INTELLIGENCE REPORT SERIES NOVEMBER 2026 OPEN ACCESS

SERIES: ECONOMIC INTELLIGENCE

The Data-Centre Power Grab: $29.4B and Who Pays (2026)

Data-centre demand added $29.4 billion to capacity costs on the largest US grid in four auctions. Analysis covers tucson, naacp, facility. Evidence-sourced,...

Reading Time50 min
Word Count9,822
Published15 November 2026
Evidence Tier Key → ✓ Established Fact ◈ Strong Evidence ⚖ Contested ✕ Misinformation ? Unknown
Contents
50 MIN READ
EN FR JP ES DE ZH AR RU PT

Data-centre demand added $29.4 billion to capacity costs on the largest US grid in four auctions. Analysis covers tucson, naacp, facility. Evidence-sourced,...

01

The Bill Arrives Before the Load
How AI’s electricity demand reaches the household meter

Global data-centre electricity use rose 17% in 2025, more than five times the 3% growth of electricity demand as a whole ✓ Established [3]. The question this report answers is not whether that load is large — it is — but who is paying to serve it. On the evidence from the United States, Ireland and Malaysia, a significant share of the cost reaches ordinary customers before the servers are switched on.

Start with the scale. The International Energy Agency estimates that data centres consumed about 415 terawatt-hours of electricity in 2024 and projects roughly 945 TWh by 2030, slightly more than the entire current consumption of Japan ✓ Established [1]. Its April 2026 update found that consumption had already reached about 485 TWh in 2025, after a single year of 17% growth, and that demand from AI-focused facilities is set to triple by the end of the decade ✓ Established [3]. The United States accounts for the largest single increment: the IEA expects American data-centre consumption to rise by up to 240 TWh, or 130%, between 2024 and 2030, while China adds about 175 TWh, a rise of 170% [1]. Few industrial sectors have ever been forecast to more than double their electricity use within six years, and none has done so while also asking utilities to build the supply on its behalf.

In the United States the numbers are already visible in national statistics. Lawrence Berkeley National Laboratory, in a report commissioned by the Department of Energy, found that data centres used 58 TWh in 2014 and 176 TWh in 2023, equal to 4.4% of all American electricity ✓ Established [4]. Its scenarios put 2028 consumption between 325 and 580 TWh, or 6.7% to 12% of national demand [4] [5]. The width of that range is itself the story: a gap of 255 TWh between the low and high cases is larger than the annual consumption of most European countries, and utilities must decide now which end of the range to build for. Whatever they choose, the cost of the decision is recovered from customers through regulated rates, which is why the forecast, not the finished building, is where the bill begins.

485 TWh
Global data-centre electricity use in 2025
IEA, April 2026 · ✓ Established
17%
Growth in data-centre demand in 2025, against 3% overall
IEA, April 2026 · ✓ Established
$725B
Planned 2026 capital spending by four hyperscalers
Financial Times via Tom’s Hardware, 2026 · ✓ Established
6.7–12%
Projected US data-centre share of electricity in 2028
LBNL, December 2024 · ✓ Established

The money behind the load is unprecedented. Amazon, Alphabet, Microsoft and Meta plan combined capital expenditure of about $725 billion in 2026, a 77% increase on the record $410 billion of 2025, according to a compilation of first-quarter guidance by the Financial Times ✓ Established [42]. Most of it is going into graphics processors, custom chips and the buildings and substations that house them [42]. Spending at that pace creates a timing mismatch with the power sector. A hyperscale campus can be built in two to three years; gas turbines ordered in 2026 are being offered for delivery slots as late as 2031, and new transmission routinely takes a decade [39]. Utilities and grid operators are therefore committing to generation and wires on the strength of announcements and interconnection requests, many of which will never be built, and recovering the cost from every customer on the system.

The channels through which data-centre demand reaches residential bills are now well documented. None of them requires a data centre to consume a single kilowatt-hour from a household’s supply. They work through the shared costs of a regulated system — capacity, wires, new plants and discounts — that are spread across every customer class in proportions set by regulators, often years before the load materialises ◈ Strong Evidence [13] [15] [16]. The four main routes are:

  • Capacity prices. Where grid operators run capacity auctions, forecast data-centre demand raises the price paid for every megawatt of reserve, and retail suppliers pass that price to all customers [11] [13].
  • Connection costs. Substations and local transmission lines built to serve a single campus are often rolled into general transmission charges rather than billed to the customer that triggered them [15].
  • New generation. Vertically integrated utilities build gas plants sized for data-centre demand and recover the capital through the regulated rate base over 30 to 40 years [35] [37].
  • Special contracts. Bespoke electricity rates negotiated with technology companies can fall below the cost of service, leaving the gap to other ratepayers [16].

Each route is legal, and each has a defensible rationale. Capacity markets exist to send scarcity signals; socialised transmission reflects the principle that a meshed grid serves everyone; new plants are needed when demand grows. The problem is cumulative and distributional. When a single customer class — data centres — drives most of the incremental load, cost-allocation rules written for diffuse demand growth shift a disproportionate share of the bill onto households and small businesses that did not cause it ◈ Strong Evidence [15] [16]. The Harvard Electricity Law Initiative describes this as utilities using their monopoly position to offer discounts to technology firms and recover the shortfall from the public [16]. Regulators in Ohio, Virginia, Texas and Georgia have begun rewriting those rules since mid-2025, which is itself an admission that the old ones did not fit [9] [20] [21] [37].

✓ Established Data-centre electricity demand grew more than five times faster than total demand in 2025

The IEA’s April 2026 analysis found data-centre consumption up 17% in 2025 to roughly 485 TWh, against 3% growth in global electricity demand, with AI-focused facilities growing faster still [3]. Its base case still projects about 945 TWh by 2030 [1]. In the United States, Berkeley Lab estimates data centres already used 4.4% of national electricity in 2023 [4].

Households are noticing. The Energy Information Administration expects the average US residential price to rise from 17.29 cents per kilowatt-hour in 2025 to 18.02 cents in 2026 ◈ Strong Evidence [19]. National averages conceal sharper local movements. A Bloomberg News analysis of 25,000 wholesale pricing nodes found prices up by as much as 267% over five years in some locations, and that more than 70% of the nodes recording increases sat within 50 miles of significant data-centre activity [17]. Wholesale costs are only the supply portion of a retail bill, typically 30% to 50%, so the retail effect is smaller than the headline [17]. But the direction is consistent across sources, and in the PJM region — 13 states and Washington DC — the market monitor has now put a figure on it [13].

