A year ago, I published The New Math of AI Infrastructure to make the environmental cost of AI tangible. For a median Gemini Apps text prompt in May 2025, Google reported 0.24 watt-hours of electricity, 0.03 grams of carbon dioxide equivalent, and 0.26 milliliters of onsite cooling-water consumption.^[1]

Google calculated the water figure using per-prompt computing energy and a prior-year fleet average for consumptive water usage effectiveness (WUE), measured in liters of water consumed per kilowatt-hour used by computing equipment. The result excludes water used to generate electricity or train the model.^[1]

Those numbers describe one kind of task, on one company’s system, at one moment in time.

The largest environmental costs sit inside AI’s industrial system. It includes chips, server racks, transmission lines, substations, cooling equipment, backup generation, gas turbines, land, water, long-term power contracts, and communities being asked to host all of it. The math runs from a fraction of a watt-hour at your keyboard to gigawatts at a single campus.

This update focuses on the US environmental, infrastructure, and public-finance footprint of AI. Its unit is physical infrastructure and public cost; labor displacement, algorithmic harms, and the global mines-to-chips supply chain require separate ledgers.

Some of the numbers I used last year have held up. The central estimate from the International Energy Agency (IEA) for global data-center electricity use in 2030 remains roughly 945 to 950 terawatt-hours across its 2025 and 2026 assessments.^[2]

Several things moved much faster than I expected: US demand forecasts, capital spending, rack density, grid costs, and the speed with which data centers became a political issue.

I also got some things wrong. I described closed-loop cooling too broadly. I treated a conditional water calculation as if it proved a local shortage.

I merged two European heat-reuse examples into a claim neither source supported. I used a simple-prompt average to estimate a research process it was never designed to measure. And I wrote about “host communities” without asking whether race, income, zoning power, and inherited pollution shape who gets asked to host the infrastructure.

Here is the new math.

The Prompt Has Become a Terrible Unit

The 0.24 watt-hour prompt became a popular shorthand because people could understand it. One query used roughly as much electricity as running a 10-watt light-emitting diode (LED) bulb for a minute and a half.^[1]

The shorthand breaks when the word “prompt” covers radically different work.

A 2026 Microsoft Research analysis published in Joule estimated a median 0.31 watt-hours per query under realistic large-scale production assumptions, with an interquartile range of 0.16 to 0.60 watt-hours. In the paper’s modeled test-time-scaling scenario, the assumed median output rose from 300 to 5,000 tokens; estimated median energy rose about 13-fold, from 0.31 to 3.91 watt-hours. The researchers found that moving only 10 percent of requests into that long category could more than double fleet-wide inference energy.^[3]

Image generation, video, multi-step research, coding agents, and systems that call other tools belong in different categories again. The IEA says video, reasoning, and agentic tasks can use hundreds or thousands of times the energy of a simple text response.^[2]

The efficiency gains are real. Google reported a 33-fold reduction in energy per median Gemini prompt in one year.^[1] The 2026 Joule paper found a plausible path to another eight- to 20-fold reduction through better hardware, software, model design, and serving practices.^[3]

When each unit gets cheaper, people use more units, invent larger tasks, and embed the system in more products. AI agents can work for minutes or hours after one human instruction. A falling cost per token can coexist with a rising electric bill.

The honest consumer label would identify the task class, approximate output size, location and time assumptions, and uncertainty. One universal “energy per prompt” number offers precision without accuracy.

The Global Forecast Held. The US Range Blew Open.

Global data-center electricity use reached about 485 terawatt-hours in 2025, up 17 percent in one year. Electricity use at AI-focused facilities grew about 50 percent. The IEA still projects roughly 950 terawatt-hours by 2030, close to three percent of global electricity use, with AI-focused demand roughly tripling.^[2]

The IEA expects US electricity use to grow by more than 420 terawatt-hours from 2026 through 2030, with data centers responsible for about half. Globally, data centers account for less than 10 percent of total electricity-demand growth from 2024 through 2030; industry output growth and electrification, electric vehicles, and air conditioning lead.^[2]

Local concentration makes the global percentage a poor guide to regional impacts.

Lawrence Berkeley National Laboratory (LBNL) estimates that US data centers used 192 terawatt-hours in 2024, or 4.7 percent of the country’s electricity. Its 2030 reference case reaches 649 terawatt-hours, or 11.8 percent. The modeled range runs from 521 to 843 terawatt-hours, equal to 9.5 to 15.3 percent of US electricity.^[4]

The high case exceeds the low case by 322 terawatt-hours. The band of uncertainty alone is larger than the 192 terawatt-hours every US data center used in 2024. The uncertainty comes from questions with no clean answers yet: how much compute gets built, how quickly chips improve, how often servers run at full load, what kinds of AI people use, and whether power constraints delay projects.^[4]

The IEA’s US estimate for 2030 is much lower, about 426 terawatt-hours.^[2] Two respected institutions differ by more than 50 percent. Anyone presenting one national forecast as settled fact is selling confidence the evidence does not support.

The checks are already being written. The world’s five largest technology companies spent more than $400 billion in capital expenditures in 2025, already more than global investment in oil and natural-gas production.^[2]

The IEA expected their spending to rise by about 75 percent in 2026.^[2] The total includes businesses beyond AI.

OpenAI said in April 2026 that Stargate had secured more than 10 gigawatts of US AI infrastructure, meeting its 2029 commitment early after adding more than three gigawatts in 90 days.^[5] The announcement did not say how much was operating.

