Grid operators and states moved to make AI data centers pay for their connections

Requests to connect AI data centers ran far ahead of the load actually energized in the first quarter of 2026. Utilities raised capital plans on that demand, PJM customers absorbed sharply higher capacity costs, and the terms for who pays were being rewritten in Texas, across the PJM region and in Washington.

Editorial snapshot: 30 April 2026

The central finding

Grid operators, governments and the largest AI buyers converged in the first quarter on making large loads carry the risk of their own connection; most of those terms were still proposals at March 31, and the mechanism each jurisdiction adopts will decide who pays for load that never arrives.

01Key findings

What changed in energy and AI in the first quarter of 2026

AI reached the energy sector from two directions this quarter. As a source of demand, it produced most of the period's dated evidence: interconnection requests, forecasts, capital plans, prices and new connection rules. As a tool inside utility, grid and oil and gas operations, it produced far less.

Capability

Cheaper AI responses came with more AI use

Google said the cost of its core AI responses fell more than 30% after it moved to Gemini 3, while its direct API traffic rose from 10 billion to more than 16 billion tokens a minute in a quarter. Efficiency per response did not reduce total use, so it is a weak basis for trimming load forecasts.

Capital

Utilities and turbine makers raised plans on AI load

AEP disclosed a $72 billion five-year plan; Dominion, about $65 billion, up from $50 billion; and Southern Company, $81 billion. GE Vernova's gas-turbine backlog and slot reservations rose from 62 GW to 83 GW in the fourth quarter of 2025. AEP's latest 28 GW of new load is not yet in its plan.

Adoption

Requests far outran load approved to energize

Large-load requests in ERCOT rose from more than 233 GW in early January to about 410 GW by March 26, about 87% of them data centers. ERCOT's March status report showed 9,042 MW with approval to energize. Requests, utility contracts and forecasts measure different things and cannot be added.

Regulation

Large loads were asked to carry their own risk

After a year in which PJM's capacity cost per MWh rose 262.3%, the governors of all 13 PJM states, the PJM board, Texas regulators and seven AI companies converged on collateral, minimum bills, take-or-pay commitments and curtailment as conditions of connection. Most of those terms were proposals or plans at March 31.

Operating question

For each large load in your plans, at what stage of commitment does it count, and who pays for what was built for it if it never arrives?

02Capability

AI agents, serving costs and the capacity already on the grid

Two changes in AI capability bear on energy work. Agents can now operate ordinary software, which brings legacy operator tools and long engineering studies within reach, and the cost of serving AI fell without slowing the growth of total use. On the grid side, attention shifted to how fully existing lines are used and how quickly new loads can be studied and connected to them.

AI agents · Generally available

Agents can now operate software interfaces

OpenAI released GPT-5.4 on March 5 with native computer use, reporting 75.0% on the OSWorld-Verified benchmark. Anthropic's Claude Opus 4.6, released Feb. 5, added agent teams in Claude Code as a research preview and, in beta, a context window of 1 million tokens. These are the features that would reach operator consoles, GIS and outage systems, and long interconnection and load-flow studies.

Limitation: The benchmark score is OpenAI's own.

Expert work · Vendor benchmark

Most expert tasks remain out of reach

On Mercor's APEX-Agents benchmark, published Jan. 21, frontier models completed fewer than 25% of hours-long professional tasks in investment banking, consulting and corporate law. Operating an interface and finishing multistep professional work to an expert's standard remain different capabilities, which matters for study and filing work that a reviewer must sign off on.

Serving cost

Cheaper AI responses, more total use

Google said the cost of its core AI responses fell after it moved to Gemini 3, while traffic through its direct API kept rising. Token volume is not power draw, so the figures show the direction of demand rather than its size in megawatts.

More than 30%Fall in Google's cost per core AI response

10B to 16B+Direct API tokens per minute, fourth quarter to first

Grid utilization

Existing lines are fully loaded only a few hours a year

Brendan Pierpont of Energy Innovation told a Brookings panel on March 2 that grid infrastructure is fully loaded only a couple of hours a year. The IEA's Electricity 2026 report estimated how much queued capacity non-firm connections and grid-enhancing technologies could unlock. Microsoft said in January it would use AI to get more output from existing lines.

Limitation: Neither estimate is specific to AI loads.

About 40%Rough average grid utilization (Pierpont)

1,200–1,600 GWAdvanced-stage queued projects worldwide that flexible connection could unlock (IEA)

Cost and consequence

Serving efficiency did not shrink total AI use in the quarter, so planners have little reason to discount AI load forecasts for it. The nearer capacity gain lies in studying, connecting and curtailing loads on lines that already exist, which is software-heavy work and the natural entry point for AI-assisted planning tools.

