Skip to content

Environmental & Health Impacts of AI

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Every prompt has a physical footprint. The chatbots, copilots, and image generators that feel weightless on screen run on warehouse-scale data centers — buildings packed with power-hungry chips that have to be manufactured, powered, and cooled around the clock. As the AI build-out accelerates, the bill is coming due in two currencies: the natural environment (electricity, water, land, and materials) and human health (the air pollution, heat, and disruption borne by the communities living next to the infrastructure).

This lesson surveys both — and the trade-offs every AI user and institution should weigh.

The short version

  • AI is driving the fastest surge in electricity demand in a generation: global data-center power use is projected to more than double to ~945 TWh by 2030 — roughly 3% of the world's electricity, about what Japan consumes today.
  • Meeting that demand is delaying the retirement of coal and gas plants, spurring a wave of new gas turbines, and reopening nuclear reactors — while Big Tech's own emissions rise despite net-zero pledges.
  • Data centers consume water — directly for cooling and indirectly through the power plants feeding them — increasingly in drought-stressed regions.
  • The pollution carries a measurable human toll: U.S. data-center air pollution is projected to cause on the order of 1,300 premature deaths and ~$20 billion in health damages per year by 2030, concentrated in the low-income and minority communities sited next to the turbines.

A footprint you can't see on screen

The mental model that AI is "just software" hides a heavy industrial reality. Behind every model are three physical demands that grow with use:

  • Compute — racks of GPUs/TPUs that must be manufactured (mining, chip fabrication) and eventually discarded.
  • Energy — electricity to run the chips, around the clock, at very high power density.
  • Cooling — water and/or still more electricity to carry away the heat those chips produce.

Training a frontier model is expensive, but the larger and growing cost is inference — answering billions of everyday queries. A single AI-generated answer can use several times the energy of a traditional web search, and that small per-query cost multiplies across billions of prompts a day.

Energy demand and the grid

How much electricity?

The International Energy Agency's Energy and AI report projects that global data-center electricity use will reach about 945 TWh by 2030 — roughly 3% of all electricity worldwide and more than double 2024 levels.

  • Data-center demand grows about 15% per year through 2030 — four times faster than total electricity demand from every other sector combined.
  • The growth is AI-specific: electricity for "accelerated servers" (the GPU clusters that run AI) climbs ~30% per year.
  • In the United States alone, data centers add about 240 TWh by 2030 (a 130% increase).
  • In the IEA's higher-growth scenario, global data-center demand exceeds 1,700 TWh by 2035 (~4.4% of world electricity).

Where that power comes from — and who pays

A demand spike this fast outruns the clean-energy build-out, with knock-on effects:

  • Fossil lock-in. Utilities are postponing the retirement of coal and gas plants and fast-tracking new gas turbines to serve data-center load — slowing, not speeding, the energy transition.
  • A nuclear scramble. Tech firms are signing deals to reopen reactors (Microsoft and Constellation's Three Mile Island restart) and to build small modular reactors (Google–Kairos, Amazon, Meta) — but new nuclear arrives slowly and won't cover near-term demand.
  • Rising emissions. Despite net-zero pledges, the largest AI firms report growing greenhouse-gas emissions — Google's were up roughly 48% and Microsoft's nearly 30% against their baselines as the build-out accelerated.
  • Cost-shifting onto households. Grid upgrades and generating capacity to serve hyperscale campuses can raise electricity bills for ordinary ratepayers, who effectively subsidize the build-out.

The efficiency paradox

Chips and data centers keep getting more efficient per computation — but demand is growing faster than efficiency, so total energy use keeps climbing. The savings are spent on doing much more AI, not on using less power: a classic Jevons paradox.

Impacts on the natural environment

Water

Data centers are thirsty. They consume fresh water two ways: directly, through evaporative cooling that boils off water to shed heat, and indirectly, through the water-cooled power plants that supply their electricity (often 80% or more of the total). The UC Riverside study Making AI Less "Thirsty" estimated that:

  • A short ChatGPT exchange of 10–50 questions can consume roughly 500 ml of water (a 16-oz bottle) once cooling and the regional power mix are counted.
  • Training GPT-3 in U.S. data centers consumed on the order of 5.4 million liters of water.

The deeper problem is where this happens: hyperscale campuses are frequently sited in hot, water-stressed regions (the U.S. Southwest, Chile, Spain), putting them in direct competition with farms and households for scarce fresh water.

