Data and research
Every figure that gets thrown around in this debate, with the document it came from, what it actually measures, and how it has moved. The daily update checks for new official releases and records changes below.
About 4.7% of all US electricity
Lawrence Berkeley National Laboratory (source 1), as of June 2026
The official federal count, published June 2026. Both sides accept it. Excludes cryptocurrency mining. The next measured year (2025) will not be published until 2027.
The lab's Reference Case has use rising 22% from 2024 to 2025. Applied to 192 TWh that is roughly 234 TWh, or about 5.6% of US electricity
Lawrence Berkeley National Laboratory, 2025 Update (source 1), as of June 2026
An estimate, not a measurement. The report states the growth rate in its conclusion and shows the yearly values only in a chart, so the 234 figure is derived. Its own stress test puts the range at roughly 20% below to 30% above.
Up 25.5% from 249.0 TWh in 2024. About 39.7% of a world total of 787.8 TWh; China 205.7 TWh, Europe 144.6 TWh
Energy Institute Statistical Review of World Energy 2026 (as reported by Forbes, August 23, 2026) (source 101), as of July 2026
A different scorekeeper with a broader count: it puts 2024 at 249 TWh where Berkeley Lab puts it at 192 TWh. Both are defensible; they define and model data centers differently. Never mix the two series in one sentence.
Up more than 15% from about 415 TWh in 2024, roughly 1.5% of world electricity. AI-focused data centers grew about 50%. The IEA puts the US at about half of world capacity
IEA, Key Questions on Energy and AI (2026) (source 100), as of 2026
The IEA's world total (485 TWh) is far below the Energy Institute's (787.8 TWh) for the same year. Same lesson: check which series a number comes from before comparing.
Reference Case growth of 29% from 2025 to 2026, applied to the 234 TWh 2025 estimate. Roughly 7% of US electricity
Lawrence Berkeley National Laboratory, 2025 Update (source 1), as of June 2026
A forecast. Nobody will have a measured 2026 national figure until 2028.
EIA forecast growth in commercial sector electricity sales for 2026, driven by data centers and manufacturing; 2.7% forecast for 2027 (Short-Term Energy Outlook, Table 7a)
EIA September 2026 Short-Term Energy Outlook (source 102), as of September 9, 2026
Commercial sector includes offices, retail and other non-industrial business uses, not just data centers, so this is an upper bound for data center-driven growth.
The report's four sensitivity scenarios span 578 to 782 TWh, which is 10.5% to 14.2% of US electricity. Its wider compounded uncertainty range is 521 to 843 TWh, or 9.5% to 15.3%
Lawrence Berkeley National Laboratory (source 1), as of June 2026
The authors publish ranges on purpose. 649 TWh is the Reference Case, not the midpoint of anything, and the two ranges are different things: do not pair the narrow band of TWh with the wide band of percentages, which is an easy mistake because both appear in the same report. Do not quote any single figure as a prediction.
From 2017 to 2023, four times faster than global electricity growth overall
International Telecommunication Union and World Benchmarking Alliance (source 83), as of June 2025
From the United Nations' own digital agency, using figures the companies report themselves. This is a growth rate, not a total. It is the clearest single answer to why this argument arrived so suddenly: the curve is bending up much faster than electricity demand in general.
About 2.1% of world electricity, and ten companies account for half of it
International Telecommunication Union and World Benchmarking Alliance (source 83), as of June 2025
Read the label carefully: this counts these companies' entire operations, not their data centers alone, so it is not a data center total and should never be quoted as one. It is useful for a different point, which is how few companies this power use sits with. The same report found 23 companies ran on 100% renewable electricity in 2023, up from 16 the year before.
A projection, starting from an estimated 448 TWh in 2025
United Nations University Institute for Water, Environment and Health (source 84), as of June 2026
A projection, not a measurement. Note the starting point before you use it: this report puts 2025 at 448 TWh, while the IEA figure on this page puts the same year at 485 TWh. Neither is wrong; they are different series. Check which one a number came from before setting two of them side by side.
Up from $28.92 before the spike, then $269.92, then $329.17, then the price ceiling
Utility Dive (source 18), as of December 2025
Wholesale standby price, not a household bill. Standby is roughly a tenth to a quarter of a home bill.