03Capital

Utility capital plans and equipment backlogs rose on the same buyers' demand

Energy's AI capital this quarter sat in utility rate base, equipment order books and hyperscaler budgets rather than venture rounds. Three large utilities disclosed bigger capital plans in February, GE Vernova reported a gas-turbine queue it expects to largely sell out 2029 and 2030 production, and four AI companies kept spending heavily on the data centers behind the load.

Utility capital plans disclosed with fourth-quarter 2025 results

Each utility defines its large-load figure differently, so the figures cannot be added or compared directly.

Utility capital plans disclosed with fourth-quarter 2025 results
Utility Capital planChangeLarge-load figure, in the company's terms
AEP (Feb. 12) Capital plan $72B for 2026–2030 Change $5B–$8B more identified but excluded Large-load figure, in the company's terms 56 GW of new load by 2030 under signed agreements Up 28 GW since late October; not yet in the plan
Dominion Energy (Feb. 23) Capital plan About $65B over five years Change Up from $50B More than 90% of the increase is in Virginia Large-load figure, in the company's terms More than 48 GW in contracting stages More than 20 GW in signed service agreements and construction letters
Southern Company (Feb. 19) Capital plan $81B Plan period not stated Change Up 31% from a year earlier Large-load figure, in the company's terms Up to 75 GW of large-load prospects Across three electric companies

AEP and Dominion figures come from company materials and earnings calls; Southern figures are as reported by The Atlanta Journal-Constitution.

Equipment

Gas-turbine slots are being booked into 2030

GE Vernova reported a total backlog of about $150 billion at the end of 2025, up $31.2 billion in a year. Management expects about 100 GW of gas turbines under contract during 2026, which would largely sell out 2029 and 2030 production, and prices new slot reservations 10 to 20 points above the existing backlog.

62 GW to 83 GWGas-turbine backlog plus slot reservations, fourth quarter of 2025

About 20 GWAnnual gas-turbine production capacity from mid-2026

Hyperscalers

Four AI companies set the demand signal

Amazon, Alphabet, Microsoft and Meta reported their first-quarter capital spending on April 29. In February, Reuters put their combined 2026 plans above $630 billion, and Alphabet said it remains compute constrained in the near term.

Limitation: Company definitions of capital spending differ, so the figures are not strictly comparable.

$44.2BAmazon

$35.7BAlphabet

$31.9BMicrosoft, fiscal third quarter

$19.84BMeta

What it signals

Rate base, turbine order books and hyperscaler budgets rest on the same few buyers. AEP's latest 28 GW is not yet in its $72 billion plan, so the next round of rate cases will carry more investment justified by the same pipeline, and the generation behind it will be bought from order books priced for scarcity.

04Products

Upstream AI tools, data center equipment and staged connection contracts

The products described in the quarter show an energy sector selling to AI more than buying it. SLB counted customers for its upstream AI assistant and data platform while building a data center equipment line. GE Vernova sold grid equipment directly to data centers. Utilities and grid operators turned connection itself into staged, priced services.

Oil and gas

Upstream AI is sold as assistants and platforms

SLB said on Jan. 23 that Tela, an agentic assistant for upstream work launched less than three months earlier, had more than a dozen customers building a geoscience foundation model with SLB, and that more than 50 customers were adopting its Lumi data and AI platform. Both are measured so far by customer counts rather than operating results.

Equipment

Suppliers are retooling for data center buyers

SLB is targeting about $1 billion in annual run-rate revenue from data center solutions by the end of 2026 and plans to double manufacturing capacity for the line. GE Vernova said more than $2 billion of its 2025 Electrification orders were signed directly with data centers, more than triple the 2024 total.

Limitation: SLB's figure is a company target.

Connection

Connection has become a staged, priced service

Dominion contracts large loads in stages, from engineering letters to construction letters of authorization to signed service agreements (ESAs), with financial commitment rising at each step. PJM plans an expedited interconnection track for loads that bring their own new generation and a connect-and-manage service, and ERCOT is moving from single to batch studies.

Status of selected offerings at March 31

Status as described by each company or grid operator at the end of the quarter.