For scale, a sustained 10-gigawatt load would equal 87.6 terawatt-hours a year, about 46 percent of LBNL’s estimate of all US data-center electricity use in 2024. OpenAI did not define whether its figure was information-technology equipment load (IT load), total facility load, grid capacity, or another measure.^{[4, 5]}

The machines are getting denser too. The IEA says the power density of AI servers increased elevenfold from 2020 through 2025 and could rise another fourfold by 2027, when an advanced rack’s peak draw could equal that of 65 households.^[2]

The Grid Bill Has Arrived

PJM Interconnection, the grid operator serving 13 states and the District of Columbia, was the early warning in the original article. That warning is now a balance-sheet item.^[6]

PJM’s July 2026 capacity auction procured 138,318 megawatts of unforced capacity, a measure adjusted for expected performance, for the 2028–29 delivery year. The price reached $325 per megawatt-day, the top of a temporary price collar approved by federal regulators.^[6]

Even at that price, the auction came up 6,831 megawatts short of PJM’s reliability requirement. It was the second consecutive auction with a shortfall.^[6]

PJM’s independent market monitor put data centers at $6.3 billion, or 38 percent, of the $16.4 billion in charges from that July auction. Across PJM’s last four base capacity auctions, the monitor attributed $29.4 billion, or 46 percent, of $63.6 billion in total capacity charges to data-center load.^[7]

For the 2025–26 delivery year, the monitor estimated that data centers accounted for $9.3 billion, or 74.7 percent, of the increase in capacity-market revenue. During the first five months of 2026, it attributed $3.8 billion, or 23.3 percent, of the increase in total wholesale electricity costs to data-center demand.^[7]

Those estimates cover wholesale markets. Household bills also include transmission, distribution, fuel adjustments, taxes, and utility regulation. A clean-power contract can still leave the grid with transmission-upgrade and reserved-capacity costs.

The utility has its own incentive. In most states, investor-owned utilities that own new power plants can earn a regulator-set return on those investments. Peskoe and Martin argue that large new plants and transmission reinforce a capital-expansion model while lower-capital options, including flexible-load agreements, may produce less profit.^[39]

On July 22, 2026, a 230-kilovolt fault in Northern Virginia caused about 3,800 megawatts of data-center load to transfer almost simultaneously from the Dominion system to backup power. PJM called it the largest load-transfer event it had experienced. PJM returned the imbalance to within its operating limit in nine minutes.^[8]

A tightly clustered group of data centers can disappear from the grid almost at once, leaving generation suddenly greater than demand. Grid planners have spent a century preparing for power plants to trip offline. Now they must also plan for several gigawatts of customer load to vanish together.

The risk runs both ways: finding enough electricity when the campuses are online, then maintaining balance and frequency when gigawatts of load transfer to backup power nearly at once.

The Tax Code Is Part of the Deal

Electric bills are one public ledger. Tax expenditures are another.

xAI’s Southaven, Mississippi, project sits on both. Mississippi’s development agency approved xAI for a data-center sales-and-use-tax exemption, while Southaven and DeSoto County supported the project through fee-in-lieu agreements.^[38]

The state announced more than $20 billion in investment and “hundreds” of permanent jobs. Its public announcement did not value either tax benefit.^[38]

Washington’s legislative auditor found that at least 38 states have tax preferences written specifically for data centers.^[38]

Virginia data-center operators reported $928.6 million in state, local, and regional sales-tax savings in fiscal 2023, $1.29 billion in fiscal 2024, and $1.94 billion in fiscal 2025. The return changes with the accounting boundary. Virginia’s 2024 legislative analysis estimated 48 cents in state revenue per state dollar exempted.^{[23, 38]}

A 2026 state model included local revenue and projected returns through 2029, producing $1.70 for every dollar exempted. Its inputs were self-reported and not independently validated, and the model excluded separate local discretionary incentives. In a separate counterfactual, the report retained an assumption that 90 percent of qualifying investment would not have occurred without the exemption.^{[23, 38]}

Georgia’s state auditor used a different counterfactual. It estimated that 30 percent of data-center activity was attributable to the exemption and 70 percent would have occurred anyway.^[38]

For fiscal 2025, the report estimated $474.2 million in forgone state revenue and $41.5 million in additional state tax collections tied to the incentive. Its estimate of broader economic value was positive. The fiscal balance was negative.^[38]

Capital investment and public return belong in separate columns. Before a vote, publish the subsidy’s present value, state and local shares, permanent jobs and wages, public infrastructure costs, “but for” assumption, clawbacks, expiration date, and enforceable community terms.

The opportunity cost belongs in the same ledger. When subsidized activity would have occurred anyway, the forgone revenue is unavailable for schools, health departments, or other public services. Any induced replacement revenue belongs in the same counterfactual.^[38]

Gas Is Filling the Construction Gap

Data centers can be built faster than new transmission, large power plants, and interconnection studies. The timing mismatch is pushing some companies toward on-site generation.

The IEA projects that installed onsite gas-fired generation for data centers could reach 15 to 27 gigawatts globally by 2030, with growth mostly concentrated in the United States. Providing grid-like reliability may require 30 to 70 percent more nameplate capacity than expected data-center demand because turbines need maintenance and backup.^[2]

Wind, solar, batteries, nuclear power, and geothermal energy all have roles. Each has a clock.

A solar project may wait years in an interconnection queue. A transmission line can take a decade.

New nuclear plants will not cover the next few years of demand. Today’s batteries shift electricity across hours. Seasonal storage requires a much larger and more expensive system.

Gas turbines are dispatchable and familiar to utilities. They also lock in fuel use, air pollution, methane leakage, and carbon emissions for equipment that can operate for decades.

The IEA says data-center emissions could double to about 350 million metric tons by 2035, around two percent of electricity-sector emissions.^[2] That percentage remains small beside the rest of the energy system. It is large enough to strain corporate climate promises and shape what gets built next.

The data center itself can also provide flexibility. Beyond uninterruptible-power-supply capacity, the IEA projects dedicated longer-duration onsite battery storage at data centers to rise from about five gigawatts worldwide in 2025 to 20 to 25 gigawatts by 2030. It expects 10 to 12 gigawatts, the majority, in the United States.^[2]

Some computing jobs can pause or move to another region. Backup systems can support the grid when contracts and controls permit it.

PJM proposed an interim framework in August 2026 that would let some large new loads connect sooner if they agree to be curtailed first during resource emergencies. The proposal still requires federal approval.^[9]

Grid access may soon come with a different bargain: firm power costs more, flexible power connects faster, and the data center pays for the difference.

