Large Load and Data Center Demand

Inside each facility, electrical and mechanical systems dominate the build cost. Applied to the current $770–910B data-center buildout, published cost splits (Turner & Townsend 54% electrical, 22% mechanical for air-cooled; Dgtl Infra 40–45% electrical and 15–20% HVAC) imply roughly $300–500B of facility-side electrical spend alone, and roughly $425–700B+ when electrical and mechanical systems are combined. That is not the same accounting boundary as utility capex, but it is the same supply chain: transformers, switchgear, cooling equipment, and skilled electrical trades. Planned 2025 U.S. investor-owned utility capex is about $208B. Data-center demand is arriving into the same equipment and labor bottlenecks the grid already faces.

The reliability consequence is already visible beyond the U.S. forecast data: NERC says more than half of North American assessment areas (13 of 23) face elevated or high resource-adequacy risk over the next decade as load growth outruns planned generation and transmission buildout.

Large load is the easiest piece to itemize, but it is not the whole story. Anchor on each market's total net peak growth over 2026-2031 and split it into drivers, and a pattern appears: in ERCOT, PJM, and MISO large load (mostly data centers) is the dominant driver, about 93.8% of their growth. Total net growth across the seven ISOs is about 110 GW, summed across different ISO peak bases (not a single coincident national peak). In CAISO and ISO-NE the growth is led by electrification, with large load a minor slice. Almost everywhere, energy efficiency and behind-the-meter solar pull the net back down.

Each bar is one ISO's total net summer-peak growth (2026-2031), split into drivers that reconcile to the total: growth drivers (large load, electrification, organic baseline) extend right of zero; peak reducers (energy efficiency, behind-the-meter solar and storage) extend left; the tick marks net growth. Within large load, segments are split by load type where the ISO publishes one (mainly ERCOT: crypto, industrial, oil and gas, hydrogen); elsewhere large load is data-center-dominated, and type shares apply each ISO's published 2031 mix to the 2026-2031 increment. A large-load stack reaching past the tick means large load exceeds the entire net rise: in PJM large load (+35 GW, almost all data centers) tops the whole net peak increase (+35 GW), so without it PJM's peak would be roughly flat (data centers alone are ~95% of net). Definitions differ by ISO and the bars are not additive into a single national peak: PJM is unrestricted (pre-DR), CAISO is managed-net, ISO-NE is summer 50/50 net (and winter-peaking, so its 0.7 GW understates electrification). EE is itemized only where the ISO forecast separates it; elsewhere it is embedded in baseline load or, for ISO-NE, outside the selected peak basis in the capacity-market passive-demand stack. Zero does not mean no EE. MISO's split is a modeled front-loaded estimate (range ~7-20 GW, central 2.0% CAGR). SPP's ~12 GW of new spot loads do not land at the system peak hour, so they are excluded and its coincident large load is ~0. Hover for each driver, peak basis, and source.

Data center location depends on workload type. Inference facilities (serving real-time user requests) cluster near fiber connectivity hubs and population centers for low latency. Training campuses (building AI models) can locate more remotely, prioritizing cheap power and land over proximity to users. Crypto mining follows the lowest-cost electricity regardless of location. Northern Virginia, Dallas-Fort Worth, and the Portland-Hillsboro corridor are the densest markets. Where these clusters overlap with congested transmission or retiring generation, grid pressure is highest.

Internet exchanges (blue) are facilities where multiple networks interconnect, providing the fiber backbone connectivity that data centers require. Their locations indicate where high-bandwidth fiber infrastructure already exists, a key factor in data center site selection alongside power availability.

Interconnection queues stretch 4-6 years, driven by a cascade of bottlenecks. The largest single delay is the study queue itself: ISOs cannot process the volume of requests. As of the end of 2025, PJM alone had about 1,100 active generator-interconnection requests totaling ~144 GW (LBNL, Queued Up 2026). When projects ahead in the queue withdraw (roughly 80% do), studies for remaining projects must be redone. Behind the study backlog, transmission infrastructure takes 5-10 years to build and permit. And equipment procurement, particularly power transformers, adds further delay.

Three categories of demand compete for the same manufacturing capacity: replacing aging infrastructure (~30-40% of demand, with over half of US distribution transformers beyond expected service life), connecting renewable generation (~30-35%, with ~2 TW in interconnection queues), and serving new large loads like data centers (~20-25%, the fastest-growing segment). The US has one domestic producer of the specialized electrical steel (GOES) used in transformer cores, and imports the majority of its power transformers.