Of $16.4 billion in charges, per the independent market monitor
Monitoring Analytics via Utility Dive (source 17), as of 2026
Measures who caused the cost, not how the bill was split. Both sides accept the number.
All but 2 GW attributed to data centers, in PJM's own load forecast
Utility Dive (source 104), as of December 2025
The state board's own figure, a 17 to 20% jump in the supply portion, phased in from June 2025
New Jersey Board of Public Utilities (source 27), as of February 2025
New Jersey is the clearest case because its board published dollar figures; most states did not. PJM reports the capacity cost as a percentage of the bill, 1.5 to 5% in some states, and does not publish a dollar-per-household number, so do not expect the two to line up.
30 of them created in 2025 and 2026
E3 (commissioned by the Data Center Coalition) (source 8), as of May 2026
Industry-funded study. Count is verifiable in state filings.
A new rate class for the largest customers: 14-year contracts and minimum charges covering the capacity they reserve, whether they use it or not
Virginia State Corporation Commission (source 9), as of February 2026
Read the 60% figure carefully if you meet it elsewhere: it is the share of those minimum charges a customer must guarantee with collateral, and only if its credit is not strong enough on its own. It is a credit requirement, not a block of generation set aside for data centers.
Against 41 cents for the average Virginia incentive. The same table gives 84 jobs against 58, and $10 million of state GDP against $9 million, per $1 million spent
JLARC (Virginia legislature's nonpartisan auditor) (source 22), as of December 2024
JLARC rates the exemption 'moderate' and says it is similar to the average incentive, so it does beat the average but not dramatically. It still does not pay for itself on the state's own books: 48 cents back per dollar given up is a loss, just a smaller one than most.
Ranked the state's second-best incentive
Cardinal News (source 28), as of June 2026
Measures wages generated, largely during construction, not money back to the treasury.
Up from $555 million in 2024; that works out to roughly $330 per Ohio household
Good Jobs First (source 20), as of 2026
The per household figure is this site's own arithmetic, the state total divided by Ohio households, not a number Good Jobs First published. It is a way to feel the size of the total, not a bill anyone receives.
Fourteen states with exemptions publish no annual total at all
Good Jobs First (source 20), as of November 2025
Counts money given away; does not fully count money later collected.
About 39% of the county budget; tax rate cut every year from $1.145 (2016) to $0.805 (2026) per $100
Loudoun County, Virginia (source 4), as of February 2025
Public budget documents. The most favorable case in the country.
Computing employment up 56%, telecom up 43%, wages flat, home values up 2 to 5% over a decade
Brookings (Bahar and Wright) (source 6), as of 2026
The one jobs study both sides cite.
Each direct job modeled to support 4.5 more
PwC for the Data Center Coalition (source 31), as of 2025
Modeled, not measured; industry-funded.
At least half of them from data center work, according to the local
Boston Globe (source 7), as of May 2026
Reported by the local union rather than from an audited filing. The often repeated additions, that this is double the hours of a decade earlier and that the local took on 600 new apprentices in one year, are not in the article cited here, so they are left out until they can be sourced.
$5.5 billion in wages, $9.1 billion to the state economy, mostly during initial construction
JLARC (source 22), as of December 2024
Both sides accept this line and the next: costs 'will likely increase' for all customers.
The GPT-3 estimate: about 700,000 liters evaporated on site, and the rest consumed off site generating the electricity; global AI water use projected at 4.2 to 6.6 billion cubic meters by 2027
UC Riverside (Li, Yang, Islam, Ren) (source 23), as of 2025
The 700,000 litre figure travels much further than the full one, and it is the on-site share alone. Quote whichever you mean, but say which: they differ by roughly eight times. Both are modelled estimates for one 2022-era model, not a measurement, and newer chips change the arithmetic.
Against its earliest facility designs, by the company's own accounting
GeekWire (source 14), as of 2026
Self reported by Microsoft and not independently audited, and it compares new designs against the company's oldest ones rather than against the industry. Claims that the newest designs use no water at all in normal operation, and that Google has committed to the same by 2030, are widely repeated but are not in this source; treat them as unverified here.