The remaining sections follow the cost. Section 2 examines why load forecasts have run so far ahead of what is actually being connected. Section 3 traces how those forecasts became capacity prices and transmission charges in PJM. Section 4 turns to water, the second resource data centres draw from public systems. Section 5 documents the revival of gas and nuclear deals built for server farms, and Section 6 compares the United States with Ireland, the Netherlands, Japan, Singapore, Malaysia and China. Sections 7 and 8 assess the regulatory response and the genuine disagreements about whether data centres raise or lower bills over the long run. The argument throughout is narrow and testable: the costs of AI’s electricity are real, they are being allocated by rules written for a different kind of growth, and the allocation, not the technology, is the policy choice.

02

Phantom Gigawatts
Load forecasts against what is actually being connected

In three years the five-year US peak-load forecast rose more than sixfold, from 24 gigawatts to 166 GW, with about 90 GW attributed to data centres ✓ Established [6]. Analysts who track the buildings themselves think the data-centre portion is overstated by roughly 25 GW ◈ Strong Evidence [6]. The difference matters because utilities build, and bill, against the forecast.

Utility load forecasts used to be among the dullest documents in the power sector. For two decades American electricity demand was essentially flat, and five-year projections moved by a few gigawatts a year. That changed abruptly. Grid Strategies, which compiles utility filings to the Federal Energy Regulatory Commission and state commissions, found the aggregate five-year forecast of peak growth rose from 24 GW to 166 GW over three years, an annual growth rate of 3.7% ✓ Established [6]. Roughly 90 GW, or about 55% of the total, was linked to data centres [6]. Six regions accounted for more than 80% of the projected growth, concentrating the effect on the customers of a handful of utilities and grid operators in Virginia, Texas, Georgia, Ohio, Arizona and the Pacific Northwest [6].

The same report contains its own warning. Comparing utility forecasts with market analysts who track chip shipments, construction starts and signed leases, Grid Strategies concluded that the data-centre portion of utility forecasts is likely overstated by about 25 GW, and that data-centre growth is unlikely to require much more than 65 GW through 2030 ◈ Strong Evidence [6]. The mechanism is not mysterious. A developer seeking a site will often file connection requests with several utilities at once, each for the full load of the planned campus. Each utility records the request in its own forecast. Aggregated nationally, a single 1 GW project can appear as three or four gigawatts of demand, and none of the utilities has an incentive to discount it, because under cost-of-service regulation a larger forecast justifies a larger capital programme and a larger rate base.

Texas offers the starkest illustration. The Electric Reliability Council of Texas reported large-load interconnection requests above 400 GW during 2026, roughly 90% of them from data centres, after a queue that nearly quadrupled in a single year ✓ Established [7] [8]. For comparison, the ERCOT system’s all-time peak demand is below 90 GW. No plausible scenario connects even half the queue within a decade. Yet each request must be studied, and some fraction will be built, and the grid operator has to plan transmission for the fraction it judges real. The cost of that judgement — whether too generous or too cautious — lands on the 27 million people served by the Texas grid, either as higher transmission charges or as reliability risk if the estimate proves too low [7] [21].

Forecasts Are Priced Before They Are Built

A load forecast is not a neutral number. In a capacity market it sets the quantity of reserve the system must buy; in a vertically integrated state it justifies new plants that enter the rate base. Either way, households begin paying for forecast data-centre demand years before the buildings exist, and they keep paying if the buildings never arrive. The test of any reform is whether it moves that risk back to the developer who created it.

Ohio provides the cleanest natural experiment in what happens when developers are asked to put money behind their requests. In July 2025 the Public Utilities Commission of Ohio approved a tariff requiring new data-centre customers above 25 MW in AEP Ohio’s territory to pay for at least 85% of their subscribed capacity for up to 12 years, whether or not they use it ✓ Established [9]. The utility subsequently told regulators that its pipeline had fallen from more than 30 GW of proposed load to 5,642 MW of signed, financially committed agreements [10]. More than four-fifths of the queue disappeared once it carried a price. That is not evidence that data-centre demand is fictitious — 5.6 GW is still several large power stations — but it is strong evidence that unpriced queues overstate it by a wide margin.

In PJM, the grid operator serving 65 million people from Illinois to New Jersey, forecast data-centre load feeds directly into the capacity auction. Between the auction held in mid-2025 and the one held in December 2025, PJM’s peak forecast for the 2027/2028 delivery year rose by about 5,250 MW, of which nearly 5,100 MW was attributed to data centres ✓ Established [11] [12]. The independent market monitor, Monitoring Analytics, has argued repeatedly that much of that forecast reflects load that is neither under construction nor contracted, and that including it raises prices for everyone [13] [14]. PJM has responded by tightening the evidence utilities must provide before load enters the forecast, but the auctions already cleared under the older rules cannot be re-run [53].

◈ Strong Evidence Forecast data-centre load, not only operating load, is raising capacity costs

Monitoring Analytics, PJM’s independent market monitor, attributes $29.4 billion of additional capacity-market revenue across the last four auctions to existing and forecast data-centre load [13]. In the December 2025 auction, nearly 5,100 MW of the 5,250 MW increase in forecast peak demand came from data centres [11] [12]. Because capacity is bought three years ahead, customers pay for forecast load whether or not it is later connected.

Forecasting errors in either direction are costly, and the risk is not symmetrical for the companies involved. If a utility overbuilds, the stranded cost sits in the rate base and is recovered from all customers for decades. If it underbuilds, it faces reliability penalties and political blame, while the data-centre developer simply moves to another state. Entergy made exactly this argument to Louisiana regulators in 2025, warning that delay risked Meta locating its largest data centre elsewhere [35] [36]. The bargaining power runs to the party with the most mobile capital. That is why the forecasting problem is ultimately a contract problem: only binding, collateralised commitments from data-centre customers convert speculative load into something a regulator can safely build for [9] [20].

The evidence therefore supports two claims that are often presented as contradictory. Data-centre demand is genuinely large and growing faster than any other source of electricity use in the developed world ✓ Established [3] [4]. And the load forecasts on which utilities are currently spending money overstate that growth by tens of gigawatts ◈ Strong Evidence [6] [10]. Both can be true at once, and the policy implication follows from holding them together. The task is not to decide whether AI demand is real; it is to ensure that the parties who make speculative requests bear the cost of the speculation, rather than the households who are offered no choice about the size of the system they pay for.

