cost Allocation Question at the Heart of Reform
The implications affect data center development economics, residential electricity bills, and the financial viability of the transmission investments required for grid reliability. State commissions are only beginning to address whether utilities should socialise those investments across all customers, allocate them specifically to the data center customers whose interconnection triggered them, or share them through some combination of both.
The Demand Response Gap
The demand response dimension of utility load planning reform deserves specific attention, because it represents the most immediate near-term tool for managing AI load variability and because the gap between the potential value of data center demand response and the current level of participation is substantial. As we have covered in our analysis of the demand response opportunity AI data centers keep refusing to take, most large AI infrastructure operators have not engaged meaningfully with demand response programmes, treating them as operational constraints rather than grid stability tools and revenue opportunities.
Utilities that are revising their load planning frameworks are, consequently, incorporating demand response from data centers as a key uncertainty variable. The difference between a planning scenario in which large AI campuses participate actively in demand response and one in which they do not is significant in terms of the reserve margins and generation capacity required to maintain reliability. Regulators in several states are beginning to examine whether large industrial customers, including data centers above certain size thresholds, should be required to participate in demand response as a condition of receiving service above a defined capacity level.
The Cost Allocation Challenge
The cost allocation dimension of utility load planning reform is politically the most contentious issue. Utilities ultimately recover transmission and generation investments triggered by AI data center load growth through rates that affect all customers in the service territory. As we have covered in our analysis of why grid decarbonisation must not compromise grid stability, the interaction between large industrial load growth and the costs borne by other ratepayers remains one of the most actively contested questions in utility regulation. The Ratepayer Protection Pledge signed at the White House in March 2026 reflects the national political importance of that issue.
The utilities that engage constructively with that policy question, working with state regulators and data center operators, are the ones that will maintain the regulatory relationships required to execute the investment programmes AI load growth demands. As we have covered in our analysis of grid decarbonisation without compromising reliability, the design of those methodologies is, ultimately, a policy question as much as an engineering one.
What Data Center Developers Should Take Away From This
The utility load planning transformation underway has direct operational implications for data center developers that go beyond the interconnection queue dynamics that are already well understood. Developers who understand how utility planning frameworks are evolving, and who engage with utilities constructively, are building the relationships that will determine their access to power in contested markets.
The most practical implication is, specifically, timing. Developers who engage with utilities on load growth plans before submitting formal interconnection applications are influencing where their projects land in the planning queue. Providing project readiness information that utilities are developing tiering frameworks to assess is, specifically, the mechanism. That influence is not guaranteed and is not a substitute for queue position. In a market where utilities are actively trying to distinguish between likely-to-proceed and speculative applications, however, a developer who makes the case for its project’s readiness before the formal process begins is in a stronger position than one who does not.
Why Demand Response Participation Is No Longer Optional
The second implication is demand response. Utilities revising their load planning frameworks are increasingly treating data center demand response participation as a planning resource that affects their reserve margin requirements and, consequently, their infrastructure investment needs. Data centers that participate in demand response programmes reduce the utility’s planning requirement and create goodwill in the regulatory relationship that translates into operational benefits over the long term. As the NERC April 2026 guidance makes clear, the expectation that large data centers will participate in grid stability mechanisms is moving from voluntary to expected, and eventually to required. The developers who establish that participation now, voluntarily and on their own terms, will have more operational flexibility when mandatory requirements arrive than those who wait to be compelled.
Ultimately, the pace of change that AI data centers are imposing on utility load planning frameworks reflects the broader challenge the AI infrastructure buildout creates for the regulatory and engineering systems that govern the US power sector. Those systems were designed for a world where the largest sources of load growth were measured in megawatts, emerged over years, and could be accommodated within planning cycles that operated on bureaucratic timescales.
The AI data center buildout has changed all of those parameters simultaneously, and the utility planning frameworks that govern the response are only beginning to catch up. The changes are not, in other words, incremental refinements to existing methodology. They are, rather, structural revisions to the foundational assumptions on which decades of utility planning practice was built. The utilities that navigate that revision most successfully, and the data center developers who work with them constructively, will consequently define the infrastructure landscape for the next decade of AI development.