Land, materials, and electronic waste

  • Construction & land. Each campus is a large industrial footprint — concrete, steel, and graded land (with embodied carbon and habitat loss) plus substations and transmission corridors.
  • Materials. The chips depend on mined silicon, copper, and rare-earth elements and on water- and energy-intensive semiconductor fabrication.
  • E-waste. AI hardware is replaced on a fast cycle. A 2024 Nature Computational Science analysis, Modeling the increase of electronic waste due to generative AI, estimated generative AI could add a cumulative 1.2–5.0 million tonnes of e-waste between 2020 and 2030 — much of it laden with lead and other toxics — though circular-economy strategies could cut that by 16–86%.

Impacts on human health

The costs above are not abstract — they land on human bodies, and not evenly.

Air pollution and its body count

To meet deadlines and bridge grid shortfalls, data centers lean on on-site fossil generation: diesel backup generators and, increasingly, gas turbines. These emit fine particulate matter (PM2.5), nitrogen oxides (NOₓ), and sulfur dioxide — pollutants linked to asthma, heart disease, lung cancer, and premature death. The 2024 UC Riverside–Caltech report The Unpaid Toll quantified the U.S. health burden using EPA methods:

  • Approximately 1,300 premature deaths per year by 2030.
  • About 600,000 asthma-symptom cases.
  • Total public-health costs approaching ~$20 billion per year — a hidden subsidy paid in clinic visits, missed school, and lost lives.

Environmental justice: who lives next to the turbines

Pollution and water stress are disproportionately sited in low-income communities and communities of color — the same environmental-justice pattern as older heavy industry.

Case study — xAI's Colossus, Memphis

Elon Musk's xAI powered its Colossus supercomputer in Boxtown, a historically Black neighborhood of South Memphis, by running a fleet of methane gas turbines — for a time without the required Clean Air Act permits. As the build-out expanded toward a second site in Southaven, Mississippi, residents and regulators counted dozens of unpermitted turbines capable of emitting on the order of 2,500 tons of NOₓ a year — likely the single largest industrial source of smog-forming pollution in greater Memphis, an area that already fails federal smog standards.

In 2026 the NAACP, represented by Earthjustice and the Southern Environmental Law Center, sued xAI for Clean Air Act violations — seeking to halt the unpermitted turbines and force best-available pollution controls. It is among the first major legal tests of who bears the health cost of the AI build-out.

Heat, noise, and local stress

Beyond air and water, neighbors of a large data center contend with the constant low-frequency hum of cooling systems and generators, waste heat, and competition for local water and grid capacity — quality-of-life and health stressors that rarely appear in a model's "cost."

What responsible use looks like

The footprint is real, but it is not a reason to abandon AI — it is a reason to use it deliberately and to demand accountability.

  • Right-size the model. Use the smallest model that does the job; reserve frontier models for tasks that need them. Don't fire off a giant model for a one-line answer.
  • Batch and reuse. Cache and reuse results; avoid needless re-runs and "let me regenerate that ten more times" habits.
  • Favor accountable providers. Prefer vendors that disclose energy, water, and carbon per workload and that match clean energy on the same grid and hour — not just buy distant offsets.
  • Ask where the campus is. Support siting that uses recycled/non-potable water and closed-loop cooling, avoids water-stressed basins, and does not concentrate pollution in already-overburdened communities.
  • Push for transparency and regulation. Disclosure standards, honest accounting, and real permitting — rather than the fast-tracked federal permitting carve-outs for data centers — are what turn "use a smaller model" from a personal gesture into structural change.

Is 'use a smaller model when you can' enough?

A recurring debate in AI ethics: is individual restraint a meaningful environmental ethic, or a personal-virtue dodge that lets the system off the hook — the AI equivalent of "use less plastic"? Both can be true. Personal choices matter at the margin; disclosure, clean-energy siting, and enforceable permits are what move the needle at scale.

Further reading

Assessment

Name the two ways a data center consumes water, and which is usually larger.
On-site cooling vs. the power supply

Direct (on-site evaporative cooling) and indirect (water used by the power plants generating its electricity). The indirect share is usually larger — often 80% or more of total water use — so a data center's water footprint depends heavily on how clean and water-efficient its grid is.

Why is the health burden of AI data centers an environmental-justice issue, not just an environmental one?
Success

The gas turbines and diesel generators that bridge grid shortfalls emit PM2.5 and NOₓ, and these facilities are disproportionately sited in low-income communities and communities of color (e.g., xAI's turbines in Boxtown, South Memphis). Those neighbors breathe the pollution and bear the asthma, heart-disease, and premature-death costs, while the benefits of the compute accrue elsewhere.

True or False: because chips keep getting more efficient, AI's total energy use is falling.
False

Efficiency per computation is improving, but total demand is rising faster because we keep doing far more AI — a Jevons paradox. The IEA projects data-center electricity use to more than double by 2030.

Ethics of AI