Status of selected offerings at March 31
Offering ProviderStatusReported scale or timing
Tela agentic assistant Provider SLB Status Commercial; launched late 2025 Reported scale or timing More than a dozen customers
Lumi data and AI platform Provider SLB Status Commercial Reported scale or timing More than 50 customers
Production System Optimization platform Provider SLB with ADNOC Status Announced by Jan. 23 Reported scale or timing No metrics disclosed
Data Center Solutions Provider SLB Status Commercial; expanding to Asia Reported scale or timing About $1B annual run-rate targeted by the end of 2026
Electrification equipment sold to data centers Provider GE Vernova Status Commercial Reported scale or timing More than $2B of 2025 orders
Staged large-load contracts Provider Dominion Energy Status In use Reported scale or timing More than 48 GW in contracting stages
Expedited Interconnection Track Provider PJM Status Planned Reported scale or timing By August 2026
Connect-and-manage service for large loads Provider PJM Status Planned Reported scale or timing By the end of 2026
Batch Zero batch studies Provider ERCOT Status Proposed Reported scale or timing board vote planned June 1; Aug. 1 target

Systems requirement

Each new connection term carries a systems requirement. Staged contracts need contract-stage tracking tied to the forecast, duplicate-request disclosure needs matching across markets, batch studies need repeatable study pipelines, and curtailable service needs telemetry, dispatch and settlement.

05Adoption

Large-load requests, contracts and forecasts diverged as PJM customers paid more

In energy, adoption this quarter meant AI load arriving on the grid, and its measures moved in different directions. Requests rose sharply, PJM cut its near-term forecast while raising the long-term outlook, and utility figures described as contracted grew on definitions ranging from letters of agreement to signed service agreements. In PJM, the cost of anticipated load had already reached customers through capacity prices.

Large-load measures at the end of the quarter

Requests, approvals, observed load, utility agreements, pipelines and forecasts measure different things. They cannot be added together or read as demand.

Large-load measures at the end of the quarter
Entity and measure Type of measureValueAs of
ERCOT requests, late March Type of measure Interconnection requests Value About 410 GW; about 87% data centers As of March 26
ERCOT requests, early January Type of measure Interconnection requests Value More than 233 GW; more than 70% data centers As of Reported Jan. 6
ERCOT new submissions Type of measure Interconnection requests Value 137 requests, about 140 GW (preliminary) As of March 13 report
ERCOT approval to energize Type of measure Operator approval Value 9,042 MW As of March 13 report
ERCOT approved loads, monthly peak Type of measure Observed consumption Value 3,801 MW Simultaneous peak of loads approved to energize As of March 13 report
AEP new load by 2030 Type of measure Utility agreements Value 56 GW Texas portion under letters of agreement As of Feb. 12
Dominion large loads Type of measure Staged contracts Value More than 48 GW; more than 20 GW signed As of December 2025
Southern Company prospects Type of measure Utility pipeline Value Up to 75 GW As of February
Microsoft in MISO Type of measure Buyer generation contracts Value 7.9 GW of new generation As of Jan. 13
PJM summer peak, 2036 Type of measure Grid-operator forecast Value 222,106 MW As of Jan. 14
U.S. load growth Type of measure National forecast (EIA) Value 1.9% in 2026; 2.5% in 2027 As of March 12

ERCOT's approval count has no stated start date, and its report labels the observed-peak month as March 2025. Microsoft says its MISO contracts are more than double its current consumption there.

What PJM wholesale customers paid in 2025

Monitoring Analytics, PJM's independent market monitor, in its 2025 State of the Market report released March 12.

What PJM wholesale customers paid in 2025
Component 2025 level or share of totalChange per MWhIncrease from 2024
Total wholesale price 2025 level or share of total $82.67/MWh Change per MWh +$27.15 Increase from 2024 48.9%
Energy 2025 level or share of total 59.6% of total Change per MWh +$16.69 Increase from 2024 51.2%
Capacity 2025 level or share of total 15.8% of total Change per MWh +$9.48 Increase from 2024 262.3%
Transmission 2025 level or share of total 22.4% of total Change per MWh +$0.80 Increase from 2024 4.5%

The monitor found the 2025/26, 2026/27 and 2027/28 capacity auctions not competitive, in significant part because of forecast data center demand, while finding the energy market competitive.

Grid operators

Forecasts now count commitments, not requests

Since March 1, ERCOT's forecasts for years after 2026 include only large loads with an executed interconnection agreement. PJM's January forecast counts only firm commitments in near-term years and derates longer-term requests; together with updated economic and electric vehicle forecasts, that lowered its near-term peaks, while its long-term outlook still rose.

Integrated utilities

Utilities count contracted load differently

AEP says all 56 GW of its new load is backed by signed agreements; its release describes the Texas portion, about 36 GW, as letters of agreement. Dominion says it does not forecast from load letters or inquiries, only from signed service agreements and construction letters. The two companies' figures are not the same measure.

PJM customers

In PJM states, the cost arrives through capacity prices

In restructured PJM states, anticipated data center load reaches customers through the capacity market rather than through a utility's rate base. The market monitor's Jan. 5 analysis attributed $21.3 billion, or 45% of the $47.2 billion cost of the last three base auctions, to forecast growth in data center load.