Data Centers Arrive on an Unequal Power Map

Data centers arrive on maps shaped by redlining, industrial zoning, highway construction, utility corridors, disinvestment, and unequal political power. The power system feeding them carries that history too.

A 2023 Nature Energy study examined 8,871 neighborhoods in 196 US urban areas. Compared with areas graded C in federal housing maps from the 1930s, redlined areas were 31 percent more likely to have an upwind fossil-fuel power plant sited within five kilometers from 2000 through 2019.^[10]

Among neighborhoods near operating plants, the redlined areas had 82 percent higher nitrogen-oxide emissions, 38 percent higher sulfur-dioxide emissions, and 63 percent higher fine-particle emissions in the study’s present-day comparison. The researchers measured plant siting and emissions burden, not actual exposure or health effects.^[10]

The US Environmental Protection Agency (EPA) found a similar disparity in its analysis of 2021 data. Among people living within three miles of more than 1,200 fossil-fuel plants covered by the Acid Rain Program and Cross-State Air Pollution Rule programs, 53 percent were people of color and 34 percent lived in households at or below twice the federal poverty level.^[10]

The national shares were 40 and 30 percent.^[10] Data-center demand inherited that generation map.

The equity result changes when the boundary changes.

A May 2026 working paper from the University of California, Berkeley (UC Berkeley) used Aterio records for 5,105 US data-center campuses. Its demographic analysis compared census tracts containing past, new, or announced campuses with tracts containing none.^[11]

Across those categories, the authors found no consistent national tendency for data centers to enter lower-income, higher-poverty, or more nonwhite tracts. Planned sites were moving into tracts with 383 people per square kilometer, compared with 2,089 elsewhere.^[11]

The time series was less tidy than the headline. Past data-center tracts were 50 percent non-Hispanic white, compared with 58 percent of other tracts.^[11]

New-site tracts were 55 percent, a difference that was not statistically significant. Planned-site tracts were 60 percent.^[11]

Sites entering through 2027 generally appeared in tracts with lower non-Hispanic white shares; the direction reversed among announcements for 2028 through 2030.^[11] It remains a working paper. Its national result is a necessary check on a blanket racial-siting claim.

Two state studies found patterns the national comparison compresses.

The Kapor Foundation began with 288 California entries from Data Center Map, removed colocated entries sharing an address, and mapped 226 addresses to CalEnviroScreen. Eighty-two percent were in tracts in the top two quintiles for diesel particulate matter, 65 percent in areas with the highest groundwater-threat classification, and 79 percent in the top quintile for hazardous waste.^[34]

The statewide analysis documents colocation with inherited hazards. It did not test race, establish that data centers caused those burdens, or weight sites by capacity.^[34]

Front and Centered cross-referenced five public lists in Washington and retained 81 individual buildings appearing in at least two. Fifty-seven percent were in tracts in the top two state deciles for the share of residents who are people of color, compared with two percent in the bottom two.^[35]

The inventory was incomplete, counted multiple buildings at one campus separately, and gave each building the same weight regardless of power capacity. It found little relationship with poverty.^[35]

Berkeley, Kapor, and Front and Centered all use census tracts. They ask different questions and count different objects: US campuses, deduplicated California addresses, and Washington buildings.

A national demographic comparison can coexist with regional clustering and inherited burden. The distance between the results comes from geography, indicators, and counting rules chosen before the analysis.

Berkeley’s pollution model adds the power map. It attributes roughly 97 percent of projected growth in grid-related carbon emissions and monetized local-air-pollution damages from 2024 through 2030 to the scale of added electricity demand.^[11]

Shifts in siting among 13 electricity regions account for about three percent.^[11] The model does not cover onsite turbines, water, noise, or fence-line exposure.

Nicholas Muller’s separate working paper estimated that electricity used by roughly 2,800 operating US data centers produced about $25 billion in gross external damages in 2025. His modeled range was $10 billion to $33 billion, with Texas and Virginia accounting for 30 percent of the national total.^[37]

The estimate covers grid-related air pollution and greenhouse gases. Observed medical spending and racial incidence sit outside its boundary.

The host tract and the people receiving the environmental bill occupy different maps.

Environmental justice asks who carries the cumulative burden, who receives the benefits, and who had power while the decision could still change. I use environmental racism to describe a documented racialized pattern in the distribution of environmental burdens or decision-making power. That analytical description is separate from a legal finding of intentional discrimination.

Memphis shows what happens when several of those maps overlap.

In a 2021 Title VI complaint over a separate oil-pipeline proposal, community organizations described Boxtown as a freedmen’s community founded by formerly enslaved people after Emancipation. The filing said ZIP code 38109 was 97 percent Black, nearly half of its households earned less than $25,000 a year, and a 2013 analysis placed southwest Memphis’s cumulative toxic-air cancer risk at about four times the national average.^[12]

Those were the complainants’ evidence, not a final EPA finding. The filing described an industrial burden that existed before AI arrived.^[12]

KeShaun Pearson grew up in southwest Memphis and leads Memphis Community Against Pollution. The group organized opposition to xAI’s turbine permit and partnered with researchers to install air monitors in Boxtown, Walker Homes, and Southaven.^[40]

“For almost a decade in southwest Memphis, we have not had a sensor to tell us the quality of the air we breathe,” Pearson said in 2026.^[40]

Three years later, xAI placed its Colossus data center at 3231 Paul R. Lowry Road in the same ZIP code. In June 2025, the Southern Environmental Law Center, on behalf of the National Association for the Advancement of Colored People (NAACP), sent xAI a notice of intent to sue.^[13]

It alleged that xAI had installed and operated at least 35 gas turbines without required permits and estimated the fleet’s potential annual nitrogen-oxide emissions at roughly 1,200 to 2,100 tons, depending on the assumptions. Those were allegations and estimates in a pre-suit notice, not findings by a court or regulator.^[13]

Shelby County later issued a construction permit to CTC Property LLC, the xAI affiliate that holds the site’s permits, covering 15 controlled turbines with 247.2 megawatts of combined capacity. In April 2026, CTC Property requested an operating permit and a reduction from 15 permitted turbines to 12. The public notice opened comment through May 8.^[13]