The report compares it to the basic yearly household water needs of 1.3 billion people
United Nations University Institute for Water, Environment and Health (source 84), as of June 2026
Projected, not measured, and global: it tells you nothing about any single site, which is the level where most water fights actually happen. The comparison to household needs across Sub-Saharan Africa is the report's own framing, and the institute's specialty is water.
$18 billion blocked outright and $46 billion delayed, counted over about two years
Data Center Watch (source 25), as of 2025
Not a calendar-2025 total: the tracker's own window runs from 2024 into early 2026. Data Center Watch is run by 10a Labs, an AI security firm whose paying clients include AI companies. 10a Labs says the tracker itself is not client funded and is kept separate from that work, which is its position, not an audited fact.
About 75 projects; opposition groups more than doubled across 49 states; 14 legislatures took up pause bills
Data Center Watch (source 25), as of April 2026
Price tag of investment that did not land, not money anyone saved.
Data Center Watch counted 45 US data center projects blocked or delayed by local opposition between April and June 2026, with regulatory scrutiny in 49 states
Data Center Watch Q2 2026 report (source 103), as of September 21, 2026
Combines blocked and delayed projects, and rests on public records and press tracking rather than a regulator database. The project-level detail sits behind a paid report, so the free version cannot be checked line by line, and Data Center Watch does not fully disclose its funders.
Heatmap Pro says data center investments totaling about $130 billion were canceled in the three months ending September 30, 2026, with roughly $260 billion canceled year to date and about $1 trillion facing meaningful opposition.
Heatmap News (Heatmap Pro), as of October 7, 2026
Measures announced dollar values of projects canceled or paused, not built capacity lost; based on Heatmap's proprietary tracker.
8% like such officials; 75% of Democrats and 44% of Republicans dislike them, survey of US adults September 28 to October 4, 2026
Pew Research Center, as of October 8, 2026
Pew's question asks about a hypothetical local data center, not a specific project; it captures political sentiment, not a vote tally.
Record quarter; full-year investment forecast near $700 billion
Industrial Info Resources (source 29), as of 2026
Commercial market data with no side in the debate.
Across 84 tracked US facilities and 7 Chinese ones, as counted in 2026
AI Data Center Index (source 30), as of September 2026
Read this as a floor, not a world share. The tracker counts facilities it can identify from public announcements, and it sees far less of China than of the US, so part of the gap it shows is a gap in visibility rather than in concrete. It is evidence that the US is building more; it is not a measurement of who holds what share of world AI compute.
Across 153 tracked projects; $14.4 billion added in August alone
ConstructConnect October 2026 Data Center Report, as of October 1, 2026
Counts tracked starts and announcements, not completed facilities, so the figure can include projects that are later canceled or delayed.
Across regions serving about 95% of the country, if customers accept short, occasional cutbacks
Duke University Nicholas Institute (source 11), as of February 2025
Depends on flexibility that most contracts do not yet require.
US and Europe; the launch named Compass Datacenters, Constellation, Emerald AI, Google, National Grid, Nebius, NVIDIA, Oracle and PADO AI
EPRI (source 12), as of 2026
Participants include the companies being tested, and the roster changes as the initiative adds members.
42 percent of forecast data center demand growth between the 2026 to 2027 and 2028 to 2029 delivery years, as of September 21, 2026
PJM Independent Market Monitor via Forbes, as of September 21, 2026
Measures announced data center load inside PJM that has been formally canceled or delayed, not total pipeline. The underlying forecast still rises overall, so the figure describes shrinkage of expected growth, not a shrinking grid load in absolute terms.
What the daily research update added or changed, newest first.
Searches every argument, question, number, source, report, update and correction on this site. Your query stays in your browser.
Before you download
The handbook is free and always will be. It takes no money from data center operators, utilities or industry groups, which is exactly why you can trust what is in it. If it is worth a dollar, now is the easy moment.
No thanks, just download itOne tap through Stripe. Your download starts either way.
Anything you wish this site had, or anything that was missing or wrong? One line is plenty, it comes straight to a person, and it is what the next version gets built from.
Nobody pays for this but readers. Chip in $1 if it was worth it.