Artificial intelligence is pushing the U.S. power grid towards its design limits. According to Goldman Sachs Research, global data center electricity demand is projected to reach 84 gigawatts (GW) by 2027, a 50% surge from 2023, with AI workloads accounting for 27% of this total. Domestically, U.S. utility forecasts for summer peak demand over the next five years have jumped from 38 GW to 128 GW, more than doubling within a single planning cycle. This figure indicates that the U.S. grid is already near its capacity limits even before the full force of the AI wave arrives.
Facing this shortfall, natural gas has emerged as the only currently viable option that can meet demand requirements in both speed and scale. Entergy is investing $3.2 billion to build three natural gas plants totaling 2.3 GW, specifically to power Meta's new data center in Louisiana. NextEra Energy, the largest U.S. renewable energy developer, is collaborating with ExxonMobil to construct a new 1.2 GW natural gas plant in the Southeastern U.S. NextEra CEO John Ketchum described this trend as the AI industry shifting towards "BYOG"—bringing your own generation.
However, the cost of large-scale bets on natural gas is becoming apparent. The capacity market clearing price for the PJM Interconnection grid, which serves a large portion of the Mid-Atlantic and Midwest, for the 2026-2027 delivery year has surged to $329 per megawatt-day, a more than tenfold increase from $28.92/MW-day two years ago. The capital cost for new combined-cycle gas turbines has also nearly doubled to approximately $2,000 per kilowatt. Meanwhile, these power plants, designed with operational lifespans of around 30 years, will profoundly impact U.S. carbon reduction commitments.
Natural gas currently supplies about 40% of U.S. electricity, with renewables and coal constituting the remainder. AI data centers require stable, 24/7, gigawatt-scale power, which is precisely the existing grid's weakness. Renewables face hard constraints: the median wait time for grid interconnection for new solar and wind projects exceeds four years. In contrast, natural gas is relatively inexpensive, abundantly available, and supported by a nationwide pipeline network, allowing new plants to become operational within three to five years.
The three plants Entergy is building for Meta's Louisiana data center alone will require 2 GW to power the computing load. The NextEra-ExxonMobil partnership bets on a "build the infrastructure first, then attract tenants" logic. John Ketchum's "BYOG" comment effectively signals a shift in the relationship between energy companies and tech giants from "supplier and buyer" to "joint infrastructure developers."
Traditional grid engineering is based on the premise of predictable loads—seasonal peaks, industrial cycles, and population growth can be modeled and planned for years in advance. AI fundamentally disrupts this logic. Training large language models involves thousands of GPUs running at high load for days or even weeks, followed by abrupt shutdowns. Each AI inference response can trigger instantaneous power spikes measuring hundreds of megawatts. Such load behaviors are unprecedented and directly challenge the dispatch curves and reserve capacity models utilities rely on.
To address this, peaker plants—designed for short-duration, high-output needs—are being deployed directly adjacent to data center campuses to absorb inference spikes that baseload plants cannot respond to quickly enough. Simultaneously, the Texas grid operator ERCOT has developed a new "Adjusted Large Load Forecast" methodology to distinguish between actual, imminent demand and "phantom demand" created by speculative interconnection applications.
The physical grid is also under strain. Transmission investment in many parts of the U.S. declined after 2015, leaving the system operating near its limits even before the AI demand surge. A U.S. Department of Energy (DOE) national transmission study concluded that transmission congestion was already severe in multiple regions prior to this recent demand increase. Demand-side data is equally striking: Texas's CenterPoint Energy reported a 700% increase in large-load interconnection requests between late 2023 and late 2024. Virginia still has a queue of up to 50 GW of data center projects awaiting grid connection.
Market prices reflect this structural tightness. The PJM capacity market clearing price soared from $28.92/MW-day to $329/MW-day within two years. The capital cost for new combined-cycle gas turbines has nearly doubled from levels seen a few years ago to around $2,000/kW. These costs will ultimately be passed downstream through electricity rate mechanisms.
The natural gas plants currently under review and construction are not merely transitional infrastructure. With average operational lifespans of about 30 years, they are set to operate well beyond the deadlines of most major net-zero emissions targets. The full lifecycle carbon emissions of natural gas are approximately 490 grams of CO₂ per kWh. In the Southern U.S., utilities plan to add about 20 GW of new natural gas capacity over the next 15 years, with data centers accounting for 65% to 85% of the forecast load growth in Virginia, South Carolina, and Georgia.
The methane issue further exacerbates carbon emission concerns. Natural gas infrastructure consistently leaks methane, which has about 80 times the warming potency of CO₂ over a 20-year period. This problem is creating a policy fault line among energy companies, large tech firms committed to net-zero, and regulators still grappling with the implications of AI's energy consumption.
Several structural policy tools are advancing, but none can address the immediate supply-demand gap. For energy storage, the 2022 Inflation Reduction Act provides a 30% tax credit for standalone storage systems entering service after 2024, offering a financial incentive for investment in battery systems. Nuclear power, a zero-carbon option capable of providing stable baseload power, is gaining increased attention. Google has partnered with NextEra in a deal to restart the 615 MW Duane Arnold nuclear facility to access 24/7 carbon-free power.
Transmission expansion remains the toughest challenge. The DOE study indicates significant transmission capacity shortfalls exist in nearly every U.S. region. Bridging these gaps will require years of coordinated investment and permitting reforms. Currently, the growth rate of AI-driven demand is comprehensively outpacing the development of its supporting infrastructure—natural gas plants, transmission upgrades, energy storage deployment, and nuclear restarts are all lagging. How this gap is closed will depend on whether policy coordination and proactive investment can arrive before more painful consequences like power shortages, data center construction delays, and rising electricity prices take hold.
Comments