The buildout then crossed the state line. Mississippi issued xAI subsidiary MZX Tech a March 2026 construction permit for 41 stationary turbines at its Southaven site, which powers Colossus 2. A July agreed order recorded 69 portable temporary turbines there and set phased removal deadlines through July 2027.^[13]

The NAACP filed a federal Clean Air Act suit over the turbines in April. On June 15, the Department of Justice moved to intervene and dismiss.^[13]

EPA’s January 2026 turbine rule gives both sides something to quote. The agency said combustion turbines covered by the federal New Source Performance Standards, including portable units, are stationary sources rather than nonroad engines. That language undercuts a categorical mobile-source loophole. It also sits in tension with Mississippi’s agreed order, which treats turbines kept on portable platforms for limited periods as mobile sources under the state plan.^{[13, 36]}

The same federal rule created an optional temporary category for qualifying small and medium turbines: no more than 24 months at one source, a 25-parts-per-million nitrogen-oxide standard when burning natural gas, and reduced monitoring and reporting. Moving or replacing turbines at the same source does not restart the clock.^[36]

Some non-major turbine sources can also avoid Title V operating permits, although EPA said new-source construction permitting generally still applies. The rule clarifies the federal framework. It does not decide which requirements applied to either xAI fleet or whether the state and local permits were lawful.^[36]

Boxtown supplies a site-specific record. A global company added fossil generation in the wider southwest Memphis industrial corridor near a historically Black community that had already spent decades living with industrial pollution.

The public fight over the permit followed the arrival and operation of turbines. That sequence puts race, inherited burden, and procedural power inside the infrastructure math without turning one case into a national prevalence estimate.

EPA defines cumulative impacts as “the totality of exposures to combinations of chemical and nonchemical stressors and their effects on health, well-being, and quality of life outcomes.”^[14] A permit limit asks what one source may emit. A cumulative review asks what happens when that source is added to everything already there.

Who pays has an equity dimension too. The US Energy Information Administration’s preliminary 2024 survey found that 55 percent of households with a Black householder and 50 percent of households with a Hispanic householder reported at least one form of household energy insecurity, compared with 26 percent of households with a non-Hispanic white householder.^[15]

The survey measures household energy insecurity; it contains no estimate of data-center effects. The disparities show why a shift of grid costs onto households could land unevenly.

That review belongs before incentives, land transfers, water commitments, and power contracts. Who lives closest? What burdens are already present?

Where will the electricity be generated? Who receives the jobs, tax revenue, and public investment? Who had power before the land, tax, water, and utility agreements were locked in?

Water Engineering Improved. Total Use Rose.

A closed liquid loop carries heat away from chips and recirculates that fluid. The heat still has to go somewhere. Some facilities transfer it to a condenser-water loop and then to a cooling tower, where makeup water replaces evaporation and blowdown.

Evaporation is consumptive use. Blowdown is a discharge, and whether a given accounting framework counts it as consumption depends on where that framework draws its return-flow boundary.

The word “closed” describes one loop. It says nothing by itself about the final heat-rejection system.^[16]

Evaporative cooling remains common. In Uptime Institute’s online 2025 operator survey, 22 percent of respondents, from a sample of 400, named evaporative cooling towers as their primary external heat-rejection method. Thirty-five percent named air-cooled chillers.^[17]

In a separate question with 512 respondents, 22 percent reported using direct liquid cooling somewhere in their data centers.^[17] These are respondent-level adoption figures rather than market shares weighted by capacity, facility count, energy, or load. They do not support the claim that engineers have “eliminated water cooling.”

Microsoft reported a fiscal 2025 global WUE of 0.27 liters per kilowatt-hour for cooling and humidification at fully owned and controlled data centers operational for all 12 months, down from 0.30 in fiscal 2024. The Americas averaged 0.34, Asia-Pacific 0.25, and Europe, the Middle East, and Africa 0.03.^[18]

Microsoft’s current page describes the numerator only as annual liters of water “used” for humidification and cooling; it does not say whether 0.27 measures withdrawal or consumption. A fleet average cannot describe one campus, one season, or one model run.^[18]

Microsoft describes its next-generation design as “zero water evaporation for cooling.” After an initial fill during construction, a closed loop circulates water without evaporation.^[19]

Administrative uses such as restrooms and kitchens remain, and Microsoft expects a nominal increase in annual energy use compared with evaporative designs. The company said pilot projects in Phoenix and Wisconsin would use the design, with the new sites beginning to come online in late 2027. Its existing fleet still uses a mix of air- and water-cooled systems.^[19]

Google makes the trade explicit. The company says water cooling can reduce energy use by about 10 percent compared with air cooling in many locations.^[20]

In 2025, Google’s data centers consumed 10.523 billion gallons of water. Across data centers, offices, and other facilities, total operational consumption reached 10.869 billion gallons, up 34 percent from 2024. Google reported 7.717 billion gallons of estimated freshwater-replenishment benefits, equal to 78 percent of its 9.947-billion-gallon freshwater consumption.^[21]

Replenishment is an estimated volumetric benefit from a portfolio of watershed projects. It does not physically replace each site’s consumption in the same basin or on the same schedule. Report it beside withdrawal and consumption.

National totals create another trap. Lawrence Berkeley National Laboratory estimated that the entire US data-center fleet, including non-AI workloads, directly consumed about 66 billion liters of water in 2023. Its electricity carried a modeled indirect consumption footprint of nearly 800 billion liters, based on regional balancing-authority grid mixes.^[22]

The indirect estimate excludes facility-specific power-purchase agreements and behind-the-meter generation. These are different, complementary boundaries.^[22]

Last year, I multiplied one gigawatt of assumed IT load running continuously for 720 hours by Microsoft’s 0.30-liter WUE. The calculation produces 216 million liters under a Microsoft metric with an undefined accounting category: withdrawal, consumption, or another “water used” boundary.^{[18, 19]}

Microsoft’s 2024 post defines WUE in the body as total annual water consumption for cooling and humidification, but its footnote calls the same 0.30 figure a global average withdrawal WUE. Its current efficiency page says only water “used.”^{[18, 19]} I treated an internally inconsistent company metric as if it were a direct measure of local water scarcity.