Definition: Counting all existing and planned data center load, the same analysis put the effect at $23.1 billion across the three auctions; the monitor repeated that figure in March.

Utility earnings

Rate base grows with the load it serves

For a cost-of-service utility, new load justifies new investment that earns an allowed return. Dominion said nearly two-thirds of its updated plan is eligible for rider recovery, subject to approval, and its investment base grows about 10% a year; AEP reaffirmed long-term operating earnings growth of 7% to 9%.

Economics

The utilities' growth guidance and the consumer protections in new large-load tariffs were being set at the same time. Minimum bills, collateral and exit fees decide whether investment made for a data center stays with that customer or moves to everyone else if the load does not arrive.

06Outcomes

Measured results came from forecasts, markets and suppliers

Published results for AI inside utility, grid and oil and gas operations in the quarter were limited to customer counts and supplier revenue. The measurable outcomes came from the power system's response to AI load: a forecast revision, the effect of a price collar on capacity costs, and the earnings of companies supplying the build-out.

PJM's 2026 summer peak forecast against its 2025 forecast

PJM's Jan. 14 long-term forecast, presented to its Planning Committee on Feb. 3. Ten-year growth is forecast at 3.6% a year, against 0.3% in the 2021 forecast.

PJM's 2026 summer peak forecast against its 2025 forecast
Forecast year Change from 2025 forecastApproximate change
2027 Change from 2025 forecast Down 2.3% Approximate change About 3,700 MW
2028 Change from 2025 forecast Down 2.6% Approximate change About 4.4 GW Both figures as reported by Utility Dive
2029 Change from 2025 forecast Down 2.7% Approximate change About 4,600 MW
2031 Change from 2025 forecast Down 0.9% Approximate change About 1,600 MW
2040 Change from 2025 forecast Up 7.8% Approximate change About 18,600 MW

PJM attributes the near-term reductions mainly to stricter vetting of large loads, with firm commitments required for near-term years, and also to updated economic and electric vehicle forecasts.

Capacity market

The price collar cut costs but added no supply

The PJM market monitor estimated that the capacity price collar reduced costs by $13.1 billion across two auctions, and that the latest auction would have cost $9.9 billion more without it. Jeff Plewes of CRA International wrote that a price cap does not bring additional megawatts and that the backstop procurement, still to be designed, carries significant uncertainty.

$13.1BCapacity costs avoided across two auctions

37.7%Cut in the latest auction's cost from the collar

Suppliers

Supplier results were commercial, not operational

SLB's digital business grew in 2025 and earned a 35% EBITDA margin; its CEO said customers were realizing drilling savings but gave no figure. GE Vernova's Power equipment backlog margins rose 11 points in 2025, which reflects scarcity pricing rather than productivity.

$2.7BSLB digital revenue in 2025, up 9%

Above $1BSLB digital annual recurring revenue, up 15%

Enterprise AI

Most CEOs report no significant financial benefit from AI yet

In PwC's 2026 Global CEO Survey, published Jan. 19 and fielded in late 2025, 56% of 4,454 chief executives across industries said they had seen no significant financial benefit from AI to date. The survey does not break out energy companies.

What it shows

The consequential measurements in the quarter were planning inputs. PJM's near-term revision of 2% to 3% cut its projected 2027 reserve shortfall to about 2.6 GW, Utility Dive reported, and capacity auctions clear against those projections.

07Corrections

Corrections to common claims about AI load, nuclear deals and power prices

Four claims that circulated in the quarter combine different measures. The distinctions change how a planner, regulator or buyer should use the figures.

Load · Requests, not demand

ERCOT's 410 GW are interconnection requests; about 9 GW had approval to energize

Claim under review: ERCOT has about 410 GW of data center demand.

The figure counts large loads seeking interconnection as of March 26, about 87% of them data centers. ERCOT's March 13 status report showed 9,042 MW approved to energize, and ERCOT's Kristi Hobbs said in December that ERCOT does not expect all requested load to materialize. Since March 1, Texas forecasts for years after 2026 count only loads with executed interconnection agreements.

Nuclear deals · Mostly future options

About 2.1 GW comes from operating plants; the rest depends on uprates and reactors planned for 2030 to 2035

Claim under review: Meta signed up to 6.6 GW of nuclear power in January.

Meta's Jan. 9 announcement combines more than 2.1 GW from Vistra's operating Perry and Davis-Besse plants in Ohio under 20-year agreements, 433 MW of uprates in the early 2030s, up to 690 MW from two TerraPower Natrium units as early as 2032 with rights to up to 2.1 GW more by 2035, and an Oklo campus of up to 1.2 GW from as early as 2030. Only the first is output from plants already running.

