The AI Demand Dilemma: Utilities Confront Speculative Growth. 9 Min Read. Getty Images:. This is the latest installment of The Breaking Points, a series that examines the critical challenges for AI infrastructure expansion.. Utilities across the US are planning transmission lines, substations and generation to meet AI-related electricity demand that may never fully materialize. But as speculative load requests flood planning systems, a growing gap is emerging between announced demand and what utilities can realistically deliver. This disconnect is forcing utilities to rethink how they plan, finance, and build infrastructure for AI-driven growth.. In Ohio, American Electric Power (AEP) received more than 30 GW of preliminary data center load requests. After implementing stricter financial requirements under the Data Center Tariff (DCT) process, speculative requests were reduced, and only about 13 GW advanced to formal load studies. Ultimately, 5.6 GW moved forward with signed Electric Service Agreements backed by financial commitments.. Related:Speed to Power: How Developers Are Restructuring for AI Demand. The AEP reduction exposed a growing problem within utility planning departments: the gap between announced AI demand and deliverable demand is increasingly widening.. “The market is moving from an ‘announced demand’ problem to a ‘deliverable demand’ problem,” said Ihab Osman, an independent strategist specializing in data center and other mission-critical infrastructure.. Utilities are increasingly sorting AI projects into tiers of credibility:. Announced projects. Formal load-study requests. Signed service agreements. Financed developments. Active construction. Fully energized facilities. The widening gap between those categories is becoming a central planning problem for grid operators. Utilities are rewriting tariffs, demanding financial guarantees and redrawing transmission plans as AI infrastructure proposals sweep across Texas, Virginia, the Carolinas, Ohio, and other major data center corridors.. AI Demand Is Breaking Traditional Planning Modes. For decades, utilities planned around gradual growth. Population climbed while industrial demand expanded incrementally. Utilities built transmission lines, substations, and generation fleets around forecasts designed to unfold slowly over years.. The AI infrastructure boom shattered that model. Utilities, regulators and grid operators are increasingly trying to distinguish committed long-term demand from speculative load requests competing for scarce power capacity.. Betsy Soehren Jones, a partner at West Monroe focused on AI, cybersecurity, and critical infrastructure risk across the utility sector, said utilities are confronting simultaneous demand growth from AI infrastructure, manufacturing, cryptocurrency mining, and Department of Defense (DOD) modernization projects.. Related:Utilities Say Data Centers Could Lower Electricity Bills. Regulators Want Proof. “It’s no longer, ‘I’ve got a singular manufacturer coming into the state anymore,’” Jones said. “You’ve got lots of different variables, so it’s got to be more portfolio-based.”. Quarterly filings from utilities, including Duke Energy, AEP, and Exelon, show companies restructuring procurement operations and customer agreements around surging AI-related electricity demand.. Companies including OpenAI, Oracle, Crusoe, CoreWeave, and xAI are pursuing campuses requiring hundreds of megawatts – and potentially gigawatts – of electricity.. Utilities are locking up transformer supply, redesigning tariffs for hyperscale customers and fighting over who pays for transmission tied to giant AI campuses. Driving many of these decisions is a growing fear: utilities could spend billions building infrastructure for projects that never fully materialize.. Utilities Reorganize Around Supply Constraints. Duke Energy created procurement subsidiaries, known internally as “ProCos,” to secure transformers and other electrical equipment through specialized financing and leasing structures as utilities confront worsening shortages and multiyear lead times.. Related:Gridlock or Growth? ERCOT Warns Texas AI Power Boom May Not Materialize. Duke also extended a $10 billion master credit facility through 2031 while raising its five-year capital plan to $103 billion, citing rising demand from data centers and other large-load customers.. AEP this month raised its five-year capital plan to $78 billion and said it expects 63 GW of incremental contracted load by 2030, much of it tied to data centers.. Those investments reflect a broader surge for electrical infrastructure. Wood Mackenzie projected the US data center electrical equipment market could swell from roughly $20 billion to $65 billion by 2030 as AI infrastructure drives demand for transformers, switchgear and power-distribution systems.. The Southeast has emerged as one of the industry’s fastest-growing AI infrastructure corridors because vertically integrated utilities can often move generation and transmission projects faster than deregulated markets struggling with interconnection backlogs and fragmented planning systems.. Utilities and regulators are deploying grid-enhancing technologies to squeeze more capacity from existing infrastructure as AI-related electricity demand outruns transmission construction timelines.. Analysis from Grid Strategies and Americans for a Clean Energy Grid (ACEG) found the US may need roughly 5,000 miles of new high-capacity transmission annually through 2035 to maintain reliability and support rising electricity demand. Fewer than 1,000 miles were built in 2024.. Reliability Planners Sound the Alarm. Earlier this month, the North American Electric Reliability Corporation (NERC) issued a Level 3 “Essential Action” alert warning that utilities and grid operators “generally did not have sufficient processes, procedures, or methods to address risks associated with computational loads,” including AI training infrastructure and large data centers.. The alert warned that rapidly changing computational loads could trigger voltage instability, oscillations, sudden load drops and violent swings in power demand occurring within seconds.. NERC also said it is developing a new “Computational Load Entity” category covering large facilities connected to the bulk power system.. Electric Reliability Council of Texas (ERCOT) working-group discussions this year have focused on how AI-oriented data centers, battery systems and large electronic loads can create rapid power fluctuations, voltage instability and other dynamic grid challenges that traditional industrial-load planning frameworks were not designed