The scenario also assumes constant full load and a site matching Microsoft’s global fleet average. Assessing risk to a town’s taps requires local withdrawal, consumption, return flow, water source, potable and reclaimed shares, peak-day demand, drought rules, utility capacity, wastewater limits, and competing uses.

Boundary discipline makes the water case more durable. Site and watershed evidence identifies the source, season, and competing demand that decide local risk.

Virginia shows why both alarm and reassurance need an address. The state’s Joint Legislative Audit and Review Commission (JLARC) estimated that data centers used 2.1 billion gallons in 2023, with just over one-third supplied by reclaimed water. Their use represented less than 0.5 percent of statewide withdrawals.^[23]

At the six utilities JLARC reviewed, data centers accounted for 0.2 to 21 percent of water use after reclaimed water was excluded. One building used 243 million gallons.^[23] The building-level dataset covered a large majority of Virginia data centers, with some figures approximated; it did not cover every building or aggregate entire campuses.

Waste Heat Has Value and a Zip Code

Nearly all electricity entering a data center ultimately becomes heat, and LBNL estimates that roughly 70 to 80 percent could be recoverable. Advanced liquid and two-phase cooling can raise dissipated-heat temperatures from roughly 30°C with typical air cooling to more than 50°C, improving reuse prospects.^[24]

Two European projects show both the opportunity and the time lag.

Meta’s operating heat-reuse system in Odense, Denmark, supplies roughly 11,000 households, according to the local district-heating utility.^[25] The Finnish project I referenced involves two Microsoft data-center sites and roughly €225 million of investment in Fortum’s heat-pump plants, supported by European Union and Finnish government funding.^[26]

Its heat-pump plants began operating on ambient air and electric boilers in 2026. Data-center waste heat is expected to enter the system in phases beginning in 2027. At full buildout, Fortum expects the project to cover about 40 percent of a two-terawatt-hour district-heating network serving 250,000 heat users.^[26]

Heat reuse needs a nearby customer, pipes, heat pumps, capital, workable regulation, and demand when the heat is available. A data center beside a low-temperature district-heating network has a real opportunity. A remote campus in a warm climate may lack a buyer.

Require a heat-reuse feasibility study for every large project. Public reporting can show delivered heat instead of nameplate potential. Local rules can require recovery where the engineering and economics work.

Volume Is Winning the Efficiency Race

AI infrastructure is becoming more efficient at the chip, model, and facility levels. Absolute use keeps rising because companies are installing vastly more equipment and asking it to do more work.

Google’s electricity demand rose 37 percent in 2025 while its operational emissions fell two percent.^[21] The company paired rapid load growth with cleaner electricity and operational changes, including clean-energy procurement at a scale few others can match.

Microsoft reported that its total Scope 1, 2, and 3 emissions rose 25 percent in fiscal 2025, driven largely by data-center construction and a change in how the company treated non-additional renewable-energy certificates.^[27] Concrete, steel, chips, buildings, and equipment replacement sit inside that number. A highly efficient operating campus still begins with a carbon-intensive construction project.

Annual clean-energy matching also differs from hourly clean power. A company can buy enough renewable energy over a year to match its annual consumption while drawing gas- or coal-generated electricity during hours when wind and solar output are low. Contracts can finance new generation. Hourly power still comes from the plants running on the local grid.^{[21, 27]}

The useful disclosure has several layers: facility electricity, task class, annual and hourly energy sources, operational carbon, construction carbon, hardware turnover, direct water, electricity-associated water, and local watershed stress.

Put those layers on a dashboard and let readers see each one.

The Benefits Need Receipts

AI’s health, disaster, and climate applications belong on the opportunity side of the ledger. Their evidence still needs a boundary.

The US Forest Service’s Potential Control Location model uses machine learning and more than 20 years of fire outcomes to help incident teams identify places where a wildfire may be contained. In evaluations of 2022 and 2023 fires, it correctly predicted more than 80 percent of successful control locations and 90 percent of unsuccessful ones.^[28]

The agency’s FireCon system combines daily fire weather, fuel dynamics, and resource deployments to estimate containment suitability across the western United States. It is restricted to trained fire professionals. The Forest Service says plainly that it cannot replace human judgment or guarantee success.^[29]

Google reports that its systems provide flood-forecasting information for more than two billion people across about 150 countries for the most significant riverine flood events.^[21] The IEA estimates that existing AI applications could produce more than 13 exajoules of energy savings by 2035, equal to about three percent of global final energy use, if adoption and implementation barriers are overcome.^[2]

“Could” is doing honest work in that sentence. Potential savings and forecasts are different claims. A model’s existence does not prove that institutions use it well.

Put benefits on the same ledger as power, water, carbon, and cost. Each campus earns its place through its local terms.

Policy Caught Up Faster Than I Expected

Several safeguards have moved into tariffs, federal proceedings, and corporate commitments.

In March 2026, Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI signed the White House Ratepayer Protection Pledge. The companies committed to build, bring, or buy new generation; pay for transmission and other delivery infrastructure; accept separate rate structures; pay contracted costs even when they use less power than expected; support grid resilience through backup generation and flexible operation; and invest in local hiring and workforce development. On July 23, the White House announced that more than 200 additional utilities, data-center developers, cooperatives, and states had joined.^[30]

A pledge is a promise. A tariff, deposit, contract, or regulatory order decides who pays when a project changes course.

The Federal Energy Regulatory Commission opened proceedings in June 2026 requiring six regional grid operators to explain whether their rules protect other customers from the cost and reliability risks of large new loads. Commissioners specifically raised cost-recovery agreements, public reporting, speculative projects, and the danger of building infrastructure for a data center that never arrives.^[31]

Virginia went further. Its GS-5 large-load rate schedule covers Dominion customers with measured or contracted demand of at least 25 megawatts on a contiguous site and a measured or expected load factor of at least 75 percent.^[32]

New customers signing service agreements on or after January 1, 2027, face 14-year terms. Minimum-demand charges generally apply at 85 percent of contracted transmission and distribution demand and 60 percent of generation demand, with collateral and exit-fee protections.^[32]

Virginia’s tariff begins a workable compact. The data center gets a path to power. The public gets a named customer, a long contract, and protection if the campus arrives late, uses less than promised, or walks away.

