Distribution investment
Growth in nominal distribution capital spending, 2019–2023. More than twice inflation. [1]
A SCIENCE FACT DOUBLE FEATURE
ATTACKOF THEDATACENTERS!
They told you AI would bankrupt your electric bill and drink the reservoirs dry. The national labs and grid researchers checked the math. The doomsday script leaves out the science.
Investigate the evidenceBig machines. Bigger headlines. A field guide for curious humans.
THE MONSTER IS FICTION.
THE RESEARCH IS REAL.
NOW SHOWINGTHE RESEARCH IS REAL.
THE FACTS
IN FULL CONTEXT
LESS PANIC ✦ BETTER ENGINEERING ✦ FOLLOW THE EVIDENCE ✦ READ THE FINE PRINT ✦ LESS PANIC ✦ BETTER ENGINEERING ✦
BEFORE YOU RUN FOR THE HILLS
Panic makes great B-movies.
Terrible energy policy.
The monster-movie version says bigger computing means inevitable environmental disaster. The evidence tells a much more interesting story: new demand can spread grid costs, flexible computing can support the grid, and modern cooling can slash water consumption.
Look past viral headlines at the research from Berkeley Lab, Duke, MIT and Nature. The blanket catastrophe narrative leaves out measurable engineering gains. Here are four encounters with the facts—and the practical choices that put those gains to work.
01The phantom
THE CASE OF THE RISING ELECTRIC BILL
The phantom
of the power bill!
PULP HYSTERIA! · ISSUE 01
SHOCK!HORROR!
THE LOAD
THAT ATE
YOUR
WALLET!
“New demand can only make everyone’s electricity more expensive!”
A FAMILIAR STORY. AN INCOMPLETE SUSPECT LIST.LAB REPORT 01 · CHECK THE HISTORICAL RECORD
MYTH VAPORIZEDMore load can mean
less cost per unit.
31states
had lower inflation-adjusted average retail electricity prices in 2024 than in 2019, in Berkeley Lab’s 48-state analysis. [1]
Growth spreads fixed grid costs over more electricity sales. In North Dakota, a mix of industrial growth—including data centers—coincided with a nearly 3¢/kWh decline in real average rates. Abundant supply helped make it possible. [1] [7]
THE ACTUAL SUSPECT BOARD · FOLLOW THE MONEY
California wildfire costs
Wildfire-related costs placed in California utility rates, 2019–2023. [7]
Natural gas prices
October 2025 versus a year earlier, as reported in WRI’s December 2025 explainer. [7]
Different measures, dates and regions; these figures are not additive shares of a national electricity bill.
ATOMIC GRID-O-METER · FIXED-COST LAB
More sales.
Same poles and wires.
Start with a fixed $100 grid-cost pool and 100 units sold. Turn the dial to see what happens when sales grow.
An arithmetic illustration, not a prediction of a household bill.
0%+25%+50%+100%
Fixed cost per unit$0.8020% lower than the $1.00 starting point
Illustration excludes energy, new construction, financing and tariff design.
02From grid invader
THE LOAD THAT LEARNED TO BEND
From grid invader
to grid stabilizer.
PULP HYSTERIA! · ISSUE 02
NOESCAPE!
IT CAME
FOR THE
POWER
GRID!
“A data center is a giant load that can never get out of the way.”
BUT WHAT IF THE LOAD… COULD BEND?LAB REPORT 02 · DUKE NICHOLAS INSTITUTE
PLOT TWIST!Timing is an
engineering resource.
Computing that can wait or move—and facilities that coordinate storage—can reduce grid demand at its most expensive moments. Duke’s February 2026 study puts dollars on that flexibility. [8]
$40–150billion
Estimated cumulative system savings through 2035 in most modeled scenarios, from avoided capital, operating and fuel costs. [8]
10–50%less investment in new
combined-cycle gas capacity
combined-cycle gas capacity
0.5–2.8%lower modeled retail prices
for other customers
for other customers
5–10 yearsutility planning horizon
examined in the study
examined in the study
Grid services, too. Duke models flexibility as a source of reserves as well as peak relief. Qualified, rapid demand reductions can provide reserve support rather than keeping additional generation spinning. Data centers’ own modeled average electricity prices fall by $2.5–$8/MWh. [8]
MISSION CONTROL · GRID STABILIZER MODE
Don’t feed the peak.
SYSTEM SAVINGS POTENTIALSTANDBY
OTHER-CUSTOMER RETAIL PRICESSTANDBY
Animation: illustrative daily load, equal computing energy. Readouts: Duke’s 2026 scenario results through 2035, not outputs calculated by this animation. Retail reductions are modeled outcomes, not individual tariff guarantees. [8]
THE ORIGINAL DISCOVERY · DUKE’S 2025 FIRST-ORDER MODEL
A little give. A lot of headroom.
Annual energy curtailedModeled new-load headroom →
22 major balancing authorities; flexible new loads generally. These percentages describe annual energy curtailed, not simply hours offline. Local network constraints require separate study. [2]
03The scale
THE CREATURE FROM THE COOLING TOWER
The scale
of the splash!