Prices · Different measure

The market monitor's per-MWh figure is a 48.9% increase

Claim under review: PJM power costs rose 56% in 2025.

Monitoring Analytics reported a total wholesale price of $82.67/MWh for 2025, up from $55.52/MWh in 2024. Press accounts of the same March 12 report gave total-cost increases of 54% (to $67 billion) and 56% (to $80.5 billion). Dollar totals rise with load as well as price, and the two press totals do not match each other.

Retail rates · Holds for PJM capacity

The evidence ties data centers to PJM's 2025 capacity prices, not to every retail rate increase

Claim under review: Data centers are raising electricity prices everywhere.

PJM's market monitor attributes the 2025 capacity price rise to data centers and found the auctions not competitive, in significant part because of forecast data center demand. At the same March 2 Brookings event, Ryan Wiser of Lawrence Berkeley National Laboratory said data centers played no role in California's increases and that the states with the largest retail increases over six years had contracting load. The two findings cover different markets, price components and periods.

08Accountability

The quarter's failures were in forecasting and process

The documented failures in energy this quarter were planning failures: a capacity auction cleared on a stale forecast, a review process was overwhelmed by requests, and new supply was withdrawn over connection costs. The quarter's AI security incidents did not involve energy systems, but they show how agents with operational access could be exploited.

Capacity auction · December 2025

PJM's auction cleared short on a year-old forecast

PJM's December 2025 auction, the first to clear short of its reliability requirement, fell more than 6.5 GW short and cleared at the $333/MW-day cap in every zone; PJM said it would otherwise have cleared at $523. It used a forecast released a year earlier, which PJM cut in January. The result prompted the federal-state principles and PJM's backstop process.

Interconnection review

ERCOT outgrew its large-load review process

ERCOT staff said in December 2025 that rules built for reviewing 40 to 50 large loads at a time had been outgrown. By March 26 the queue stood at about 410 GW. ERCOT is moving to batch studies, and in Texas a forecast rule took effect and an interconnection rule was proposed.

New supply

A gas unit was withdrawn over interconnection costs

A gas unit of about 1.5 GW in PJM's ATSI zone was withdrawn over high interconnection costs, CRA International reported in February. The cancellation removed planned supply from the region whose shortfall prompted January's intervention.

AI security · Risk

Agents' tool and file access was the path for exploits

Flaws in Anthropic's Git MCP server, fixed in December and disclosed Jan. 20, allowed remote code execution through indirect prompt injection, and a file-exfiltration flaw reached Anthropic's Cowork agent the same month. Neither involved energy systems; the risk for utilities is that the same kinds of tool and file permissions would reach operational systems.

Who is accountable

Markets and commissions now put a price on planning errors about AI load. For most energy companies that error, more than an AI tool failing in operations, is the larger near-term AI-related risk, and it belongs to whoever signs off on the forecast.

09Regulation

Large-load connection rules in the US and AI obligations in the EU

U.S. rules in the quarter concerned who pays for AI's load, and several institutions converged on the same conditions: size thresholds, financial commitment, commitment-based forecasts, curtailment and self-supply. The concrete changes came from Texas and PJM, while federal action was political and FERC's rulemaking remained pending. EU rules address different questions: how data centers report energy use and how AI used in grid operation is controlled.

Conditions for connecting large loads, by jurisdiction

PJM: the Jan. 16 board decisional letter and the Jan. 15 principles. Texas: the Public Utility Commission of Texas (PUCT) proposal of March 12 and ERCOT rules. Virginia: Dominion's terms under a November 2025 state order. Federal: DOE's October 2025 proposal to FERC and the March 4 pledge.

Conditions for connecting large loads, by jurisdiction
Condition PJMTexasVirginiaFederal and buyers
Size threshold PJM 50 MW at one interconnection point Texas 75 MW and above Virginia 25 MW, for minimum demand charges Federal and buyers 20 MW in DOE's proposal
Financial commitment PJM ESA or credit support for forecast inclusion Texas $50,000/MW security plus a non-refundable $50,000/MW fee Virginia Firm terms, exit fees and added collateral Federal and buyers Full participant funding (DOE); take-or-pay (pledge)
Forecast inclusion PJM Firm commitments in near-term years Texas Executed interconnection agreements, after 2026 Virginia Signed ESAs and construction letters Federal and buyers Verifiable commitments (principles)
Curtailment PJM Connect-and-manage by the end of 2026 Texas ERCOT may order curtailment or backup generation In energy emergencies Virginia — No comparable figure Federal and buyers Backup generation for emergencies (pledge)
Self-supply PJM Expedited track for loads bringing new generation Texas — No comparable figure Virginia — No comparable figure Federal and buyers Build, bring or buy new generation (pledge)
Duplicate requests PJM Forecast reforms for 2027 Texas Disclosure of similar requests elsewhere Virginia — No comparable figure Federal and buyers — No comparable figure

Blank cells mean no comparable term was reported. Most terms were proposed or planned at March 31.

