Environmental and Resource Externalities of AI

Energy Consumption: The Scale of AI's Appetite

Data Center Electricity Demand

In 2023, U.S. data centers consumed 4.4% of total U.S. electricity (176 TWh). By 2026, that figure is anticipated to reach 6% (260 TWh). By 2028, projections range between 325 and 580 TWh (6.7-12.0%) of total U.S. electricity. Globally, data center electricity consumption is expected to approach 1,050 TWh by 2026, placing data centers 5th in global electricity consumption—between Japan and Russia as sovereign nations.

Regional concentration is extreme. Data centers consumed 26% of total electricity in Virginia in 2023, 15% in North Dakota, 12% in Nebraska, and 11% in both Iowa and Oregon. AI's share of data center power is currently 5-15% but is projected to reach 35-50% by 2030, with one forecast predicting AI will consume over half of all data center electricity by 2028.

Per-Query Energy Costs

A single ChatGPT query using GPT-4o consumes roughly 0.0026 kWh—approximately 9 times more energy than a Google Search at 0.0003 kWh. Training GPT-3 consumed approximately 1,287 megawatt-hours of electricity, but for deployed AI services, inference costs dominate: one year of inference for a ChatGPT-like service produces 25 times the carbon emissions of the model's initial training. With over 700 million weekly ChatGPT users as of 2025, the aggregate energy cost of inference dwarfs training by orders of magnitude.

Grid Infrastructure Crisis

Goldman Sachs estimates approximately $720 billion is needed for U.S. grid upgrades through 2030 to accommodate data center growth. Total U.S. data center capacity was approximately 25-35 GW in 2024. Over 8.9 GW across 105 projects target operation by end of 2026, with 47 already under construction. Goldman Sachs expects demand to grow to approximately 92 GW by 2027 and 219 GW by 2030. The growth rate is approximately 17% yearly between 2025 and 2028, placing unprecedented strain on a grid that was not built for concentrated industrial loads of this magnitude.

Water Consumption: The Hidden Resource Crisis

Corporate Water Footprints

Google consumed 6.4 billion gallons (24.2 billion liters) globally in 2023, with 95% used by data centers. Google's Council Bluffs, Iowa facility alone consumed 1 billion gallons (3.8 billion liters) in 2024—the most of any single data center. Microsoft consumed approximately 1.69 billion gallons, a 34% increase from the previous year, with 42% drawn from areas experiencing water stress. Meta consumed 813 million gallons (3.1 billion liters), with 95% used by data centers.

Meta's Newton County, Georgia facility uses approximately 500,000 gallons per day—roughly 10% of the small community's total water supply. Newton County is on track for a water deficit by 2030. Nearby residents reported their well water taps running dry after Meta built next door, spending $5,000+ on water problems, with well replacement costs reaching $25,000.

Projections

AI systems' water footprint could reach 312.5 to 764.6 billion liters in 2025. Morgan Stanley predicts global data center water use will rise 11 times to 1,068 billion liters per year by 2028. Training GPT-3 alone directly consumed an estimated 700,000 liters of freshwater. A single ChatGPT conversation (10-50 queries) consumes approximately 500ml of water. Two-thirds of new U.S. data centers are being built in high water-stress areas, including Arizona, Nevada, and Texas, placing tech industry water consumption in direct competition with agriculture and residential needs.

Microsoft's Zero-Water Cooling

Microsoft introduced zero-water cooling designs starting August 2024 that could save 33 million gallons per data center annually. While a meaningful engineering advance, it addresses a fraction of the industry's water footprint and does not retrofit existing facilities. The broader industry has not adopted similar designs, and the expansion rate of conventional water-cooled data centers far outpaces the deployment of zero-water alternatives.

The Nuclear Revival: Powering AI with Atoms

Microsoft and Three Mile Island

Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Unit 1 of the Three Mile Island Nuclear Generating Station in Pennsylvania—the first time a retired American nuclear plant is being brought back to life for a single commercial customer. The reactor will provide 835 MW of carbon-free electricity, enough to power approximately 800,000 homes. The U.S. Department of Energy closed a $1 billion federal loan to Constellation Energy, with total project costs estimated at $1.6 billion. The reactor is expected to return to service in 2027, approximately one year ahead of the original schedule.