PULP HYSTERIA! · ISSUE 03
THIRSTFROM SPACE!
IT DRANK
THE WHOLE
LAKE!
“Every chatbot query needs a bottle of water. Forever.”
ONE NUMBER CANNOT DESCRIBE EVERY MACHINE.LAB REPORT 03 · THE QUERY NUMBER IS NOT A CONSTANT
MYTH VAPORIZEDModels change.
So does the water math.
In Knowable, UC Riverside researcher Shaolei Ren says his 2024 estimate—500 mL for drafting a short email with GPT-4—is already outdated. [5]
About five drops
Google’s 2025 estimate for the median Gemini Apps text prompt: 0.26 mL. A company-reported, model-specific measurement—not a universal AI average. [16]
THE GREAT AMERICAN WATER TANK
Big numbers need a scale.
National annual volumes in billions of U.S. gallons. Select a subject to inspect the evidence.
Data centers17.4B
Direct consumption · 2023 · Berkeley Lab
Golf irrigation~531B
Water applied · 2024 · GCSAA
Residential outdoors~2,920B
Outdoor use, mainly landscaping · EPA daily estimate × 365
NATIONAL VOLUME · DIRECT CONSUMPTION17.4 billion gal.
Berkeley Lab’s estimate for on-site data-center water consumption in 2023. Cooling is a far smaller national volume than familiar outdoor water uses.
Read the source [4] ↗Read the scale correctly: the bars share a linear scale from zero, but compare different measures and years: consumption, irrigation applied and outdoor water use. They are context, not equivalent water footprints. [4] [10] [11]
The national story is smaller
than the monster headline.
ITIF estimates that direct cooling plus electricity-related consumption amounts to less than 1% of U.S. consumption, using Berkeley Lab and USGS figures. This is an approximate comparison across assessments. The electricity-related component—about 211 billion gallons in Berkeley Lab’s 2023 estimate—is additional to the 17.4B direct figure. [9] [4]
04Revenge of
SCIENCE TO THE RESCUE
Revenge of
the engineers!
PULP HYSTERIA! · ISSUE 04
SCIENCESTRIKES BACK!
THE THIRST
THAT COULD
NOT BE
TAMED!
“You can’t cool powerful computers without endlessly evaporating fresh water.”
THE ENGINEERS WOULD LIKE A WORD.LAB REPORT 04 · CHANGE THE COOLING DESIGN
DOOMSDAY CANCELLED*A liquid loop can
use its water again.
Direct-to-chip cooling circulates coolant through a sealed loop. Pair it with dry heat rejection and heat moves into the air without evaporating the cooling water. Microsoft has announced new designs that use this approach. [13]
113°F coolant
NVIDIA describes 45°C liquid loops that make dry cooling practical for more operating conditions. Warmer coolant widens the gap between air and fluid, helping fans carry heat away. [13]
*The supposed engineering impossibility is cancelled. Performance depends on the chosen equipment and climate.
RECIRCULATE
Keep the coolant.
Closed primary loops reuse fluid. Dry or suitably designed hybrid heat rejection cuts evaporative losses; the full cooling system determines the result. [5] [13]
RECLAIM
Save drinking water.
In 2023, just over one-third of water used by Virginia data centers was reclaimed wastewater. That is a statewide finding, not a Northern Virginia-only percentage. [12]
RECOVER
Reuse the blowdown.
Reverse osmosis and zero-liquid-discharge systems recover water from wastewater streams so it can be reused. They complement cooling design; they do not capture water already evaporated. [9]
PEER-REVIEWED POSSIBILITY · NATURE, 2025
Design isn’t a footnote.
It’s a lever.
Advanced cooling versus an air-cooled reference, in a life-cycle comparison. [6]
15–20%less energy demand
31–52%less blue-water consumption
15–21%less greenhouse gas emissions
Modeled case comparisons, not fleet-wide measurements. Blue water means surface- and groundwater consumption. Authors were Microsoft or WSP employees. [6]
THE REAL VILLAIN · AVOIDABLE BAD DECISIONS
Don’t blame the technology
for bad design or bad actors.
With data centers—just like housing, hospitals, factories or highways—poor siting, sloppy engineering and unfair cost allocation can produce bad results. A thirsty cooling design in a stressed watershed, or a contract that dumps new infrastructure costs on homeowners, deserves scrutiny. Those are choices we can demand be made better.
Modern coolant loops, water reclamation and grid-interactive operation give us the tools to conserve freshwater, flatten peaks and lower system costs. The evidence supports building better, rather than treating every server building as the same monster.
Grid reality. Supply constraints can raise prices; PJM has experienced real capacity-cost pressure. Duke’s headroom estimates do not replace a local network study. [1] [2] [7]
Carbon reality. MIT finds that shifting to cheap hours can increase emissions on fossil-heavy grids. Duke models lower long-run emissions as investment shifts, including fewer new gas plants; these are different models, not a refutation of MIT. [3] [8]
Water reality. National scale does not settle local summer demand. The full cooling chain, reclaimed-water source, power-generation mix and watershed all matter. A sealed chip loop alone does not describe facility-wide consumption. [4] [5]
THE EARTHLING’S FIELD CHECKLIST
Good engineering.