03

How the Cost Travels
Capacity auctions, connection charges and confidential discounts

PJM’s December 2025 capacity auction cleared at its price cap of $333.44 per megawatt-day, cost $16.4 billion and still fell 6,623 MW short of the reliability requirement for the first time in its history ✓ Established [11] [12]. The market monitor attributes $29.4 billion of capacity revenue over four auctions to data-centre load ✓ Established [13].

The PJM capacity market is the single clearest window onto who pays for data-centre growth, because its prices are public and its market monitor publishes attributions. Capacity payments compensate power plants for being available at peak, and the cost is recovered from every customer through retail supply charges. In the auction for the 2027/2028 delivery year, held in December 2025, the price cleared at the FERC-approved cap of $333.44 per megawatt-day across the entire footprint ✓ Established [11] [12]. The total bill was $16.4 billion [11]. Despite the ceiling price, the auction procured 6,623 MW less unforced capacity than PJM’s reliability requirement — the first time the whole system had cleared short [11] [12]. A market at its cap and still short is a market in which demand has outrun supply faster than any price signal can correct.

The market monitor has quantified how much of this is due to data centres. In reports through 2026, Monitoring Analytics found that existing and forecast data-centre load increased capacity-market revenues by a combined $29.4 billion across the four most recent auctions ✓ Established [13]. In the energy market, it estimated that data centres accounted for about 9% of total PJM wholesale costs through July 2026, equal to $10.48 per megawatt-hour, at a time when total wholesale costs rose 46% to $56.7 billion in the first seven months of the year [13]. The monitor’s conclusion is direct: the total will continue to grow until the issues associated with adding large data-centre loads are addressed [13]. For a typical household, capacity charges are a line item it has never heard of, folded into a supply rate that rises at the next annual reset.

$29.4B
Capacity revenue added by data-centre load over four PJM auctions
Monitoring Analytics, 2026 · ✓ Established
$4.3B
Data-centre connection costs passed to PJM customers in 2024
Union of Concerned Scientists, Sept 2025 · ✓ Established
267%
Maximum five-year wholesale price rise near data-centre clusters
Bloomberg News, Sept 2025 · ◈ Strong Evidence
6,623 MW
Shortfall against PJM’s reliability requirement, 2027/2028
PJM, December 2025 · ✓ Established

The second route runs through the wires. The Union of Concerned Scientists examined 130 transmission projects approved in 2024 to connect data centres in seven PJM states and found that more than 95% passed their costs to local utility customers rather than to the data centres that required them ✓ Established [15]. The total was $4.3 billion in a single year. By state, Virginia accounted for $1.9 billion, Ohio $1.3 billion, Pennsylvania $491.7 million, Illinois $239 million, West Virginia $215.8 million, Maryland $107.5 million and New Jersey $14.5 million [15]. These costs do not appear in the capacity auction; they sit in transmission rates, which are recovered from all customers in a utility zone. UCS noted that they come on top of an estimated $9.3 billion in data-centre-related capacity costs for PJM customers in the delivery year from June 2025 [15].

The mechanism is a gap between federal and state regulation. Connection facilities for a large load are frequently classified as network transmission, which FERC-approved formula rates recover from the whole zone, rather than as customer-specific facilities charged to the end user [15]. Mike Jacobs, a senior energy manager at UCS, described the result as technology companies exploiting an outdated regulatory process to pass billions of dollars of costs to families [15]. Utilities respond that connection lines often strengthen the wider network and that assigning every dollar to the triggering customer would deter investment. Both points have merit, but the UCS data suggest that the default has been to socialise almost everything, and that the exception — 5% of projects charged directly — is the outcome that a cost-causation principle would treat as the norm ◈ Strong Evidence [15].

The solution is to make sure that people who want to build data centers are serious enough about it to bring their own generation.

— Joseph Bowring, President, Monitoring Analytics (PJM independent market monitor), August 2025 [54]

The third route is the least visible. Researchers at the Harvard Electricity Law Initiative reviewed utility filings and identified dozens of special contracts between utilities and data-centre operators, negotiated outside the normal rate case ✓ Established [16]. When the contract rate is below the cost of serving the customer, the utility can recover the difference from other ratepayers. Many of the contracts are filed under confidentiality claims, which makes the size of any subsidy impossible for the public to measure [16]. The authors, Ari Peskoe and Eliza Martin, recommend that state commissions move new data-centre service away from bespoke contracts and onto published tariffs that apply to every customer of the same size [16]. Several commissions have since done so, as Section 7 shows, but the contracts already signed run for a decade or more.

Bloomberg’s pricing analysis shows where these mechanisms converge geographically. Of the 25,000 wholesale nodes it examined across seven grid operators, more than 70% of those recording price increases were within 50 miles of significant data-centre activity, and some locations saw increases of up to 267% over five years ◈ Strong Evidence [17]. Proximity does not prove causation — fuel prices, plant retirements and transmission congestion also move nodal prices — and Bloomberg itself stressed that wholesale prices are only part of a retail bill [17]. But the spatial pattern is consistent with the market monitor’s attribution and with the transmission data. In the regions with the most data-centre construction, the price of electricity at the point where it is delivered has risen fastest. The summary below sets out the four cost paths and the evidence attached to each:

  • Capacity: $29.4 billion over four PJM auctions attributed to data-centre load by the market monitor ✓ Established [13].
  • Energy: about 9% of PJM wholesale costs through July 2026, or $10.48 per megawatt-hour ✓ Established [13].
  • Transmission: $4.3 billion of 2024 connection costs, over 95% socialised across seven states ✓ Established [15].
  • Contracts: dozens of confidential special rates whose subsidy, if any, cannot be publicly measured ◈ Strong Evidence [16].

None of this means every data centre raises bills, or that every increase in PJM is caused by data centres. Gas prices, the retirement of coal plants, and slow interconnection of new generation all contribute, and Section 8 examines evidence that load growth has historically lowered average rates ⚖ Contested [18]. What the PJM record establishes is narrower and harder to dispute. In the one large market where an independent monitor publishes attributions, data-centre demand — much of it forecast rather than operating — is the largest single driver of a capacity-price spike that every customer pays, and the rules in force until 2025 offered households no protection from it ✓ Established [13] [14].

04

The Water Line
Cooling, drought and the disclosure gap

US data centres consumed an estimated 17 billion gallons of water directly for cooling in 2023, and the indirect water used to generate their electricity was about 12 times larger ✓ Established [22]. Google alone reported freshwater consumption of 10.9 billion gallons in 2025, up 34% in a year ✓ Established [23].