Amazon's Small Modular Reactor Fleet

Amazon plans 12 small modular reactors in Washington State built by Energy Northwest using X-energy reactor designs. Phase 1 provides 320 MW capacity; full buildout reaches up to 960 MW. Amazon invested $500 million (Series C) in X-energy. In Virginia, Amazon signed an agreement with Dominion Energy to develop SMRs near the existing North Anna nuclear power station, bringing at least 300 MW to a region where Dominion projects power demands will increase by 85% over the next 15 years. Amazon and X-energy also signed agreements with South Korean partners Doosan Enerbility and Korea Hydro & Nuclear Power to deploy more than 5 gigawatts of new nuclear energy across the U.S. by 2039.

Google and Kairos Power

Google signed a Master Plant Development Agreement with Kairos Power for a fleet of advanced nuclear projects totaling 500 MW by 2035. The Hermes 2 reactor in Oak Ridge, Tennessee—built through a PPA between Kairos Power and the Tennessee Valley Authority—increased output from 28 MW to 50 MW from a single reactor, scheduled to begin operations in 2030. In 2025, Kairos Power commenced safety-related construction of the Hermes demonstration reactor—the first non-water-cooled reactor approved for construction in the U.S. in over 50 years.

Carbon Emissions: The Climate Contradiction

Industry-Wide Emissions

AI data centers generated 105 million metric tons of CO2 in the 12 months ending August 2026, accounting for 2.18% of national emissions and surpassing the aviation industry's carbon footprint. A January 2026 study estimated data centers emitted between 32.6 million and 79.7 million tonnes of CO2 in 2025—roughly equivalent to the annual emissions of a small European country. Goldman Sachs forecasts that about 60% of increasing data center electricity demand will be met by burning fossil fuels, increasing global carbon emissions by about 220 million tons.

Company-Level Failures

Google's emissions rose 11% in 2024 to 11.5 million tonnes—a 51% increase from 2019 levels. Electricity consumption increased 27% in 2024 to support AI workloads. Microsoft's emissions increased almost 30% since announcing its carbon-negative goal, largely because of AI data center buildouts. Microsoft acknowledges it will not meet its 2030 carbon-negative goal on the original timeline. Training GPT-3 alone emitted roughly 500 metric tons of CO2—equivalent to driving a car from New York to San Francisco approximately 438 times.

The Greenwashing Report

A February 2026 report by climate analyst Ketan Joshi, commissioned by Beyond Fossil Fuels, Friends of the Earth U.S., Stand.earth, and the Green Web Foundation, found that 74% of industry claims about AI's climate benefits are unproven. Not a single example was found where consumer generative AI systems led to material, verifiable, substantial emissions reductions. Only 26% of claims cited published academic papers; 36% cited no evidence at all. Companies routinely blur differences between generative AI (high environmental cost) and traditional machine learning (lower footprint). Microsoft announced it will stop buying "non-additive, unbundled renewable energy certificates"—an implicit admission that the industry's primary green energy accounting method was effectively meaningless.

Community Opposition: The Backlash

The Scale of Resistance

In Q2 2025 alone, 20 data center projects were blocked or delayed, affecting $98 billion in potential investment—more than all disruptions tracked since 2023. Between 2023 and late March 2025, $64 billion worth of projects were blocked or delayed. At least 25 data center projects were canceled in 2025 following local opposition; up to 25 more were canceled in the first three weeks of 2026. Opposition groups were successful in blocking or delaying two out of every three projects they protested. Nearly 200 community groups across more than two dozen states are now actively opposing data center projects.

Specific Conflicts

In Richmond, Virginia, DC Blox withdrew its $500 million proposal after community opposition. In Maryland, a proposal to convert an abandoned mall sparked 20,000 petition signatures. In Tucson, Arizona, council members unanimously discontinued discussions on Project Blue. Prince William County, Virginia approved the "Digital Gateway" in December 2023—the largest data center corridor in the world with over 22 million square feet across 2,100+ acres. Northern Virginia now has over 10,000 diesel generators surrounding data centers. Loudoun County, Virginia amended its zoning ordinance in March 2025 to prevent new data centers from being approved on a by-right basis. Virginia state legislator Sen. Danica Roem is pushing to limit data centers to industrial areas and remove their sales tax exemption.

Public Opinion

Only 44% of Americans would welcome a data center nearby—less popular than gas plants, wind farms, or even nuclear facilities. Water use is the No. 1 reason cited for local opposition, mentioned for more than 40% of contested projects.