Put it in the permit.
The science gives us a playbook. Here are five practical questions that turn promising engineering into accountable projects.
0 / 5 questions ready for mission control
An editorial checklist for evaluating proposals, not a project certification or prediction.
NO SECRET ALIEN KNOWLEDGE REQUIRED
Roll the
research credits.
Start with Knowable for water, WRI for bills and Duke’s 2026 report for the economics of flexibility. Follow the source links to inspect the evidence. Reviewed October 17, 2026.
BERKELEY LAB + BRATTLE · JOURNAL ARTICLE & BRIEFING · 2025
Factors Influencing Recent Trends in Retail Electricity Prices in the United States ↗
The foundation for the bill explainer. Historical state-level relationships have multiple drivers and do not guarantee future results.
DUKE NICHOLAS INSTITUTE · RESEARCH REPORT · FEBRUARY 2025
Rethinking Load Growth ↗
The source of the 76 / 98 / 126 GW chart. A first-order analysis of flexible new loads, with important network and implementation limits.
MIT RESEARCHERS · iSCIENCE · 2026
Flexible data centers reduce power system costs but can increase emissions ↗
Workload shifting lowered modeled system costs, while its carbon effect depended on regional generation. This supersedes the 2025 working-paper version.
LAWRENCE BERKELEY NATIONAL LABORATORY · DECEMBER 2024
2024 United States Data Center Energy Usage Report ↗
The source of our 2023 direct and electricity-related water estimates. Includes separate estimates for facility cooling and electricity generation, with methods and future-demand scenarios.
KNOWABLE MAGAZINE / ANNUAL REVIEWS · AUGUST 24, 2026
How much of a problem is AI’s water use? ↗
Katarina Zimmer’s accessible explainer puts water claims in context while taking drought risk seriously. It discusses why query estimates change and what cooling technologies can do.
NATURE · MAY 2025 · INDUSTRY-AFFILIATED AUTHORS
Using life cycle assessment to drive innovation for sustainable cool clouds ↗
The basis for the cooling-reduction ranges. Comparative life-cycle modeling shows what better designs can achieve under stated assumptions.
WORLD RESOURCES INSTITUTE · EXPLAINER · 2025
What’s Really Driving Up US Electricity Prices? ↗
A strong general-reader companion to Berkeley Lab. It describes both the potential to spread fixed costs and the real capacity-market pressure associated with data-center growth in PJM.
DUKE NICHOLAS INSTITUTE · FEBRUARY 17, 2026
Data Centers and Generation Capacity over the Next Decade ↗
Ross and Ewing model flexibility, reserve provision and generation investment through 2035. The $40–150B headline covers most modeled scenarios; the full scenario range is approximately $7–150B, versus inflexible operation.
ITIF · POLICY REPORT · JULY 6, 2026
The Data Center Water Problem Is Soluble ↗
Robin Gaster’s policy report surveys water-saving technologies and governance, including an approximate national-consumption comparison. A policy argument, distinct from a peer-reviewed engineering study.
U.S. EPA · WATERSENSE
Residential outdoor water use ↗
EPA’s estimate of nearly 8B gallons/day of residential outdoor use is annualized in the chart. Most of this water is used for landscaping.
GCSAA · GOLF COURSE ENVIRONMENTAL PROFILE · 2025
Water Use and Conservation Practices on U.S. Golf Courses ↗
The industry’s 2024 estimate is 1.63M acre-feet applied for irrigation, converted here to approximately 531B U.S. gallons. This is water applied, not a measured consumptive footprint.
VIRGINIA JLARC · REPORT 598 · DECEMBER 2024
Data Centers in Virginia ↗
The state audit reports just over one-third reclaimed wastewater in Virginia data-center water use in 2023 (printed page 62).
NVIDIA · 2026 / MICROSOFT · 2024
Warm liquid loops and zero-evaporation cooling designs ↗
Primary vendor explanations of 45°C coolant and new cooling designs. These establish specific technical pathways, not adoption across the whole industry.
LINCOLN INSTITUTE OF LAND POLICY · LAND LINES
Data Drain: The Land and Water Impacts of the AI Boom ↗
A land-and-water perspective that links to golf and landscape estimates while stressing watershed planning. It is included as context, not as an endorsement of this site’s thesis.
THE WASHINGTON POST · SHANNON OSAKA · OCTOBER 25, 2025
The real reason electricity prices are rising, and it’s not data centers ↗
Mainstream coverage of Berkeley Lab’s price research. A newspaper subscription may be required.
GOOGLE CLOUD · AUGUST 2025
Measuring the environmental impact of AI inference ↗
Google’s measurement of the median Gemini Apps text prompt, including the 0.26mL water figure. Company-reported and specific to that workload and period.
THE END? THE CHOICES ARE STILL OURS.
Doomsday cancelled.
Science gets the sequel.
Less panic. More progress. Better infrastructure—by design.
Return to The Human Precipice →