Electricity is the resource that dominates the data-centre debate, but water is the one that most often turns local opinion. Servers convert almost all the power they draw into heat, and the cheapest way to remove heat in most climates is evaporation. Berkeley Lab’s 2024 report estimated that US data centres consumed about 17 billion gallons, or 64 billion litres, directly through cooling in 2023, and projected that the figure could double or even quadruple by 2028 as hyperscale and AI facilities expand ✓ Established [22] [4]. That direct figure is the smaller part. The laboratory estimated that the water consumed upstream — in the thermal power stations generating the electricity data centres use — was roughly 12 times greater than on-site cooling in 2023 [22]. A data centre on a gas- or coal-heavy grid draws on rivers and aquifers twice: once at its own cooling towers, and again at the power plant.

Corporate disclosures confirm the trajectory. Google’s 2026 environmental report shows total electricity consumption up 37% in 2025, its largest annual increase, and more than 250% higher than in 2019 ✓ Established [23]. Freshwater consumption reached 10.9 billion gallons, up 34% on 2024, with data centres responsible for most of the increase [23]. In 2024 the company’s campus at Council Bluffs, Iowa, consumed about 1 billion gallons on its own [22]. Google is among the more efficient operators — it reports a fleet-wide power usage effectiveness of 1.09, meaning only 9% of energy goes to overheads such as cooling [23]. The industry average, according to the Uptime Institute’s 2025 survey, has been stuck at 1.54 for six consecutive years [24]. Efficiency gains at the best facilities are being swamped by the sheer volume of new capacity.

✓ Established The water used to generate data-centre electricity is about 12 times the water used on site

Berkeley Lab estimated direct cooling consumption by US data centres at about 17 billion gallons in 2023 and indirect consumption at power plants at roughly 12 times that [22]. Company disclosures generally report only the on-site figure, so the public accounting of data-centre water use captures a small fraction of the total ✓ Established [22] [23].

The local politics of water are sharper than the national statistics suggest, because data centres cluster where land and power are cheap, and those places are often dry. In August 2025 the Tucson City Council voted unanimously to reject Project Blue, a proposed data-centre campus linked to Amazon, after months of public opposition centred on water use in a desert city ✓ Established [25]. In Aragón in northern Spain, three proposed Amazon data centres adjacent to existing facilities were licensed to use an estimated 755,720 cubic metres of water a year, enough to irrigate 233 hectares of farmland [26]. In Colón, in the Mexican state of Querétaro, residents of a water-stressed municipality have watched hyperscale campuses arrive while household supplies remain unreliable [27].

Latin America has produced the most consequential confrontations. In February 2024 a Chilean environmental court partially revoked the permit for a Google data centre in Cerrillos, near Santiago, ordering the project to account for the effects of climate change on the local aquifer ✓ Established [28]. In Uruguay, Google’s plans for a campus in Canelones met public protest during the country’s 2023 drought, when the original design was reported to require 7.6 million litres of water a day, equivalent to the domestic consumption of 55,000 people [29]. In both cases the company revised its plans towards air-cooled designs that use far less water but more electricity [29]. That trade-off — water for power — is the central engineering choice of the sector, and it means that reducing one burden usually increases the other.

Water Is Where Consent Breaks

Electricity costs are diffuse and arrive through tariffs most people never read. Water use is local, visible and finite, and in drought years it competes directly with households and farms. Tucson, Cerrillos and Canelones show that the communities most likely to refuse a data centre are those asked to share an aquifer, not a grid. Operators that cannot disclose site-level water use in advance should expect that refusal to spread.

Disclosure is the core weakness. Most operators report water consumption at the corporate level, sometimes by region, rarely by site, and almost never in a form that local planning authorities can verify before approval [22]. Researchers writing for The Conversation noted that companies seldom say exactly how much water an individual facility uses, and that the indirect consumption at power plants is generally absent from corporate reporting [22]. Without site-level numbers, a municipality weighing a data-centre application has no basis for comparing it with the needs of agriculture or housing. The gap is widest in the United States, where most of the growth is taking place and where site-level water reporting is voluntary ◈ Strong Evidence [22].

The industry’s response has been to promise replenishment. Microsoft committed in January 2026 to reduce the water-use intensity of its data centres by 40% by the end of the decade and to replenish more water than it consumes through leak repair and wetland restoration ✓ Established [41]. Google reports similar replenishment programmes [23]. Replenishment is a corporate accounting concept; it offsets consumption in one basin with restoration in another, which is of limited comfort to the specific town whose aquifer is drawn down. Closed-loop and direct-to-chip liquid cooling can cut on-site water use to near zero, and newer AI facilities increasingly use them. But such systems require more electricity for chillers in hot climates, returning the problem to the grid and, through the grid, to thermal power stations and their own cooling water [22] [24].

Water therefore belongs in the same analytical frame as electricity. Both are public or semi-public systems with regulated or administered prices. Both are being asked to absorb a new industrial customer whose demand is large, concentrated and fast-growing. And in both cases the distribution of the burden depends on rules — water rights, permit conditions, disclosure obligations — that were written before hyperscale data centres existed. The difference is speed of feedback. A household may take years to connect a higher electricity bill to a data centre 50 miles away; it notices immediately when a town is asked to choose between a server farm and its wells ◈ Strong Evidence [25] [28].

05

Gas Turbines and Reactor Restarts
The fossil and nuclear revival built for server farms

Louisiana approved three gas plants totalling about 2.3 GW for a single Meta campus; Georgia approved 9,885 MW of new capacity largely for data centres; and Microsoft is paying to restart a reactor at Three Mile Island ✓ Established [35] [37] [30]. The IEA expects gas-fired generation for data centres to grow by 175 TWh by 2030 ✓ Established [1].

For a decade the dominant story in US electricity was the retirement of fossil capacity. Data centres have reversed it in several states. In August 2025 the Louisiana Public Service Commission voted 4–1 to let Entergy build three gas-fired plants, with roughly 2.3 GW of capacity and an estimated cost above $3 billion, to serve Meta’s Hyperion campus in Richland Parish — the company’s largest data centre ✓ Established [35]. The settlement also included new transmission and up to 1.5 GW of solar procurement [35]. The Union of Concerned Scientists criticised the approval as rushed and lacking adequate protections for other ratepayers if Meta’s demand fell short of projections over the plants’ multi-decade life [36]. The plants will be recovered through Entergy Louisiana’s rate base; Meta’s contract terms are what stand between that capital and the utility’s other customers.

