PJM’s 15GW Emergency Capacity Request Exposes AI Infrastructure Bottleneck
The nation's largest grid operator seeks emergency power supplies as data center demand outpaces infrastructure buildout by 5-7 years, threatening US compute competitiveness.
PJM Interconnection requested 15 gigawatts of emergency power capacity on April 10, a move that transforms the AI data center boom from an energy planning challenge into a systemic infrastructure crisis across 13 states serving 65 million people.
The emergency request from the nation’s largest grid operator crystallizes a fundamental mismatch: Data Centers deploy in 2-3 years while transmission upgrades require 7-10 years. The December 2025 capacity auction fell 5.2% short of reliability requirements for the 2027-2028 delivery year—the first shortfall in PJM’s history, per Jackson Walker. Starting June 2026, ratepayers across the region will absorb an additional $1.4 billion in capacity market costs driven primarily by data center load growth.
The Infrastructure Constraint
Capacity prices surged more than tenfold in two years, from $28.92 per megawatt-day in the 2024/25 delivery year to $329.17 in 2026/27, according to IEEFA. Data centers accounted for $6.5 billion—40% of the $16.4 billion total—in the December 2025 auction. PJM now forecasts data centers will require more than 30 gigawatts of peak capacity by 2030, while summer peak load is expected to grow 3.6% annually to 222 GW by 2036—a 10-year increase of 65.7 GW.
The physical bottleneck is worsening. Grid connection requests now take 4-7 years in key regions like Virginia, with approximately 11 GW of announced 2026 data center capacity still lacking construction starts, Tech-Insider reported. Two terawatts of capacity—almost twice the currently installed base across the entire US grid—sits stuck in interconnection queues.
“Data centers are connecting to the grid faster than we can build new power supply.”
— NRDC analysis
Debbie Lew of the Energy Systems Integration Group framed the timeline gap plainly: “You can build a data center in two to three years, but it might take at least seven years to build new or upgraded transmission facilities.” The constraint is now physical Infrastructure rather than capital availability—a shift that fundamentally alters investment dynamics across the energy sector.
Regulatory Response and Market Impacts
The White House National Energy Dominance Council and all 13 PJM state governors jointly called for an emergency auction by September 2026 in a January 16 statement, as documented by Jackson Walker. PJM’s expedited interconnection track targets August 2026 implementation, with FERC approval sought by April 28. PJM Board Chair David Mills emphasised the proposal represents grid optimization rather than data center prohibition: “This is not a yes/no to data centers.”
The cost burden extends beyond wholesale markets. NRDC analysis projects electricity costs could increase by over $100 billion through 2033 due to data center load growth, potentially adding $70 monthly to household bills. The economic impact is concentrated in states with high data center density—Virginia, Ohio, Pennsylvania—where capacity charges represent a growing share of retail rates.
Replacing thermal generation with renewables creates a capacity value gap. Each megawatt of retired coal or gas plant requires approximately 5.2 MW of solar or 14 MW of onshore wind to provide equivalent reliability. From 2011-2023, PJM retired 54.2 GW of thermal capacity; an additional 24-58 GW faces retirement risk by 2030. Meanwhile, 86 GW of new utility-scale capacity is planned for 2026—51% solar, 28% battery storage, 14% wind—but intermittency means firm capacity additions lag nameplate figures.
AI Workload Intensity Drives Demand
US data centers are projected to consume 580 TWh annually by 2028, accounting for 12% of the nation’s total electricity use, per ERM citing Lawrence Berkeley National Lab data. Global data center power demand could reach 68 GW by 2027 and 327 GW by 2030, compared to 88 GW of total global capacity in 2022, according to RAND Corporation research.
Single hyperscale AI data centers now require 100-300 MW of continuous power. Individual training runs could demand 1 GW by 2028 and 8 GW by 2030—equivalent to a large nuclear plant dedicated to a single computational task. This represents a step-change from traditional enterprise data centers, which typically operate at 5-20 MW.
Developer behaviour is already shifting. Wood Mackenzie data shows data center developers added only 25 GW to project pipelines in Q4 2025—half the previous quarter’s volume—signaling that grid constraints are now binding on expansion plans, Fortune reported. Ben Hertz-Shargel of Wood Mackenzie noted utilities “just don’t necessarily have either the grid capacity or the generating capacity to be able to build it fast enough to accommodate these new large energy demand centers.”
Geopolitical Compute Competition
The infrastructure bottleneck carries strategic implications beyond energy markets. China and Gulf states are building AI infrastructure with integrated power generation—avoiding the interconnection queue constraints that now bind US deployments. RAND analysis highlights this divergence: jurisdictions with state-directed energy development can co-locate generation and compute infrastructure on accelerated timelines, while US projects navigate multi-year regulatory processes across federal, state, and regional bodies.
The constraint is particularly acute for AI training workloads, which require sustained high power density and cannot easily shift across geographic regions. Cloud inference workloads can distribute globally, but the frontier model training that determines technological leadership requires concentrated power delivery—exactly what PJM’s congested grid struggles to provide.
- Firm capacity generation—gas peakers, nuclear SMRs, geothermal—commands premium pricing as intermittent renewables prove insufficient for reliability requirements
- Grid modernization and transmission upgrades: 2 TW in interconnection queues requires hundreds of billions in network investment
- Energy storage at utility scale: battery deployments comprise 28% of planned 2026 capacity additions, bridging renewable intermittency
- On-site generation for data centers: “bring your own grid” approaches bypass interconnection delays, creating demand for modular power solutions
What to Watch
FERC’s April 28 decision on PJM’s expedited interconnection framework will determine whether projects entering the queue in 2026 reach operation by 2033 or face longer delays. The September emergency auction—if approved—will test whether capacity pricing can clear sufficient supply or if physical constraints render price signals ineffective. Summer 2026 grid performance across PJM will provide the first real-world test of whether the current capacity portfolio can maintain reliability under accelerating data center load growth.
Longer-term, the mismatch between AI deployment cycles and grid expansion timelines is forcing a fundamental reassessment of infrastructure planning. PJM’s crisis is regional in geography but systemic in implication—every major grid operator faces the same collision between legacy planning processes and exponential compute demand growth. Whether the US adapts faster than competitors with more centralized energy governance will shape the geography of AI development for the next decade.