Federal · Pending at March 31

FERC's large-load rule was still pending

DOE directed FERC in October 2025 to take final action on large-load interconnection by April 30, 2026. The advance notice of proposed rulemaking included a 20 MW threshold and full participant funding of network upgrades. Its claim of federal jurisdiction is the main dispute with states, whose regulators say they keep authority over retail load interconnection of any size.

PJM region · Non-binding

13 governors and the White House set principles for PJM

On Jan. 15, Interior Secretary Doug Burgum and Energy Secretary Chris Wright, for the National Energy Dominance Council, and the governors of all 13 PJM states signed principles calling for a reliability backstop auction no later than September, an extended price collar, backstop costs assigned first to load-serving entities whose data centers did not self-procure or accept curtailment, and a 150-day limit on interconnection studies. PJM tariff changes still require FERC filings.

Texas · Proposed; comments due April 17

Texas proposed upfront costs and 80% forfeiture

The PUCT's proposed rule under Senate Bill 6 would charge study fees of $100,000 for loads of 75 to 250 MW and $300,000 above that, require all direct interconnection costs upfront, and forfeit 80% of remaining security to the utility's rate base if a project withdraws or misses a milestone by six months. Greenberg Traurig reported concerns from customers and from transmission utilities, whose earnings would fall because forfeited security and fees offset rate base.

Buyers · Voluntary

Seven AI companies pledged to pay their own way

Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed the White House Ratepayer Protection Pledge on March 4, committing to build, bring or buy new generation, pay for delivery upgrades, accept separate rates and pay for power and infrastructure whether or not they use it. The White House described no penalties or monitoring to enforce it.

EU AI Act · Applies Aug. 2, 2026 under current law

AI in grid operation is high-risk under the AI Act

Annex III classes AI used as a safety component in the management and operation of the supply of electricity, gas and heating as high-risk. Council and Parliament positions on the Digital Omnibus, adopted March 13 and March 26, would move those obligations to Dec. 2, 2027; negotiations had not concluded by March 31.

EU data centers · Annual report due May 15

EU data centers report energy performance every May

The Energy Efficiency Directive requires owners and operators of data centers with at least 500 kW of installed IT power demand to report energy performance each year by May 15 to a European database. On March 26 the Commission opened a call for feedback on a draft EU-wide data center rating scheme, running to April 23.

Scheduled dates as of March 31

Scheduled dates as of March 31
Item DateJurisdictionStatus at March 31
PUCT forecast rule (Project 58480) Date March 1, 2026 Jurisdiction Texas Status at March 31 In effect
Comments on the PUCT large-load rule (Project 58481) Date April 17, 2026 Jurisdiction Texas Status at March 31 Proposal
FERC final action on large loads, as DOE requested Date April 30, 2026 Jurisdiction U.S. federal Status at March 31 Pending
Data center energy reporting, 500 kW and above Date May 15, annually Jurisdiction EU Status at March 31 In law
ERCOT board vote on Batch Zero Date June 1, 2026 Jurisdiction ERCOT Status at March 31 Planned
PJM 2028/29 base capacity auction Date Opens June 30, 2026 Jurisdiction PJM Status at March 31 Scheduled
ERCOT Batch Zero protocols Date Aug. 1, 2026 (target) Jurisdiction ERCOT Status at March 31 Planned
AI Act high-risk obligations, including grid operation Date Aug. 2, 2026 Jurisdiction EU Status at March 31 In law; deferral to Dec. 2, 2027 proposed
PJM Expedited Interconnection Track Date August 2026 Jurisdiction PJM Status at March 31 Planned
PJM reliability backstop auction Date By September 2026 Jurisdiction PJM Status at March 31 Requested by the principles
PJM connect-and-manage for large loads Date End of 2026 Jurisdiction PJM Status at March 31 Planned
Dominion minimum demand charges, 25 MW and above Date January 2027 Jurisdiction Virginia Status at March 31 Approved

What applies

At March 31, Texas' forecast rule was the only new large-load rule already in effect, and Virginia's minimum demand charges were approved to start in January 2027. The other large-load terms were proposals, plans or requests with dates between April 2026 and 2027, which leaves the terms large loads will actually face to filings and votes still ahead.

10Voices

Where officials, utilities, buyers and researchers agree and differ

The positions converge on three points: data centers should pay for the power and grid capacity they cause, unscreened requests overstate demand, and supply is the binding constraint. They divide on what is raising bills, what PJM's forecast cut means, which mechanism should make data centers pay, how contracted load should be counted and who has jurisdiction. Interests differ as well: utilities and equipment suppliers benefit from high forecasts, and buyers from faster connection and from deflecting blame for rising bills.

