Back

ACC Education Center • Updated August 28, 2026

AI's physical footprint is growing faster than the rules around it.

The buildout is driving new electricity demand, water pressure, material use, and community conflict. AI can also improve forecasting and cleaner systems. This guide separates what is known, what is projected, and what responsible infrastructure requires.

Education CenterAI and Data Centers
Last reviewed: August 28, 2026Personal use, energy, water, materials, communities, and solutions28-32 minute read

The issue in one minute

The AI boom is now an infrastructure story.

AI services run through physical facilities filled with specialized chips. Those facilities need power every hour, shed heat, draw on water and supply chains, and connect to local grids and communities. The impact is not one universal number per prompt. It changes with the model, hardware, location, cooling system, power source, and time of day.

The newest evidence points in three directions at once: ordinary text prompts can be small per use, heavier tasks can require far more energy, and total infrastructure demand is rising quickly. Measurement standards, public reporting, efficiency methods, and community protections are also becoming more concrete.

Field File 003 research • Personal AI use

Does one person's AI use really matter? The honest answer is more useful than either guilt or dismissal.

Read the evidence →

Field File 003 • Personal AI use

One person's text prompts can be small. “Irrelevant” is still too strong.

For a person who mainly drafts, summarizes, brainstorms, or asks questions in text, direct electricity use can be modest. That does not make every AI task small, and it does not make the combined infrastructure footprint disappear.

The proposed claim needed one important correction: the amount of data someone moves through the internet is not a reliable measure of environmental impact. Bytes describe traffic. Watt-hours describe electricity. Emissions and water depend on the power supply, location, cooling, hardware, provider, task, output length, and measurement boundary.

ACC evidence verdict

The idea is partly right, but the strongest wording outruns the evidence.

Supported with caveats
Routine text-based AI can use little electricity per prompt, and efficiency per task is improving quickly.
Misleading as stated
“Infinitesimal and irrelevant compared with everyday data use” compares unlike units and treats every AI task as if it had the same footprint.
Not knowable from a username or bill
A person's exact footprint cannot be calculated honestly without the provider, model, task mix, output length, location, and lifecycle boundary.

A transparent text-only scenario

This illustration applies Google's company-reported full-stack median of 0.24 watt-hours for one Gemini Apps text prompt in May 2025. The television comparisons use the same study's 100-watt reference. The arithmetic is transparent; the result is not a universal AI average.

One median text prompt0.24 Wh

Less electricity than about nine seconds of a 100-watt television in this Google measurement.

20 text prompts in a day4.8 Wh

About three minutes of a 100-watt television. This is an ACC calculation from Google's per-prompt result.

100 text prompts in a day24 Wh

About 14 minutes of a 100-watt television. It still describes text prompts only.

This is a scale illustration, not an AI footprint calculator.

It does not describe every provider, ChatGPT or another named service, model training, user-device electricity, storage, image generation, video generation, deep reasoning, agents, or a full business workflow. Google's authors also show why published estimates can diverge when studies include different parts of the serving system.

Review the Google measurement →
Use the right unit.

Bytes measure data volume. Watt-hours measure electricity. Grams of CO2-equivalent measure climate emissions. Liters or milliliters measure water. A responsible comparison keeps the task, provider, period, location, and lifecycle boundary visible.

Task type can change the answer

Usually lighter

Simple text

Short drafting, summarizing, classification, and question-answering can be relatively low-energy when an efficient model fits the task.

Potentially much heavier

Reasoning and agents

The IEA reports that some reasoning and agentic tasks can use hundreds or thousands of times more energy per query than simple text.

Compute-intensive media

Images and video

Generative image work is generally more energy-intensive than text in measured model comparisons. Video generation is among the high-energy uses highlighted by the IEA.

A prompt count alone can hide the biggest difference.

One hundred short text prompts and one hundred long reasoning, agentic, image, or video jobs are not equivalent workloads. The International Energy Agency identifies efficiency gains, adoption, and the changing mix of tasks as the three central drivers of AI electricity demand.

Review the IEA findings →

Why published numbers differ

Two disclosures can be useful without being directly comparable.

Google reported a median Gemini Apps text prompt using 0.24 Wh, 0.03 gCO2e, and 0.26 mL of water in May 2025. Mistral reported 1.14 gCO2e and 45 mL of water for a 400-token Le Chat response in its lifecycle study. The products, periods, power systems, and accounting boundaries differ, so this is not a provider ranking.

