U.S. electricity demand forecasts that project a surge from AI data centers may be significantly inflated by phantom interconnection requests and optimistic tech industry extrapolations, driving a build-out of gas-fired power plants that could cost ratepayers billions for capacity that never materializes. Regional grid operators and consulting firms have published eye-popping numbers—PJM's summer peak climbing by 70,000 megawatts over 15 years, ERCOT's near doubling by 2030, and a nationwide 128-gigawatt jump by 2029—yet a growing chorus of energy researchers warns that the underlying assumptions are riddled with speculation. The disconnect between headline forecasts and verifiable commitments is already reshaping utility investment plans and consumer electricity rates, sparking an urgent debate over who should bear the financial risk of a potential overbuild.
The Numbers Fueling the Hype
Every major U.S. grid operator now points to AI-ready hyperscale data centers as the primary catalyst for a historic reversal in electricity demand trends. PJM Interconnection, serving 65 million people from Washington, D.C. to Chicago, expects its summer peak to jump roughly 42%—adding 70,000 MW over the next 15 years—an increase equivalent to the output of dozens of large power plants. ERCOT, the Texas grid, has publicly warned that peak demand could nearly double from 2024 levels by 2030, pushing into the mid-100s of gigawatts.
National-scale analyses amplify the message. Grid Strategies, a consulting firm, found that five-year peak-load forecasts across utility filings grew by 128 GW through 2029, concentrated in data-center-heavy regions like northern Virginia, Georgia, and the Pacific Northwest. ICF's 2025 scenario models a 25% higher national electricity demand by 2030 and up to 78% by 2050, with AI workloads as a primary accelerator. These projections have become the evidentiary backbone for utility proposals to build new gas-fired generation and transmission, and for tech companies to negotiate preferential energy deals. But critics argue that the forecasts treat raw interconnection queue volumes as committed load—a methodological flaw that overstates reality.
The Phantom Data Center Problem
"One data center may shop around a few different locations before it decides where it finally wants to be," Cathy Kunkel, a consultant at the Institute for Energy Economics and Financial Analysis, told Straight Arrow News. Utility interconnection queues are cluttered with duplicate and speculative requests from companies that file multiple applications in parallel while negotiating tax breaks and site options. Neither utilities nor grid operators routinely reconcile these duplicates across service territories, so the raw "requests" megawatt count balloons well beyond what will ever be built.
The Wall Street Journal recently reported that some utility projections for future power demand are multiples higher than existing peak demand—a sign of how phantom data centers distort planning. Jonathan Koomey, a former Lawrence Berkeley National Laboratory scientist who documented the dot-com era's overhyped electricity forecasts, sees a direct parallel. "It turned out that across the board, these claims were vast exaggerations," he said, referring to 1990s predictions that computers would consume half of all U.S. electricity. "Both the utilities and the tech companies have an incentive to embrace the rapid growth forecast for electricity use."
Sean O'Leary of the Ohio River Valley Institute put it bluntly: "Many data centers that are talked about and proposed and in some cases even announced will never get built." Yet utilities are already acting on those shaky signals.
Why the Tech Industry's Own Economics Throw Cold Water on the Forecasts
Beyond speculative siting, the financial viability of AI pure-plays remains questionable—a point often lost in the hype. OpenAI reported a $5 billion loss in 2024, while Anthropic, the maker of Claude, was found to be losing money on its paid subscribers according to one analysis. Major cloud providers like Microsoft, Google, and Amazon have deep pockets and strategic reasons to build out enterprise AI capacity, but the surge of investment into unprofitable startups raises doubts about how many of the biggest multi-gigawatt proposals will actually break ground.
"The forecasts that we're seeing right now are basically what the tech industry wants to happen and what they're selling to their investors," Kunkel said. And even if demand materializes, physical constraints may cap growth. AI accelerators—primarily Nvidia GPUs—have at times been supply-limited. Alex de Vries, founder of digiconomist and a researcher at VU Amsterdam, noted that tech companies are already utilizing all the advanced chips manufacturers can produce. "You can't really predict much further than like one to two years into the future" because "the supply chain capacity is fully utilized," he told SAN. A London Economics International analysis supports this view, concluding that chip bottlenecks mean many proposed data centers will never obtain the necessary hardware.
Efficiency: A Wildcard That Could Slash Demand
On the consumption side, breakthroughs in AI model efficiency could dramatically reshape demand. The release of DeepSeek, a Chinese AI model claiming similar performance to ChatGPT while using up to 90% less electricity, sent shockwaves through the industry. If such gains prove replicable and broadly adopted, the energy intensity of AI workloads could be orders of magnitude lower than current extrapolations assume.
However, efficiency alone won't cap grid stress. Cheaper per-query compute costs typically unleash more consumption—the Jevons paradox—so total energy draw could still climb. Alex de Vries flagged that newer "reasoning" models capable of advanced analysis are more power-hungry than conventional large language models, and AI-generated video far outstrips text interactions. The net effect of efficiency improvements remains uncertain, but it exposes the fragility of 10- and 15-year load forecasts that assume fixed per-workload energy intensity.