Federal and state officials

The officials behind the PJM principles, and the state regulators and consumer advocates contesting how large loads should pay and who decides.

Chris Wright

U.S. Secretary of Energy

“High electricity prices are a choice”

On Jan. 16 he named PJM the region most at risk and backed an emergency auction for more than $15 billion of baseload generation. Interior Secretary Doug Burgum said technology companies, not taxpayers, would fund the plants.

Federal government · DOE release

Wes Moore

Governor of Maryland

“We cannot build a 21st-century economy on an energy market that blocks new supply”

His Jan. 16 release presented the principles as freezing new-generation incentive costs through the end of the decade and assigning them to data centers through special rates.

State government · governor's release

National Association of Regulatory Utility Commissioners

Association of state utility regulators

“State public utility commissions retain jurisdiction over retail load interconnections, regardless of size”

Its November 2025 comments on DOE's proposal asked FERC to state that it is not asserting jurisdiction over distribution interconnections, generation facilities or end-use sales. DOE's proposal claims federal jurisdiction over large-load interconnection.

State regulators · FERC docket comments

Pennsylvania Office of Consumer Advocate and Delaware Division of the Public Advocate

State consumer advocates; Pennsylvania's office is led by Consumer Advocate Darryl Lawrence

Large loads should pay 100% of the network upgrades assigned to them with no crediting, face an 85% minimum demand charge and connect only if they bring enough capacity.

Filed in November 2025 on DOE's proposal. On transmission, the advocates and the PJM market monitor take opposite positions.

Consumer advocates · FERC docket comments

Grid operators and the PJM market monitor

The institutions running the two grids at the center of the quarter, and the independent monitor of PJM's markets.

Jason Connell

Vice President of Planning, PJM

Load growth in PJM's forecast is still faster than at any time in recent decades.

Said with the Jan. 14 forecast, which lowered near-term peaks and raised the long-term outlook.

Grid operator · PJM publication

Kristi Hobbs

Vice President of System Planning and Weatherization, ERCOT

ERCOT has outgrown the process it built for reviewing large loads and does not expect all requested load to materialize.

Speaking to ERCOT's board in December 2025. By March 26 the queue was about 410 GW.

Grid operator · reported remarks

Joseph Bowring

President, Monitoring Analytics, PJM's independent market monitor

“it's nothing but data centers”

On PJM's 2025 capacity price rise, rejecting EVs and electrification as explanations. He wants data centers to bring their own new generation or buy it through a separate auction, opposes off-grid co-location and argues that transmission costs should be shared by all load.

Market monitor · Brookings event

Utilities

Two large utilities that count the same kind of demand in different ways.

Bill Fehrman

CEO, American Electric Power

Called the growth a generational load phenomenon and said AEP's 56 GW is not speculative because all of it is backed by signed agreements.

AEP's results release describes the Texas portion, about 36 GW, as resting on signed letters of agreement.

Utility · earnings call, as reported by T&D World

Bob Blue

CEO, Dominion Energy

Said data center demand can be a win for Virginia, customers and the company, and that Dominion forecasts only from signed service agreements and construction letters, not from load letters, inquiries or engineering letters.

A difference in method from AEP rather than opposition to data centers. Dominion also says its Virginia residential rates are about 4% below the national average.

Utility · earnings call transcript

Large-load buyers

A buyer's January commitment to pay the full cost of serving its data centers.

Brad Smith

Vice Chair and President, Microsoft

Committed Microsoft to ask utilities and commissions to set its rates high enough to cover the full cost of serving its data centers, including infrastructure added for them, so that cost is not passed to residential customers.

He attributed recent residential rate increases in part to inflation, supply-chain constraints and grid upgrades. Microsoft backs a proposed Very Large Customer tariff in Wisconsin.

Buyer · company statement

Equipment and services suppliers

Suppliers with a direct commercial interest in AI load, whose own caveats are informative.

Scott Strazik

CEO, GE Vernova

Cautioned that the plan for about 100 GW under contract in 2026 does not assume many long-dated framework agreements close this year, that nuclear timing will be slower and that changes to market mechanisms will take time.

Said alongside a backlog of about $150 billion and a gas-turbine queue that management expects to largely sell out 2029 and 2030 production.

Supplier · earnings call

Olivier Le Peuch

CEO, SLB

“The opportunity is growing faster than anticipated”

On SLB's data center business, in the same call in which he said near-term oil oversupply may keep pressure on prices in the first half of 2026 and operators would stay cautious.