Developers facing 4-6 year interconnection queues are responding in two ways. The choice between them is reshaping how the grid evolves.

Why this matters for grid planning: Behind-the-meter generation is largely invisible to ISO operators. A data center with 500 MW of on-site gas may appear as zero load to the grid (fully self-served) or as 500 MW of sudden demand if the on-site generation trips. This creates phantom reliability risk that existing planning models do not account for.

Why this matters for clean energy: Most hyperscalers claim 100% renewable electricity via annual renewable energy credit matching. Behind-the-meter gas generation sits outside the utility's obligation. Virginia's Clean Economy Act mandates 100% carbon-free electricity from Dominion by 2045, but BTM gas at data centers in Virginia effectively circumvents this mandate.

Why it's happening: Speed to power. GPU-dense AI facilities generate an estimated $10-12 million per MW annually in compute revenue (general-purpose colocation is significantly lower). A 4-6 year interconnection queue represents tens of billions in lost revenue. Temporary or modular on-site gas (turbine-on-trailer units, pipeline-connected modular generation) can come online in months; permanent large-scale gas plants take longer but still typically beat utility interconnection timelines. 90% of these projects were announced in 2025 alone.

Five forces are converging to make the brute-force path harder and the flexible path more attractive:

None of these forces has stopped Path 1: 90% of the 56 GW in BTM announcements came in 2025, after all five were active. But they are creating conditions where the flexible path becomes viable for developers willing to trade some speed for lower political and regulatory risk.

Does data center flexibility actually work?

Data centers are technically capable of meaningful flexibility, but economically incentivized against it. Compute economics are roughly two orders of magnitude larger than flexibility value on a per-MW-year basis, so flexibility only activates at scale when it is a condition of interconnection, not a voluntary economic choice.

How flexibility trades for speed

Three models are emerging where developers trade flexibility commitments for faster grid access: deploying on-site storage that benefits both the data center and the grid, embedding demand response into power purchase agreements, or bringing their own accredited capacity to offset new utility infrastructure. This last approach is known as Bring-Your-Own-Capacity (BYOC): the data center developer directly procures generation capacity (clean energy contracts, batteries, or on-site resources) and makes it available to the wholesale market, rather than relying on the utility to build new infrastructure.

The Camus/Princeton ZERO Lab analysis found that a 500 MW data center using flexible grid connection plus bring-your-own-capacity can reach full operation in approximately 2 years, 3-5 years faster than traditional interconnection. The combined model offsets nearly 100% of the traditional firm interconnection cost ($764M per GW).

Making other loads more flexible can also accelerate data center interconnection. Residential air conditioning (load factor 36%) creates the sharpest peak demand. Reducing residential peak through smart thermostats, managed EV charging, or demand response programs frees grid capacity that new loads can use. Active managed EV charging alone saves up to $400 per EV annually in avoided distribution upgrades (Brattle Group, January 2026).

For developers, the core question is not whether flexibility is conceptually valuable. It is which path gets power online fastest, through which counterparty, and with what tradeoff.

Cost-reflective tariffs (the AEP Ohio model) are not a direct path to power, but they change the economics developers face. In AEP Ohio's queue: ~30 GW reported earlier, and 36 sites totaling about 13 GW signed up for formal studies.

Not every market is at the same stage. Some have operational grid-cooperative mechanisms. Others are being forced toward them by political and regulatory pressure. The distinction matters: proven traction is a different signal than policy-forced transition.

Where Path 2 has real traction

Where the market is being forced toward Path 2

Status reflects current evidence as of early 2026. "Active pilot" means an operational or approved mechanism with committed capital. "Proven policy lever" means demonstrated demand-side impact from regulatory action. "Pressure zone" means high political/regulatory pressure with mechanism design underway but not yet proven.

Utilities size transmission and distribution infrastructure for peak demand. Different customer classes contribute to that peak differently. Residential load is the peakiest (load factor ~36% in PJM, driven by air conditioning and heating). Commercial is moderately peaky (~46%). Data centers have one of the highest load factors of any customer class (~80%, nearly flat across all hours), comparable to continuous-process industries.

This creates a cost allocation tension. When high load factor customers like data centers grow, they increase total energy consumption without proportionally increasing peak demand. As baseload grows, the fixed costs of peak infrastructure spread across more kilowatt-hours, but the system is still sized for peak. Customers with lower load factors, particularly residential, can see rising costs even if their own consumption patterns have not changed, because the system is being expanded to serve growing baseload demand.