Georgia went further. On 19 December 2025 the Georgia Public Service Commission unanimously authorised Georgia Power to procure 9,885 MW of new capacity, including five new gas units at Plant Bowen, Plant McIntosh and Plant Wansley, alongside batteries, solar-plus-storage and purchase agreements ✓ Established [37]. The construction investment alone was put at $16.3 billion [37]. The Southern Environmental Law Center estimated that the plan would cost customers $50 billion to $60 billion over the life of the resources and argued that it lacked sufficient customer protections ◈ Strong Evidence [38]. Georgia Power agreed that revenue from new large-load customers would be used to put downward pressure on rates after its current base-rate freeze expires [37]. Whether that promise holds depends on how much of the forecast load materialises.

The gas revival is constrained by manufacturing as much as by permits. GE Vernova reported 116 GW of gas-turbine backlog and slot reservations at the end of the second quarter of 2026, up from 83 GW at the end of 2025, and is now selling delivery slots for 2031 ✓ Established [39]. Siemens Energy held a 69 GW backlog and Mitsubishi Heavy Industries 35 GW [39]. Roughly a fifth of GE Vernova’s backlog and a quarter of Siemens Energy’s gas-turbine order book is tied directly to data-centre projects, and prices for new orders rose 10 to 20 percentage points per kilowatt in six months [39]. When turbines are scarce, the developers with the deepest pockets secure them first. Utilities building for ordinary customers then pay more, and wait longer, for the same machines.

2019
Singapore moratorium — Singapore pauses new data-centre approvals after the sector’s power and land use outgrows planning assumptions [48].
2022
Dublin connections halted — EirGrid says no new data-centre connections in the Dublin region until 2028 because of grid constraints [44].
Mar 2024
Nuclear campus sale — Talen Energy sells its 960 MW data-centre campus beside the Susquehanna nuclear plant to Amazon Web Services for $650 million [32].
Sep 2024
Three Mile Island deal — Microsoft signs a 20-year agreement with Constellation to restart the 835 MW Unit 1 reactor, renamed the Crane Clean Energy Center [30].
Nov 2024
FERC blocks expansion — FERC rejects an amended agreement that would have raised co-located load at Susquehanna from 300 MW to 480 MW [32].
Dec 2024
Berkeley Lab report — The DOE-commissioned study finds data centres used 4.4% of US electricity in 2023 and could reach 12% by 2028 [4] [5].
Jun 2025
Texas SB 6 — Texas requires large loads to share connection costs and accept curtailment during grid emergencies [21].
Jul 2025
Ohio data-centre tariff — Regulators require AEP Ohio data centres above 25 MW to pay for at least 85% of contracted capacity for up to 12 years [9].
Aug 2025
Louisiana gas approval — Regulators approve three Entergy gas plants, about 2.3 GW, for Meta’s Hyperion campus [35].
Dec 2025
PJM clears short — The capacity auction hits its price cap and falls 6,623 MW below the reliability requirement; FERC orders new co-location rules [11] [33].
2026
Meta nuclear portfolio — Meta signs agreements with Vistra, TerraPower and Oklo for up to 6.6 GW of nuclear capacity by 2035 [34].

Speed has pushed some developers to bypass the grid altogether. In Memphis, xAI installed gas turbines to power its Colossus supercomputer before obtaining air permits; the Southern Environmental Law Center said at least 35 turbines had operated at the site, with a combined capacity of more than 400 MW, and the company later received a permit for 15 ✓ Established [40]. Civil-rights and environmental groups, including the NAACP, challenged the permit, arguing that the plant added nitrogen-oxide pollution to neighbourhoods in South Memphis that already carried a heavy industrial burden [40]. The case shows the other cost of self-supply. Behind-the-meter generation can relieve the grid of the load, but it moves the environmental cost from the bill to the air, and onto the communities that live closest to the site.

Nuclear power is the industry’s preferred answer, and the deals are now large enough to matter. Constellation is investing $1.6 billion to restart Three Mile Island Unit 1 under a 20-year power purchase agreement with Microsoft, adding about 835 MW, and the restart is now expected in 2027, ahead of the original 2028 target ✓ Established [30] [31]. Meta has assembled a portfolio of up to 6.6 GW by 2035: 20-year agreements with Vistra for 2,176 MW from existing plants in Ohio and Pennsylvania plus 433 MW of uprates, funding for TerraPower Natrium reactors of up to 690 MW as early as 2032, and prepayment for a 1.2 GW Oklo campus in Pike County, Ohio [34]. Google’s agreement with Kairos Power targets 500 MW of advanced reactors by 2035, and Amazon is working with X-energy on more than 5 GW by 2039 [34].

The distributional question is whether these deals add supply or merely reallocate it. A restarted reactor is genuinely new capacity for the grid. Contracts for the output of existing plants are not: they move power that was previously sold into the wholesale market to a single buyer, and if the plant is physically co-located with a data centre, the grid loses that capacity at peak while still providing back-up [32] [33]. That was the substance of FERC’s November 2024 rejection of the expanded Talen–Amazon arrangement at Susquehanna, after neighbouring utilities argued it would shift transmission costs to other customers ✓ Established [32]. In December 2025 FERC ordered PJM to create three new transmission services for co-located load, aiming to make such arrangements pay a defined share of the grid they rely on [33].

Small modular reactors are unlikely to change the picture this decade. The earliest TerraPower delivery date in Meta’s portfolio is 2032, the Kairos programme runs to 2035 and the X-energy target to 2039 ◈ Strong Evidence [34]. Until then, the incremental electricity for data centres in the United States will come overwhelmingly from gas and from existing plants, a pattern the IEA’s projections already anticipate: about 175 TWh of additional gas-fired generation to serve data centres by 2030, alongside renewables meeting roughly half of the growth [1]. For households, the significant fact is not the fuel but the contract structure. A gas plant built by a utility for a data centre and placed in the rate base is a 40-year obligation, and it is only as safe for other customers as the data-centre contract that sits beside it [36] [38].

06

Seven Grids, One Problem
How Ireland, the Netherlands, Japan, Singapore, Malaysia and China compare

Ireland’s data centres used 23% of all metered electricity in 2025, up from 5% in 2015 ✓ Established [43]. No other country has gone as far, but the same tension between a mobile industry and a fixed grid appears from Amsterdam to Johor.

Ireland is the case that every other grid operator studies. According to Central Statistics Office data reported by the Irish Times, data centres consumed 7,663 gigawatt-hours in 2025, up 10% from 6,973 GWh in 2024, and their share of metered electricity rose to 23% from 22% ✓ Established [43]. In 2015 the figure was 5% [43]. Total national consumption was just under 33,000 GWh, so data centres now use nearly as much electricity as every home in the country combined [43]. The response came early. In January 2022 EirGrid said there would be no new data-centre connections in the Dublin region until 2028 [44]. By August 2025 RTÉ’s Prime Time was reporting that the government had been warned of rising household bills as data centres strained the grid [45].