A common rulebook would keep communities from inventing terms campus by campus while negotiating with companies that have larger budgets than some states.

Community pressure is changing the policy map. Data Center Watch, an opposition tracker, counted at least 75 US projects worth about $130 billion blocked or delayed during the first quarter of 2026. Its figure combines blocked and delayed projects; the public summary does not publish project-level methods, so treat it as the group’s measure of project friction.^[41]

Community infrastructure is forming too. Erin Brockovich’s volunteer-run initiative gives residents a place to report concerns and follow local legislation. Its resident-submission map includes operating, proposed, and rumored sites; facility counts and prevalence require other sources.^[42]

In July, New York paused discretionary environmental permits for hyperscale data-center applications not already deemed complete for up to one year while it develops a statewide environmental review. The order also calls for a community-investment framework with organized labor at the table, wage standards, local hiring, apprenticeships, and direct community benefits.^[41]

Put the Risk on the Right Invoice

First, charge large loads for connection risk and reward flexibility. Require deposits, minimum payments, take-or-pay provisions, exit fees, and direct funding for dedicated transmission and distribution upgrades. Faster interconnection can come with batteries, scheduled computing, non-firm service, and curtailment rights during grid emergencies. The customer creating the risk belongs on the contract.

Second, publish every environmental boundary and require a cumulative-impact review. Before incentives or permits, publish analysis at the most granular reliable geography, using block-group indicators where valid. Include race, income, existing pollution, health burdens, residential proximity, nighttime noise, heat, water source, peak-day demand, wastewater chemistry, and every onsite or enabling power source. Compare alternative sites and designs.

Then report annual withdrawal, consumption, WUE, cooling architecture, final heat rejection, electricity use, hourly and annual power sources, construction emissions, and hardware turnover. Power usage effectiveness (PUE) is total facility electricity divided by electricity delivered to computing equipment. Include drought scenarios, cumulative campus demand, wastewater and brine plans, and enforceable limits on potable water in stressed basins.

A screening map starts the inquiry; causation and discrimination require separate evidence.

Third, count the physical chain and put community power and benefits in the contract. The impact statement starts with mines, chip fabrication, concrete, steel, substations, transmission, backup systems, water treatment, and equipment disposal. Start public review while assumptions, alternatives, and cost allocation can still change. Publish the full incentive agreement and its present value before approval.

Disclose rate treatment, emergency-service needs, noise rules, water commitments, and responsibility for each infrastructure upgrade. Separate state and local revenue forgone, permanent jobs from construction jobs, public infrastructure costs, the “but for” assumption, and subsidy per permanent job.

Require annual performance reports, clawbacks, an expiration date, and enforceable power, water, pollution, and community-benefit conditions. Give affected communities money and authority for independent technical review. Permit compliance and health protections are the floor; community benefits should be additional and enforceable.

Foundations and public-interest funders can pay for independent engineering, legal counsel, health monitoring, and resident organizing before the permit hearing.

Fourth, give users task-class information. A simple text question, a long reasoning job, an image, a video, and an autonomous agent need separate ranges. Product-level estimates can show those ranges and their assumptions. False precision will make people distrust the disclosure.

My Own Accounting

Last year, I estimated that researching and writing the article used 15 to 20 watt-hours of electricity and roughly the same number of milliliters of water. I multiplied an estimated number of interactions by a median simple-prompt figure.

I withdraw that estimate.

The work involved web searches, long outputs, document review, revisions, and different systems in unknown locations. An LBNL assessment found more than 10,000-fold variation in liters consumed per workload, driven by more than 1,000-fold differences in water consumed per kilowatt-hour of server electricity and roughly tenfold differences in server workload efficiency.^[33]

It ranks server efficiency, grid water intensity, utilization, cooling, infrastructure efficiency, climate, inactive servers, and refresh cycles among the key determinants.^[33] The platforms did not provide the telemetry needed to calculate my share.

The accurate answer is: I do not know.

I used AI again for this update. It helped collect primary sources, compare forecasts, check arithmetic, and challenge my claims. I am responsible for every sentence that survives publication.

The Verdict

The per-prompt debate asked whether using AI was environmentally irresponsible. That question was always too personal and too small.

The material decisions now sit in utility commissions, economic-development boards, zoning meetings, interconnection queues, corporate capital budgets, cooling designs, and long-term contracts. The global figure may land near three percent of electricity in 2030. At one local water utility, data centers can claim 21 percent of demand.^{[2, 23]}

AI already supports flood forecasts and wildfire-containment decisions. The campuses behind it can raise power costs, extend fossil-fuel use, consume scarce water, absorb public subsidies, and leave communities holding assets built for a customer that never arrived.

An environmental racism inquiry changes the unit again. A national tract analysis finds no consistent racial siting pattern. California data-center addresses cluster in tracts already burdened by diesel pollution and hazardous waste.

A Washington building inventory shows a strong racial concentration. Berkeley’s grid model puts most projected pollution growth on the scale of added demand.

In Memphis, the turbine dispute shows how cumulative burden and procedural power can converge at one site. These results describe different geographies, indicators, and parts of the physical chain.

Name those boundaries before drawing the conclusion. Then ask whether race, income, and inherited pollution help predict who bears the next increment, where the enabling power is generated, and whether the people living there had power before the deal was done.

Build with discipline. Count the whole system and publish the assumptions. Put the risk on the large-load customer, give the community enforceable terms, and use the machines for work worth doing.

Behind the meter

Equipment or generation on the customer’s side of the utility meter. A data center’s backup generators and batteries usually sit behind the meter.

Capacity market

A market that pays power resources to be available in a future delivery year, separate from payment for the electricity they actually produce.