Supplier · earnings call

Independent researchers and analysts

Researchers and analysts who dispute parts of the load and price story, each speaking to a specific market or period.

Ryan Wiser

Lawrence Berkeley National Laboratory

Retail price increases vary widely by state; data centers played no role in California's increases, and the states with the largest increases over the past six years had contracting load.

Speaks to national retail prices over several years, where Bowring speaks to PJM's wholesale capacity prices in 2025.

Researcher · Brookings event

Abe Silverman

Johns Hopkins University; former New Jersey Board of Public Utilities staff

Data centers and the capacity prices they produced hit Maryland hardest; he is skeptical of demand from an industry that does not yet know what it wants to be and says data centers should pay their own way.

Reads PJM's forecast cut as a sign of immature demand, where Jefferies reads it as delay.

Researcher · Brookings event

Brendan Pierpont

Energy Innovation

Puts the long-term rise in transmission and distribution costs ahead of data centers as a driver of bills, and favors flexibility and fuller use of existing infrastructure.

Energy Innovation is an advocacy think tank.

Policy researcher · Brookings event

Jefferies

Equity research analysts

Read PJM's near-term forecast revisions as delays and pushouts rather than weaker demand, and said Microsoft's commitment would pressure other hyperscalers to follow.

PJM attributes the cuts mainly to stricter vetting of large loads, and also to updated economic and electric vehicle forecasts.

Sell-side analysts · analyst note, as reported

11Decisions

Three decisions for energy leaders

Each decision follows from a gap the quarter exposed between how load is counted, who pays for it and which systems support it.

Head of load forecasting, utility or grid operator

Decide which commitment stage counts in the forecast

AEP's 56 GW includes load under letters of agreement, Dominion counts only signed service agreements and construction letters, ERCOT counts only executed interconnection agreements after 2026, and PJM requires firm commitments in near-term years. Commissions and market monitors are now asking how forecasts were built.

At which stage of commitment does a large load enter our forecast, and could we defend that choice to our commission and market monitor?

Chief financial officer, regulated utility

Decide where the risk of unrealized load sits

Minimum demand charges, collateral, exit fees, participant funding and rider recovery each place the cost of load that never arrives on a different party. Utility growth guidance was set while those terms were still being negotiated.

If a contracted large load does not arrive, which customers pay for the investment made for it under the tariffs we have filed or plan to file?

Chief technology or information officer

Decide where AI enters planning work and what it can reach

Agents that operate software became generally available in the quarter, and the documented exploits ran through the tool and file access agents had been granted. The new connection rules create study, contract-tracking and curtailment work on fixed cycles, which is where AI-assisted tools would contribute first and where access to operational systems has to be bounded.

Which planning and study workflows will we automate first, and which operational systems, if any, will those agents be allowed to reach?

12Actions

Four actions for the second quarter

  1. 1

    Tag every large-load request by commitment stage

    Record each request's stage, from inquiry and load letter through engineering and construction letters to service and interconnection agreements, and report totals by stage. The result is a forecast that reconciles to ERCOT and PJM inclusion rules and holds up in a rate case.

    Load forecasting lead

  2. 2

    Keep a dated calendar of large-load terms

    Track the Texas rule; ERCOT's Batch Zero vote on June 1; PJM's expedited track, backstop auction and connect-and-manage service; Virginia's January 2027 minimum demand charges; and the EU AI Act dates, with an owner for each. The result is early warning on which terms become binding and when.

    Regulatory affairs

  3. 3

    Make large-load studies repeatable on a batch cycle

    Standardize study inputs and outputs so a batch can be rerun on ERCOT's six-month cadence, and match requests against similar filings elsewhere, as the Texas proposal would require. The result is study throughput that keeps pace with batch deadlines.

    Interconnection engineering

  4. 4

    Inventory AI that can reach grid-operation systems

    List every AI system and agent connector with access to operator consoles, GIS or outage systems, test each against indirect prompt injection before widening permissions, and classify systems used in EU electricity or gas operations against AI Act Annex III. The result is a register that serves both security review and EU high-risk documentation, which applies from Aug. 2, 2026, under current law, or from Dec. 2, 2027, if the proposed deferral is adopted.

    OT security and AI governance

13References

References

Sources are publications through April 30, 2026. Those published after March 31 are used only for what they report about the first quarter; those published before January are background.

Grid operators and power markets

Regulation and policy

Utilities and energy suppliers

AI companies and buyers

AI capability and security

Research and analysis

Thoughtful engineering for what’s next.

Have a question about what this means for your team? Let’s talk it through.

Start a conversation

Have a project in mind, or just exploring an idea? Let’s talk.
Fill out the form or give us a call.

413-570-3299

What can we help you with?