The comparison across jurisdictions shows a common sequence — rapid growth, a grid constraint, a pause, then a set of conditions for resumption — at very different scales. The key reference points are:

  • United States: 176 TWh in 2023, 4.4% of national demand, rising to 6.7–12% by 2028; costs allocated through state commissions and regional markets ✓ Established [4].
  • Ireland: 7,663 GWh in 2025, 23% of metered demand; Dublin connections effectively paused since 2022 ✓ Established [43] [44].
  • Netherlands: 5,100 GWh in 2024, 4.6% of national consumption, up 37% since 2021, concentrated around Amsterdam ✓ Established [46].
  • Japan: data-centre peak demand forecast at 6.16 GW by fiscal 2034, in a system whose overall peak grows just 0.4% a year ◈ Strong Evidence [47].
  • Malaysia: data centres projected to need 8,391 MW by 2030, more than half of it in Johor ◈ Strong Evidence [50].

The Netherlands shows the same dynamic at a smaller scale. Statistics Netherlands found that data centres consumed 5,100 GWh in 2024, 4.6% of national electricity, compared with 3.3% in 2021, an increase of 37% in three years ✓ Established [46]. There are about 200 data centres in the country, most of them around Amsterdam, which concentrates the load on a single regional network [46]. Japan, by contrast, is at the start of the curve. The Organization for Cross-regional Coordination of Transmission Operators forecasts data-centre and semiconductor peak demand rising from 560 MW in fiscal 2025 to 7.15 GW in fiscal 2034, of which 6.16 GW is data centres [47]. In a system whose total peak is forecast to grow by only 0.4% a year, that concentrated load accounts for most of the growth, and its location will determine where new transmission is needed ◈ Strong Evidence [47].

Singapore took the most deliberate route. Constrained by land and by its dependence on imported gas, it imposed a moratorium on new data centres in 2019, when the city-state hosted more than 70 facilities with about 1.4 GW of capacity ✓ Established [48]. When it reopened, it did so through allocation rather than first-come connection: the May 2024 Green Data Centre Roadmap committed to at least 300 MW of additional capacity in the near term, with a further 200 MW or more available to operators deploying green energy [48]. The model treats grid capacity as a scarce public asset to be auctioned against efficiency and decarbonisation criteria. It has also pushed demand across the causeway into Malaysia, where the effects are now arriving on the other side of the border.

Malaysia’s experience shows what happens when the overflow lands on a cheaper grid. Projections compiled by industry bodies have Malaysian data centres needing 8,391 MW by 2030, with Johor accounting for more than half ◈ Strong Evidence [50]. From July 2025, the base electricity tariff for Peninsular Malaysia rose from 39.96 sen to 45.62 sen per kilowatt-hour under a new regulatory period, and operators estimated the change would add 10% to 14% to data-centre electricity costs [49]. For an industry drawn to Johor partly by cheap power, that is a material change, and it shows the trade-off: tariffs that protect households also erode the cost advantage that attracted the industry.

The Pause Is the Universal First Response

Singapore in 2019, Dublin in 2022 and Central Ohio’s moratorium before July 2025 all followed the same pattern: growth outran the grid, connections stopped, and resumption came with conditions attached. The conditions — allocation, minimum payments, efficiency standards — are what distinguish a durable policy from a temporary halt. Jurisdictions that skip directly to conditions avoid the cost of the pause.

China is the outlier in both scale and method. The IEA expects Chinese data-centre consumption to rise by about 175 TWh, or 170%, between 2024 and 2030 ✓ Established [1]. Rather than letting developers choose sites, Beijing’s East Data, West Computing programme, launched in 2022, designated eight national computing hubs and ten data-centre clusters, steering computation towards western provinces where land, renewable electricity and cooling are abundant [51]. The approach treats data-centre location as an extension of grid planning rather than as a market decision to be accommodated after the fact. It has its own costs, above all the long-distance transmission needed to link western clusters to eastern users, but it avoids the American pattern of speculative requests clustering in a few congested utility territories ◈ Strong Evidence [51].

Across all seven systems, two variables explain most of the difference in household exposure. The first is whether data-centre growth is large relative to the grid: at 23% of demand Ireland has no room to absorb mistakes, while at 4.6% the Netherlands can still treat the issue as local congestion [43] [46]. The second is whether connection is allocated or first-come. Singapore and China allocate; the United States, until 2025, largely did not [48] [51]. Where capacity is allocated, the costs of growth can be attached to the allocation. Where it is first-come, the costs fall on whoever happens to share the grid, and in practice that has meant households in Virginia, Ohio, Pennsylvania and Dublin ◈ Strong Evidence [15] [45].

07

Rewriting the Tariff
Minimum payments, long contracts and the right to curtail

Between June and December 2025, regulators in Texas, Ohio, Virginia and Georgia and the federal regulator in Washington all changed the terms on which data centres connect ✓ Established [21] [9] [20] [37] [33]. The common principle is that the customer who causes the cost should carry the risk.

The regulatory response has converged on a small set of instruments. The first is the minimum-take tariff. Ohio’s, approved in July 2025, requires AEP Ohio data-centre customers above 25 MW to pay for at least 85% of their subscribed capacity for up to 12 years ✓ Established [9]. Virginia’s State Corporation Commission went further in November 2025, creating a new GS-5 rate class for customers of 25 MW or more, effective January 2027, with 14-year contracts and minimum charges of 85% of contracted transmission and distribution demand and 60% of generation demand [20]. These terms convert a connection request from an option into an obligation. A developer who files in four states under such tariffs must be prepared to pay in all four, which is precisely the discipline that removed more than 24 GW from AEP Ohio’s pipeline [10].

The same Virginia order shows the limits of tariff design as household protection. The commission approved a Dominion Energy rate increase that adds $11.24 a month to the typical residential bill in 2026 and a further $2.36 in 2027, a total of $13.60, or about 9%, on an average bill of $149.92 ✓ Established [20]. Dominion had asked for $19.57 [20]. The commission cited inflation and policy costs that affect all utilities, and the GS-5 class does not take effect until 2027 [20]. In other words, the rules that assign more cost to data centres arrived after much of the growth they are meant to price, and households in Virginia are paying for the transition. Tariff reform protects against future cost-shifting; it does little for costs already in the rate base.