Interconnection and curtailment

Interconnection is the technical and contractual process for connecting a generator or large user to the grid. Curtailment is a required or voluntary reduction in electricity use during stressed conditions.

Cumulative impact

The totality of exposures to combinations of chemical and nonchemical stressors and their effects on health, well-being, and quality of life outcomes. A cumulative-impact assessment can include existing and proposed sources, community vulnerability, and the distribution of burdens and benefits.

Environmental justice

Fair treatment and meaningful involvement of affected people in environmental laws and decisions, including attention to how environmental burdens and benefits are distributed.

Environmental racism (as used here)

A pattern in which environmental burden or decision-making power is distributed along racial lines through public policy, markets, zoning, infrastructure, enforcement, or institutional practice. A legal finding of intentional discrimination requires separate evidence.

Power scale

A gigawatt is one billion watts. A terawatt-hour is one billion kilowatt-hours. One gigawatt drawn continuously for a year equals 8.76 terawatt-hours.

IT load and load factor

IT load is electricity used by computing, networking, and storage equipment, excluding facility overhead such as cooling and power conversion. Load factor is average use divided by peak or contracted demand over a period.

Power usage effectiveness (PUE)

Total facility electricity divided by electricity delivered to computing equipment. A PUE of 1.20 means the facility uses 20 percent beyond the IT load for cooling, power conversion, lighting, and other systems. PUE is an efficiency ratio. Absolute electricity use, computing efficiency, and water performance require separate measures.

Scope 1, 2, and 3 emissions

Direct company emissions; emissions associated with purchased energy; and value-chain emissions such as construction, manufacturing, suppliers, and product use.

Water usage effectiveness (WUE)

Annual water attributed to site operations in liters divided by annual IT electricity in kilowatt-hours. The numerator may refer to withdrawal, consumption, or simply water “used” for specified functions such as humidification and cooling; disclosures do not always distinguish among them. State the numerator category, facility boundary, and included uses before comparing WUE values.

Water withdrawal and consumption

Withdrawal is water removed from a surface-water or groundwater source. Consumption is the portion evaporated or otherwise unavailable for immediate reuse in the local water environment.

Site and electricity-associated water

Site water is withdrawn or consumed inside the data-center boundary. Electricity-associated water is attributed to generation. Each figure needs to identify withdrawal or consumption and whether electricity attribution follows the physical grid mix, contractual procurement, or another boundary.

Closed-loop and zero water evaporation

A closed loop recirculates coolant through the IT or secondary cooling system; the final heat-rejection system may still use evaporative cooling. “Zero water evaporation for cooling” says the operating cooling system does not evaporate water. Initial fill, administrative water, total site withdrawal, and electricity-associated water remain separate quantities.