Utilities and the Guaranteed Payday
Despite the red flags, U.S. utilities are moving aggressively to lock in new generation—overwhelmingly gas-fired. Reuters reports that about 114,000 MW of new gas power plants are currently proposed or in planning, often pitched as "firm" capacity to serve AI loads sited adjacent to data centers. For example, Entergy Louisiana won state approval to construct three new gas plants dedicated to a Meta data center, adding enough electricity for two cities the size of New Orleans.
Herein lies the consumer-protection threat: utilities typically recover capital costs plus a guaranteed rate of return through customer bills, regardless of whether the assets are ultimately needed. Critics contend this creates a powerful incentive to overbuild. "The utility is going to be able to recover the cost of that power plant plus a commercial rate of return, whether or not that plant is ultimately needed or not," O'Leary said. Tyson Slocum, energy program director at Public Citizen, added that utilities and tech firms "intentionally create a sense of hype and panic around the energy consumption of AI, because both had a shared financial interest in doing so."
Meanwhile, hyperscalers are pursuing direct deals—long-term power purchase agreements, investments in nuclear refurbishments, and behind-the-meter generation—that could mitigate grid strain but also concentrate bargaining power. Google's nuclear fusion investment and Microsoft's push to revive Three Mile Island are long-term speculative bets; in the interim, gas is the go-to.
Technical Handcuffs: Transmission, Cooling, and Permitting
Even if demand forecasts are accurate, the grid's physical capacity to deliver power on the timelines envisioned is doubtful. High-voltage transmission lines, substations, and interconnects take years to site, permit, and build. PJM's own interconnection queue backlog has become notorious, delaying both renewable and fossil projects. Data centers demand continuous, high-capacity power, requiring dispatchable sources or contracts that guarantee availability around the clock—short-duration batteries paired with solar won't suffice without substantial long-duration storage or flexible gas/nuclear capacity.
Cooling and water add another layer of complexity. Large AI facilities need massive amounts of water for evaporative cooling in many designs, stressing local aquifers and triggering environmental reviews. Hydrogen-ready gas turbines and small modular reactors are frequently touted as solutions, but both face multi-year permitting and commercialization hurdles.
The Ratepayer Reckoning Already Underway
Consumers in high-growth data center corridors are already feeling the pinch. Capacity auction price spikes in PJM and localized rate increases signal that the mismatch between forecasted demand and available supply is trickling down to bills. If utilities overbuild based on inflated projections, those costs will be spread across captive ratepayers for decades—a scenario Jonathan Koomey warns is a repeat of the dot-com bubble's infrastructure overhang.
For IT decision-makers and organizations reliant on cloud services, the fallout could reshape where workloads are hosted. Hyperscalers may increasingly concentrate new capacity in regions with firm, long-term energy deals, leading to cost arbitrage and latency shifts. Localized outages or constrained capacity could become a material factor in site selection for Windows-based and enterprise workloads. Meanwhile, communities weighing data center proposals face a delicate calculus: tax revenue and jobs versus potential water scarcity and higher electricity rates if speculative requests drive overdevelopment.
Policy Fixes to De-Risk the Forecasts
A suite of policy interventions could prevent a costly overbuild while still ensuring reliability if AI demand materializes. First, interconnection queue reform—requiring meaningful deposits, enforceable milestones, and cross-utility reconciliation of duplicate requests—would bring queue numbers closer to committed reality. The Federal Energy Regulatory Commission (FERC) has begun exploring such measures. Second, risk-weighted planning that treats queue megawatts probabilistically, rather than as deterministic load, could result in staged generation procurement better matched to actual need.
Demand-response agreements with hyperscalers offer another lever. Major cloud providers have piloted automated curtailment; formalizing these into contract language could shave peak requirements. Ironically, AI-assisted grid studies are now being tested by PJM and others to speed interconnection analysis—a case of AI helping solve a problem it helped create. Lastly, robust ratepayer protections must require utilities to prove prudence before passing through costs of speculative assets, with developers bearing the risk for projects that never connect.
What Comes Next
The AI energy debate remains a high-stakes balancing act. On one hand, ignoring the upside risk—the possibility that AI and electrification truly generate the high-load scenarios—could lead to brownouts, economic disruption, and a slower energy transition. On the other, allowing utility rate bases to swell on the back of duplicate interconnection requests and unverified extrapolations risks transferring billions in costs to households and small businesses.
The pragmatic path is disciplined planning: verify developer commitments through binding power purchase agreements and construction milestones, not just queue entries. For every eye-popping headline forecast, ask whether it accounts for duplicate applications, chip supply constraints, and the Jevons effect. As Koomey noted, after the dot-com bubble, electricity demand growth turned out to be far tamer than predicted. The marker that separates a credible forecast from hype isn't the size of the number—it's whether the megawatts are backed by deposits and contracts that put real money on the line. For now, consumers and policymakers should watch utility filings for those binding signals, and remain skeptical of simple extrapolations from a queue that is clogged with phantoms.