◈ Strong Evidence Minimum-payment tariffs sharply reduce speculative data-centre demand

After Ohio required data centres above 25 MW to pay for at least 85% of contracted capacity for up to 12 years, AEP Ohio reported its pipeline fell from more than 30 GW to 5,642 MW of signed, financially committed agreements [9] [10]. Virginia’s GS-5 class applies similar logic with 14-year contracts from 2027 [20]. The evidence suggests that pricing the request is the most effective single filter on phantom load.

Texas chose operational controls rather than price alone. Senate Bill 6, signed on 20 June 2025, requires large loads above 75 MW to contribute to the cost of their interconnection studies and facilities, and requires new large loads to install equipment allowing ERCOT to curtail them during firm load-shed events and declared emergencies ✓ Established [21]. The provision matters because it addresses reliability as well as cost: a curtailable data centre can be connected to a grid that could not otherwise serve it at peak. Research from Duke University’s Nicholas Institute found that the US grid could integrate 76 GW of new load — about 10% of aggregate peak demand — if those loads accepted curtailment for an average of 0.25% of their annual operating hours [52]. Flexibility, in that framing, is a substitute for new capacity, and therefore for new costs on other customers.

At federal level, FERC’s December 2025 order on co-location found PJM’s existing rules unjust and unreasonable and directed the creation of firm, non-firm and interim transmission services for loads located at generating plants ✓ Established [33]. The aim is to replace the ad hoc treatment that produced the Talen–Amazon dispute with defined rates for the grid services a co-located data centre still uses [32] [33]. PJM, separately, has developed a backstop procurement of about 6 GW aimed at the reliability gap created by data-centre demand, and has tightened the evidence required before large loads enter its forecast [53]. The market monitor’s preferred solution is blunter: that data centres should bring their own generation rather than draw on a system that, in Bowring’s words, has simply no new capacity to meet new loads [14] [54].

Especially when tech companies are so profitable, it’s both unfair and politically unrealistic for our industry to ask the public to shoulder added electricity costs for AI.

— Brad Smith, Vice Chair and President, Microsoft, January 2026 [41]

The industry has begun to make similar commitments itself. In January 2026 Microsoft announced a set of pledges under which it will ask utilities and regulators to set rates high enough for the company to cover 100% of the electricity costs of its data centres, including grid upgrades, so that residential tariffs do not rise because of its facilities ✓ Established [41]. A voluntary pledge is not a tariff: it binds one company, depends on regulators accepting the proposed rates, and cannot be enforced by the households it is meant to protect. But it concedes the central empirical point of this report — that, absent intervention, costs do shift — and it raises the political price for competitors who decline to match it. The risk assessment below grades the main approaches against the cost-shifting they are meant to prevent:

ApproachResidual risk to householdsAssessment
Minimum-take tariffs (Ohio, Virginia)
High
Effective against future speculative load, as AEP Ohio’s 30 GW to 5.6 GW pipeline shows, but does not touch costs already in the rate base [10] [20].
Utility-built gas plants for single customers
High
A 30 to 40 year asset protected only by a shorter contract; Louisiana and Georgia approvals drew criticism for weak ratepayer safeguards [36] [38].
Curtailment and flexibility rules (Texas)
Medium
Reduces the need for new peak capacity, with Duke estimating 76 GW of headroom at 0.25% curtailment, but depends on enforcement in emergencies [21] [52].
Co-location and own generation
Medium
Removes load from the grid but can shift transmission costs and, with on-site gas, local pollution, as Susquehanna and Memphis show [32] [40].
Socialised connection costs (status quo)
Critical
More than 95% of 2024 PJM connection projects charged to all customers; the default that most reforms are trying to end [15].

The table points to a hierarchy of protections. Rules that price the request — minimum-take tariffs and collateral — are the most effective against phantom load and the cheapest to administer ◈ Strong Evidence [9] [10]. Rules that price the connection — assigning substations and dedicated lines to the customer that requires them — address the $4.3 billion problem that UCS documented [15]. Rules that require flexibility reduce how much new capacity anyone needs to build [52]. Utility-built plants dedicated to single customers carry the highest residual risk, because the asset outlives the contract and the regulator must forecast the data-centre industry’s fortunes for four decades [36] [38]. The most robust jurisdictions will combine all three priced protections and avoid the fourth unless contract terms match the asset life.

What none of these reforms yet provides is transparency comparable to the scale of the money involved. Special contracts remain confidential in many states [16]. Utilities report data-centre load to regulators but rarely in a form that allows outsiders to check forecasts against signed agreements, which is why AEP Ohio’s disclosure of its pipeline before and after the tariff is so unusual and so valuable [10]. Water use is reported at the corporate level, not by site [22]. A household asked to pay for the AI build-out has a reasonable claim to know how much, and on what terms its utility has agreed to serve the companies driving it. The jurisdictions that have published that information — PJM through its market monitor, Ireland through its statistics office — are the ones where the debate is most grounded in evidence ◈ Strong Evidence [13] [43].

08

Do Data Centres Raise Bills
The genuine disagreement, and what the evidence supports

A Berkeley Lab study of 2019–2024 found that states with the fastest load growth saw real retail prices fall, because fixed costs were spread over more sales ⚖ Contested [18]. The PJM record shows the opposite pressure in a capacity-constrained market ✓ Established [13]. Both findings are correct, and the difference between them is the key to the policy question.

The strongest argument that data centres lower bills comes from Lawrence Berkeley National Laboratory. Its analysis of state retail prices between 2019 and 2024, published in November 2025, found that states with the highest load growth experienced reductions in inflation-adjusted average prices, while states with shrinking demand generally saw prices rise ⚖ Contested [18]. The logic is simple: much of a utility’s cost is fixed, and spreading those costs over more kilowatt-hours lowers the price of each one. In the authors’ words, spreading fixed costs over more demand naturally exerts downward pressure on retail prices [18]. The study found that the main driver of price increases was spending on distribution and transmission, much of it for refurbishment and replacement rather than for new load [18].

The same study contains two caveats that the debate often omits. The negative relationship between load growth and prices was smaller and lost statistical significance when residential prices were analysed on their own; most of the benefit accrued to commercial customers [18]. And the authors stated explicitly that whether these historical relationships extend to a future of significant, nationwide load growth is unclear ◈ Strong Evidence [18]. The 2019–2024 period was one in which most new load connected to grids with spare capacity. Fixed-cost spreading works when the system already exists. It does not work when the new load requires new plants, new lines and capacity bought at a capped auction price — which is the situation PJM has been in since 2024 ✓ Established [11] [13].