1. 1. Google, “Measuring the Environmental Impact of Delivering AI at Google Scale” (2025), PDF.
2. 2. International Energy Agency, World Energy Outlook Special Report: Key Questions on Energy and AI (April 16, 2026), Executive Summary and full report. The 2025 baseline of 485 TWh, the 17 percent growth rate, the 50 percent growth at AI-focused facilities, the 950 TWh projection for 2030, the capital-expenditure figures, rack power density, onsite gas capacity, battery storage, the 350 million tonne emissions estimate, and the 13 exajoule savings estimate all come from this report. The 945 TWh central estimate for 2030 and the 426 TWh US estimate for 2030 come from the earlier IEA report Energy and AI (April 10, 2025), source. US demand growth of more than 420 TWh from 2026 through 2030 comes from IEA, Electricity 2026, “Demand”.
3. 3. Microsoft Research, “Energy Use of AI Inference: Efficiency Pathways and Test-Time Scaling,” Joule (2026), publication page and full paper.
4. 4. Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update (June 2026), report page.
5. 5. OpenAI, “Building the Compute Infrastructure for the Intelligence Age” (April 29, 2026), source.
6. 6. PJM Interconnection, “PJM Capacity Auction Procures 138,318 MW of Generation Resources as Work Continues to Address Growing Electricity Demand” (July 14, 2026), source; PJM, 2028–29 Base Residual Auction Results.
7. 7. Monitoring Analytics, PJM Market Monitor Report presentation (July 28, 2026), PDF, for the July auction and four-auction figures; PJM Market Monitoring Report presentation (June 24, 2026), PDF, for the earlier delivery-year and wholesale-cost figures.
8. 8. PJM Interconnection, “July 22 Dominion Load Transfer Event” (July 31, 2026), PDF.
9. 9. PJM Interconnection, “PJM Proposes Framework to Connect Data Centers Without Compromising Reliability, Affordability” (August 13, 2026), source.
10. 10. Lara J. Cushing et al., “Historical Red-Lining Is Associated with Fossil Fuel Power Plant Siting and Present-Day Inequalities in Air Pollutant Emissions,” Nature Energy 8 (2023), article; US Environmental Protection Agency, “Affected Communities,” 2023 Power Sector Programs Progress Report, source.
11. 11. Danae Hernandez-Cortes, Kyle Meng, and Paige Weber, “The Environmental Costs and Geography of U.S. Data Center Expansion,” Energy Institute at Haas Working Paper 363 (May 2026), abstract and paper.
12. 12. “Title VI Complaint Regarding the Tennessee Department of Environment and Conservation” (May 16, 2021), complaint, reproduced in the White House Environmental Justice Advisory Council’s May 2021 public-meeting EPA record.
13. 13. NAACP and Southern Environmental Law Center, “Notice of Intent to Sue for Violations of the Clean Air Act” (June 17, 2025), NAACP summary and letter with appendices; Shelby County Health Department, CTC Property LLC construction permit (July 2, 2025), copy reproduced in a South Carolina Public Service Commission docket, and operating-permit public notice (April 7, 2026), county copy and Daily Memphian reprint; Mississippi Department of Environmental Quality, MZX Tech LLC permit record and agreed order (July 30, 2026); NAACP v. X.AI Corp., No. 3:26-cv-00074 (N.D. Miss.), complaint (April 14, 2026); US Department of Justice, motion for intervention and dismissal (June 15, 2026).
14. 14. US Environmental Protection Agency, “Cumulative Impacts Explained,” updated October 31, 2025, source.
15. 15. US Energy Information Administration, “Household Energy Insecurity, 2024,” preliminary 2024 Residential Energy Consumption Survey, Table HC11.1 (March 2026), PDF.
16. 16. US Department of Energy, Federal Energy Management Program, “Cooling Water Efficiency Opportunities for Federal Data Centers,” source.
17. 17. Uptime Institute, Cooling Systems Survey 2025: DLC Adoption Remains Slow and Steady, UII Data Report 181 (July 2025), PDF.
18. 18. Microsoft, “Measuring Energy and Water Efficiency for Microsoft Datacenters,” fiscal 2025 data, source. Microsoft defines its current WUE numerator as annual liters of water “used for humidification and cooling” but does not classify the fiscal 2025 figure as withdrawal or consumption.
19. 19. Microsoft, “Sustainable by Design: Next-Generation Datacenters Consume Zero Water for Cooling” (December 9, 2024), source. The body calls the fiscal 2024 WUE numerator total annual water consumption; a footnote calls the same 0.30 L/kWh value Microsoft’s global average withdrawal WUE.
20. 20. Google, “Google’s Water Stewardship Commitments for Local Communities” (June 3, 2026), source.
21. 21. Google, 2026 Environmental Report, report page and PDF.
22. 22. Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (December 2024), report page.
23. 23. Virginia Joint Legislative Audit and Review Commission, Data Centers in Virginia (December 2024), report and summary.
24. 24. Lawrence Berkeley National Laboratory, “Avoiding Waste Heat through AI Infrastructure Thermal Integration” (2026), source.
25. 25. Fjernvarme Fyn, “Glædens dag for Fjernvarme Fyns kunder: Udskældt prisloft forsvinder helt” (January 2025), source.
26. 26. Fortum, “Fortum Has Started Heat Production at Two Large Data Centre Sites in Finland” (May 6, 2026), source.
27. 27. Microsoft, 2026 Environmental Sustainability Report, report page; Microsoft, “Responsibly Building the AI Future” (July 9, 2026), source.
28. 28. US Forest Service, “Potential Control Location Suitability Model,” updated May 14, 2026, source.
29. 29. US Forest Service, “FireCon: Daily Fire Containment Suitability,” updated May 14, 2026, source.
30. 30. The White House, “President Donald J. Trump Advances Energy Affordability with the Ratepayer Protection Pledge” (March 4, 2026), source; “President Trump’s Ratepayer Protection Pledge Secures American AI Dominance, Protects Consumers” (July 23, 2026), source; pledge page.
31. 31. Federal Energy Regulatory Commission, “FERC Launches Aggressive, Targeted Action to Speed Large Load Integration” (June 18, 2026), source; Commissioner Rosner, remarks on the large-load orders.
32. 32. Virginia State Corporation Commission, Final Order, Case PUR-2025-00058 (November 25, 2025), PDF and news release.
33. 33. Lawrence Berkeley National Laboratory, The Water Use of Data Center Workloads: A Review and Assessment of Key Determinants, Resources, Conservation and Recycling 219 (2025), report page and DOI.
34. 34. Kapor Foundation, The Unequal Burden of Data Centers: An Examination of the Environmental and Public Health Impacts on Communities in California (December 9, 2025), report page and PDF.
35. 35. Front and Centered, Communities Centered: Frontline Perspectives on Hyperscale Data Centers in Washington State (July 2026), report page and PDF.
36. 36. US Environmental Protection Agency, “New Source Performance Standards Review for Stationary Combustion Turbines and Stationary Gas Turbines,” final rule, 91 Fed. Reg. 1910, 1925–27, 1960, 1979–80, 1987–88 (January 15, 2026), source; correction, 91 Fed. Reg. 43561–64 (July 16, 2026), source.
37. 37. Nicholas Z. Muller, Measuring the Impact of Data Centers in the United States Economy: Monetary Damage from Air Pollution and Greenhouse Gas Emissions, NBER Working Paper 35100 (April 2026), source.
38. 38. Mississippi Development Authority, “Tech Leader xAI Investing More Than $20 Billion in Southaven” (January 8, 2026), source; DeSoto County Board of Supervisors, March 2, 2026 agenda, p. 4; Virginia Department of Taxation and Virginia Economic Development Partnership, Biennial Data Center Retail Sales and Use Tax Exemption Report (January 2, 2026), PDF, report pp. 3–7; Carl Vinson Institute of Government, University of Georgia, for the Georgia Department of Audits and Accounts, Tax Incentive Evaluation: Georgia Data Center Sales & Use Tax Exemption (December 2025), PDF, pp. 3–7, 38–39; Washington Joint Legislative Audit and Review Committee, “Data Centers in Urban Counties,” preliminary report (July 2026), source.
39. 39. Ari Peskoe and Eliza Martin, Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power (Harvard Environmental and Energy Law Program, March 2025), pp. 25, 30–32, PDF.
40. 40. Memphis Community Against Pollution, “Take Action,” campaign page; Ashli Blow, “Memphis Leaders Celebrate xAI, but Will Its ‘Burden’ Go Unchecked?” MLK50: Justice Through Journalism (July 22, 2024), source; Katherine Burgess, “Inside the Memphis Chamber of Commerce’s Push for Elon Musk’s xAI Data Center,” MLK50 (August 22, 2025), source; Michael Finch II, “Southwest Memphis Air Monitors Reveal ‘No Relief’ from Elevated Pollution, Researchers Find,” MLK50 (June 9, 2026), source.
41. 41. Data Center Watch, “Q1 2026: Data Center Watch Report,” public summary; the public summary does not include the project-level dataset. New York Governor Kathy Hochul, “First Statewide Moratorium on New Hyperscale Data Centers Launched by Governor Kathy Hochul” (July 14, 2026), source.
42. 42. Brockovich AI Data Center Reporting, community issue map, FAQ, and terms. The volunteer-run initiative says its map is built from community reports and can include operating, proposed, rumored, and under-construction projects. It describes entries as an awareness starting point that may be approximate, incomplete, or outdated.