to manage.. ERCOT stakeholders have also debated whether large AI facilities and on-site battery systems can reduce demand quickly enough to support grid stability during stressed operating conditions.. Regulators Rewrite the Rules. Utilities and regulators are rapidly rewriting the economics of large-load grid access.. At the Federal Energy Regulatory Commission (FERC), officials have opened proceedings examining how massive AI-related loads should connect to transmission systems and whether hyperscale customers should directly fund grid upgrades tied to their projects.. The debates intensified after proposals involving co-located generation projects inside the PJM Interconnection footprint raised questions about who receives priority access to constrained power infrastructure.. In Ohio, AEP proposed requiring large new data centers to commit to paying for at least 90% of the electricity capacity they request for a decade before the utility would build supporting infrastructure.. “We need accurate plans and solid commitments from large data center customers so the right facilities are built at the right time,” AEP Ohio President Marc Reitter said when the company filed the proposal.. Utilities across multiple states are now rolling out specialized large-load tariffs requiring hyperscale customers to post collateral, guarantee minimum electricity usage levels or directly fund portions of transmission expansion.. The shift reflects growing concern that utilities could otherwise build billions of dollars in infrastructure around projects that stall, shrink or disappear before full deployment.. The Fight Over ‘Paper Megawatts’. In Texas, the forecasting problem has become impossible to ignore.. ERCOT recently warned regulators that preliminary AI-related load forecasts may overstate how much proposed demand will materialize on projected timelines. The concern is not that AI demand will disappear, but that many proposed projects may never advance as quickly – or as far – as their initial plans suggest.. ERCOT said it may revise projections using “historical realization rates” and other independent metrics as planners struggle to separate speculative requests from buildable projects.. The problem resembles earlier renewable interconnection queues, where speculative projects often flooded planning systems faster than developers could secure financing, equipment or permits.. ERCOT is also developing a new “batch study” framework to group large-load projects that meet stricter requirements into coordinated transmission studies.. ERCOT disclosed that average peak consumption for some proposed large-load projects reached only 49.8% of requested megawatt levels in recent planning exercises.. Jones said utilities are becoming more skeptical of large-load announcements because projects increasingly compete for constrained equipment and supply-chain capacity.. “Do they have the permits? Is the project funding secured?” Jones said. “More importantly, do they have the supplies to actually build it?”. Tyler Demetriou, an associate attorney at the Southern Environmental Law Center, said utilities and hyperscale developers can both benefit from aggressive long-term load projections.. Utilities can justify new transmission and generation investments that earn regulated returns, while developers may shop projects across multiple jurisdictions before committing to a final site, he said.. “Without regulatory intervention, neither of those parties is incentivized to take a realistic look at things,” Demetriou said.. Osman said utilities are increasingly focused on execution risk rather than headline megawatt figures.. “A serious AI infrastructure project is not defined by the number of megawatts claimed,” Osman said. “It is defined by the chain behind it: site control, interconnection path, utility engagement, collateral and milestones, transformer and switchgear access, transmission feasibility, financing, permitting, cooling strategy, EPC credibility, and a commissioning path that can survive contact with reality.”. Utilities built grids around relatively predictable growth. AI infrastructure introduces a different problem: gigawatt-scale demand tied to volatile capital cycles, uncertain deployment timelines and intense competition for scarce power capacity.. “The growth is going to come, period,” Jones said. “I just think it’s the certainty of when it’s going to hit.”. The question is, who absorbs the cost if projected demand doesn’t materialize on schedule? Collateral requirements, minimum-demand agreements, phased energization schedules and transmission security arrangements are designed to prevent ordinary ratepayers from absorbing the cost of overbuilt infrastructure if large AI campuses stall, shrink or disappear.. Who Pays If the Load Never Arrives?. Jones said federal industrial policy and AI expansion goals are colliding with utility cost-recovery systems built for slower, more predictable growth.. “The financial mechanisms to enable all of that are stuck in decades-old rate-making policy,” Jones said.. In a recent filing before the Public Utility Commission of Texas (PUCT), a coalition of retail electricity providers argued that a large commercial customer reducing 1 MW of peak demand can save roughly $68,550 annually under current ERCOT transmission pricing structures, while a retail provider managing the same reduction across residential customers saves only about $94 annually.. Consumer advocates and utility commissions are increasingly questioning whether residential customers could absorb the cost of transmission lines, substations and generation projects built primarily for hyperscale customers.. That concern sits at the center of a recent complaint filed by the Maryland Office of People’s Counsel (OPC) against transmission planning decisions within PJM. The filing argues projected data center growth is already driving billions of dollars in transmission expansion and warns consumers could face substantial costs if expected load growth fails to materialize.. In several PJM states, recent capacity-market increases tied to rising data center demand have already pushed residential electricity bills higher.. Jones said policymakers often frame the problem as an AI data center issue when utilities are simultaneously managing load growth tied to manufacturing reshoring, military modernization and other industrial expansion.. “The biggest misconception is that it’s just data centers,” Jones said.. Utilities built the modern grid around demand they could forecast. The AI buildout is forcing them to plan around demand that may never fully exist.. About the Author