The disagreement is therefore not about facts but about regimes. In a system with surplus capacity and cost-reflective connection charges, a large new customer that pays its own way lowers average costs for everyone else, as the Berkeley Lab evidence suggests ⚖ Contested [18]. In a system that is short of capacity, prices scarcity through a capped auction and socialises connection costs, the same customer raises costs for everyone else, as the PJM evidence shows ✓ Established [13] [15]. The United States in 2026 contains both regimes. Much of the current political argument consists of each side citing the evidence from the regime that suits it. The more useful question is which regime a given utility territory is in, and what rules would move it from the second to the first.

The Case That Data Centres Lower Bills

Fixed-cost spreading
States with the fastest load growth saw real average prices fall between 2019 and 2024 [18].
Contracted revenue
Minimum-take tariffs and 14-year contracts turn data-centre load into stable revenue that can support the wider system [9] [20].
Flexibility headroom
Duke estimates 76 GW of new load could connect with minimal new capacity if it accepts brief curtailment [52].
Private capital for supply
Microsoft’s Crane restart and Meta’s nuclear portfolio bring privately financed generation onto the grid [30] [34].
Industry commitments
Microsoft has pledged to cover 100% of its data-centre electricity costs, including grid upgrades [41].

The Case That Data Centres Raise Bills

Capacity prices
Data-centre load added $29.4 billion to PJM capacity revenue over four auctions [13].
Socialised connections
More than 95% of 2024 PJM data-centre connection costs, $4.3 billion, went to all customers [15].
Residential exclusion
The load-growth price benefit was not statistically significant for residential customers [18].
Stranded-asset risk
Gas plants built for single campuses outlive their contracts, leaving long-term cost risk with ratepayers [36] [38].
Phantom forecasts
Utility forecasts may overstate data-centre demand by about 25 GW, and customers pay for forecast load [6].

Several other claims in the debate remain contested on the evidence. Whether utility forecasts are systematically too high is disputed: Grid Strategies estimates an overstatement of about 25 GW, while the ERCOT queue and the IEA’s 2026 upgrade both point to demand still accelerating ⚖ Contested [6] [3] [7]. Whether gas plants are a necessary bridge is contested: the IEA’s projections include 175 TWh of additional gas generation for data centres, while UCS and SELC argue that Louisiana and Georgia approved fossil capacity without adequate protection against demand falling short ⚖ Contested [1] [36] [38]. Whether flexibility can substitute for new capacity is contested too: the Duke analysis shows large theoretical headroom, but data-centre operators have historically treated uninterrupted uptime as their core product, and few have yet accepted curtailment at scale ⚖ Contested [52] [21].

What can be said with confidence is narrower but more useful. Four findings hold across the sources examined in this report:

  • The load is real. Data-centre electricity use grew 17% in 2025 and is on course to roughly double by 2030 ✓ Established [3] [1].
  • The forecasts run ahead of it. Where requests were priced, as in Ohio, more than four-fifths of the pipeline disappeared ✓ Established [10].
  • The costs have shifted. In PJM, capacity and connection costs driven by data centres have been spread across all customers ✓ Established [13] [15].
  • The rules are changing. Texas, Ohio, Virginia, Georgia and FERC all rewrote connection terms in 2025, but after much of the cost was incurred ✓ Established [21] [9] [20] [33].

The structural insight is that the price of AI’s electricity is being set less by engineers than by rate designers. The same 1 GW campus can lower or raise a household’s bill depending on whether its connection is priced, whether its forecast is backed by collateral, whether it can be curtailed at peak, and whether the plant built to serve it is matched by a contract of equal length ◈ Strong Evidence [9] [18] [52]. These are administrative choices made by state commissions and grid operators, most of them in proceedings that few households know exist. The technology companies have already conceded the principle that they should pay their own way [41]. The remaining task is to write that principle into tariffs that do not depend on a press release.

Allocation, Not Technology, Decides the Bill

Data centres are neither inherently good nor inherently bad for electricity customers. They are very large, very fast-growing loads whose effect on bills depends on rules about who pays for capacity, connection and risk. Where those rules price the request and the connection, households can benefit from spreading fixed costs. Where they do not, households underwrite the AI build-out. The evidence from 2025 is that most of the US grid was in the second category, and is only now moving towards the first.

The data-centre power grab is therefore less a seizure than a default. Rules written for a century of diffuse, slow demand growth assumed that new load would pay for itself through additional sales, and that socialising the cost of the grid was fair because everyone used it in roughly similar ways ◈ Strong Evidence [16] [18]. A single industry asking for tens of gigawatts in a few years breaks both assumptions. The jurisdictions that have adapted fastest — Singapore with allocation, Ohio with minimum payments, Texas with curtailment — have not stopped data centres from being built [48] [10] [21]. They have changed who carries the risk if the forecasts prove wrong. That shift, more than any new reactor or turbine order, is what will determine whether households pay for AI’s electricity or merely share a grid with it.

SRC

Primary Sources

All factual claims in this report are sourced to specific, verifiable publications. Projections are clearly distinguished from empirical findings.

Cite This Report

APA
OsakaWire Intelligence. (2026, November 15). The Data-Centre Power Grab: $29.4B and Who Pays (2026). Retrieved from https://osakawire.com/en/the-data-centre-power-grab-who-pays-for-ais-electricity/
CHICAGO
OsakaWire Intelligence. "The Data-Centre Power Grab: $29.4B and Who Pays (2026)." OsakaWire. November 15, 2026. https://osakawire.com/en/the-data-centre-power-grab-who-pays-for-ais-electricity/
PLAIN
"The Data-Centre Power Grab: $29.4B and Who Pays (2026)" — OsakaWire Intelligence, 15 November 2026. osakawire.com/en/the-data-centre-power-grab-who-pays-for-ais-electricity/

Embed This Report

<blockquote class="ow-embed" cite="https://osakawire.com/en/the-data-centre-power-grab-who-pays-for-ais-electricity/" data-lang="en">
  <p>Data-centre demand added $29.4 billion to capacity costs on the largest US grid in four auctions. Analysis covers tucson, naacp, facility. Evidence-sourced,...</p>
  <footer>— <cite><a href="https://osakawire.com/en/the-data-centre-power-grab-who-pays-for-ais-electricity/">OsakaWire Intelligence · The Data-Centre Power Grab: $29.4B and Who Pays (2026)</a></cite></footer>
</blockquote>
<script async src="https://osakawire.com/embed.js"></script>