Public Health Consequences

Diesel Generator Pollution

Data center backup diesel generators emit nitrogen oxides, fine particulate matter (PM2.5), carbon monoxide, and sulfur dioxide. In Virginia, data centers' backup generators could release 9,000 tons of nitrogen oxides—equal to about half of what has typically been emitted annually in Northern Virginia by all sources. Northern Virginia has over 10,000 diesel generators at data center sites, producing 20 GW of backup power capacity.

Projected Health Costs

By 2030, U.S. data centers could contribute to nearly 1,300 deaths annually, resulting in a public health burden of more than $20 billion. In California, health costs attributed to data center facilities tripled from about $45 billion to $155 billion between 2019 and 2023. Continuous noise exposure from generators and cooling systems is linked to hypertension, cardiovascular disease, tinnitus, type 2 diabetes, cognitive impairment, and sleep disturbance. Generators can reach 100 dB(A)—loud enough to cause hearing damage with prolonged exposure.

Environmental Justice

Health threats are particularly severe in low-income communities historically exposed to environmental pollutants. Federal Opportunity Zone policies have incentivized data center construction in these communities, concentrating pollution in areas with the least political power to resist. Microsoft has committed to phasing out petroleum-based diesel generators by 2030, but the rest of the industry has made no such commitment.

Semiconductor Manufacturing: The Upstream Footprint

TSMC's Water Consumption

TSMC consumed 101 million cubic meters of water in 2023. Semiconductor production can require up to 10 million gallons of ultrapure water per day. TSMC's Arizona fab uses 4.75 million gallons per day; with water reclamation at 90% recycling, demand drops to fewer than 1.2 million gallons per day. TSMC missed its target of reducing unit water consumption by 2.7%, instead increasing it by 25.2% in 2023. Water usage across semiconductor manufacturing is forecast to double by 2035.

Chemical Contamination

PFAS ("forever chemicals") remain essential in the semiconductor manufacturing process. The industry has begun phasing out PFOS and PFOA, but PFAS have been found in fab wastewater discharges in some states. TSMC's Taichung Zero Waste Manufacturing Center can reduce outsourced waste processing by 130,000 metric tons per year, but this represents an incremental improvement against a rapidly growing waste stream.

GPU Embodied Emissions

An NVIDIA H100 baseboard (8 cards) carries 1,312 kg CO2e of embodied emissions—approximately 164 kg CO2e per card. The breakdown: high bandwidth memory 42%, integrated circuits 25%, thermal components 18%. NVIDIA's B200 achieved a 24% reduction in embodied carbon emissions across large workloads compared to H100, reducing carbon intensity from 0.66 to 0.50 gCO2e per exaflop. But production volumes are increasing far faster than per-unit efficiency gains.

E-Waste: The Downstream Crisis

AI-Specific E-Waste Projections

Global e-waste reached 62 million tonnes in 2022 (82% rise from 2010), projected to reach 82 million tonnes annually by 2030. AI is accelerating this trend: projections for additional AI-specific e-waste by 2030 range from a conservative 1.2 million tonnes to an upper bound of 16.1 million tonnes if deployment accelerates at the upper end. A Nature Computational Science study projected up to 5 million tonnes of additional AI e-waste by 2030.

Hardware Obsolescence Cycles

AI-specific hardware becomes obsolete in 2-3 years, compared to 5-7 years for traditional servers. Major cloud providers refresh AI training clusters every 12-18 months, compared to traditional 3-5 year lifecycles. Between 20 and 70 million hard disk drives in the U.S. reach end of life each year. Only 17-22% of global e-waste is documented to be collected and properly recycled.

E-Waste Colonialism

Much of this e-waste is shipped to the Global South—Ghana, Nigeria, India—for "recycling" under hazardous conditions. Workers, often children, extract valuable metals using dangerous methods without proper protection. AI hardware requires cobalt, lithium, and rare earth metals; mining these materials causes environmental destruction and human rights abuses, primarily in African nations and China. A full life cycle assessment of generative AI services found that end-user terminals and networks account for 85% of metal and mineral consumption and 45% of the carbon footprint—costs borne by consumers, not AI companies. Extending hardware lifespan, refurbishing, and redesigning for recyclability could reduce e-waste generation by up to 86% in a best-case scenario, but no major AI company has committed to such measures.