The mechanism is concrete. PJM bills capacity and transmission costs to each utility based on its load during a small handful of system-peak hours: five summer-afternoon peak hours for capacity (the customer-level number is called a Peak Load Contribution, or PLC), and the single highest hour of the year for transmission (a Network Service Peak Load, or NSPL). Whatever a customer pulls during those hours becomes their share of the bill for the next year. Other ISOs use related mechanisms: ERCOT allocates on 4CP (the four highest summer peaks), MISO and ISO-NE on a single annual peak, and CAISO on monthly peaks. The specifics differ, but in each case a few hours determine how infrastructure costs get allocated across customer classes.

For most residential customers in PJM, the allocator-hour load is not measured directly. Advanced metering infrastructure (AMI) reached about 84% of US meters by 2024, but parts of the Midwest and East Coast still lag, so the typical PJM utility computes a proxy PLC from the residential class's average summer load profile and scales it by each customer's annual kilowatt-hours. The 1.46x summer-peak ratio above gets baked into every residential customer's per-kWh capacity allocation regardless of their own behavior at peak. A family on vacation during the 5 PJM peak hours and a household running smart thermostats or scheduled EV charging both pay the same class-based PLC, even if their actual peak demand was below the class average. Data centers, by contrast, are interval-metered and get an individual PLC from their actual demand during those 5 hours: for a flat 1.0x load, that comes out roughly equal to the annual average.

New transmission and generation built specifically because of data-center load growth is recovered through this same allocation method, which spreads costs across all classes in proportion to their existing PLC share. The Union of Concerned Scientists counted $4.4 billion in 2024 PJM transmission upgrades attributable to data-center interconnections, socialized across customers in seven states (IL, MD, NJ, OH, PA, VA, WV). A Dominion witness testified at the Virginia SCC that of the $7.6 billion in planned new transmission infrastructure identified in Dominion's 2024 Integrated Resource Plan, residential customers will pay 55%, including for projects that serve only data centers. PJM's Independent Market Monitor attributed 40% of the December 2024 capacity auction's $16.4 billion in clearing costs (around $6.5 billion) to data-center load. Those costs flow to every customer in PJM via their utility's capacity rider, regardless of how flat or peaky the customer's own load is.

This is why every state with significant data-center load growth is now creating separate rate classes. Virginia's State Corporation Commission approved a new GS-5 rate class (general service for customers above 25 MW) in November 2025, requiring 85% minimum demand for transmission and distribution and 60% minimum for generation under 14-year contracts. Ohio's PUCO approved a parallel AEP Ohio tariff in July 2025: 85% minimum billing for up to 12 years for customers above 25 MW, with a three-year exit notice or a fee equal to three years of energy use. Oregon enacted the POWER Act in June 2025, requiring investor-owned utilities to create a separate large-load customer category. Pennsylvania's PUC issued a tentative model tariff for large-load customers in November 2025. PUCO Chair Jenifer French said the AEP Ohio order “safeguards non-data center customers on an industrial and residential level.” Virginia's SCC said the GS-5 class will “help insulate ratepayers from the costs around the rapid build-out and construction of infrastructure to support businesses such as data centers.” Both orders explicitly acknowledged the cost-shift risk the standard methodology was producing. The 78% of Americans concerned about data centers raising their energy bills are responding to a real and acknowledged risk, not a hypothetical.

Three policy approaches are being tested across states. Each has trade-offs, and none has been in place long enough to draw definitive conclusions about long-term outcomes.

Data centers are the grid's largest new load, its most politically contested infrastructure category, and technically capable of meaningful flexibility. The evidence shows they are not yet choosing flexibility voluntarily. What moves the needle is compulsion through rate design, interconnection conditions, and regulatory frameworks, not market incentives alone.

The central question is not whether data center flexibility is technically possible. EPRI proved it is. The question is whether the regulatory and market frameworks that compel flexibility can scale faster than the political backlash that forecloses the opportunity entirely.

If you reference this page in research, regulatory filings, or journalism, the citations below resolve to a stable URL.

Balgeman, C. (2026). Data Centers and the US Grid: Where Load Is Growing and Why It's Constrained. Grid Flexibility. https://gridflexibility.fyi/data-centers

@misc{balgeman2026datacenters,
  author       = {Balgeman, Corey},
  title        = {Data Centers and the {US} Grid: Where Load Is Growing and Why It's Constrained},
  year         = {2026},
  howpublished = {Grid Flexibility},
  url          = {https://gridflexibility.fyi/data-centers},
  note         = {Last updated 2026-05-02}
}