Review the Mistral lifecycle study →
  • GoogleProduction measurement for a provider-specific median text prompt; includes serving equipment, provisioned idle capacity, data center overhead, and reported emissions and cooling water.
  • MistralProvider-specific 400-token response; includes upstream impacts such as server manufacturing and excludes the user's terminal.

The practical answer is proportion, not panic

What one person can do
  • Use ordinary software when it already solves the problem.
  • Default to text and a fit-for-purpose model for routine work.
  • Ask for only the length and detail the job needs.
  • Reuse strong outputs instead of regenerating them.
  • Reserve long reasoning, agents, images, and video for work that benefits from them.
What the system must do
  • Publish comparable energy, water, emissions, and lifecycle data.
  • Make efficient modes and smaller fit-for-purpose models easy to choose.
  • Use cleaner electricity and water-smart cooling without hiding tradeoffs.
  • Prevent infrastructure costs and local impacts from being shifted onto communities.
  • Report total growth alongside efficiency per task.

ACC conclusion

Use AI intentionally, not anxiously. Keep personal use in perspective and keep asking providers for the missing evidence.

See responsible solutions →

The physical buildout

AI has become a utility-scale planning issue.

A single prompt is one useful lens for personal use, but it is not the whole system. The larger issue is the combined growth of model training, everyday inference, cloud services, specialized chips, cooling systems, and new data center campuses.

The International Energy Agency estimates that global data center electricity demand reached 485 terawatt-hours in 2025. It also estimates that AI-focused data center electricity use grew 50% that year. Its updated outlook reaches roughly 950 TWh by 2030, close to 3% of global electricity use. The global share is still smaller than many major sectors, but data centers are concentrated. A single region can feel intense grid, water, land, and cost pressure long before the worldwide percentage looks dramatic.

February 2026A lifecycle standard arrives

ITU-T L.1801 sets guidance for measuring raw materials, production, training, inference, use, and end-of-life impacts.

Review the standard →
June 2026A new U.S. bottom-up forecast

Lawrence Berkeley National Laboratory estimates a 649 TWh reference case for 2030, with a 521-843 TWh uncertainty range.

Review the report →
July 28, 2026 guidanceEU reporting becomes more usable

The European Commission added public guidance for its energy-and-water database while continuing work on ratings and minimum performance standards.

Review EU reporting →
July 2026Lifecycle tradeoffs sharpen

A new review finds that embodied emissions can exceed half of emissions for large AI data centers and that water and carbon goals can conflict.

Read the review →
July 28, 2026 • working paperFlexible AI load gets a planning framework

A University of Chicago analysis uses 49.4 million production inference requests to estimate how much load mixed-use and inference-heavy facilities could reliably reduce.

Review the working paper →

Read the numbers carefully

Four distinctions keep the evidence honest.

All data centers vs. AI-focused facilities
Many reports measure the whole data center sector. AI is a growing part of that total, not the entire total.
Estimate vs. projection
An estimate describes a recent year using available data. A projection models a possible future under stated assumptions.
Operational vs. embodied impact
Operational impact comes from running and cooling equipment. Embodied impact comes from chips, concrete, steel, construction, and replacement.
Direct vs. indirect water
Direct water is used at a facility. Indirect water is used elsewhere to generate electricity and manufacture equipment.

Electricity and the grid

Fast growth matters most where the grid is already tight.

Electricity use is measured over time. Grid operators must also meet the highest moment of demand. AI campuses create challenges in both places: they can consume enormous amounts of annual energy and request very large, dense, always-available connections.

Global data center electricity demandIEA estimates for 2024 and 2025; 2030 is a projection.
2024 estimate
415 TWh
2025 estimate
485 TWh
2030 projection
~950 TWh

One terawatt-hour is one billion kilowatt-hours. The chart compares the scale of the sector; it does not claim that all data center electricity is used by AI.

Efficiency per task is improving, but total demand is still rising.

The IEA estimates that electricity per AI task has fallen by at least tenfold each year in recent years. At the same time, adoption is expanding and video, reasoning, and agentic tasks can use hundreds or thousands of times more energy than simple text. A more efficient task does not guarantee a smaller system-wide footprint when the number and intensity of tasks grow faster.

Lawrence Berkeley National Laboratory's June 2026 bottom-up model places all U.S. data centers, not AI alone, at a 649 TWh reference case in 2030. Its compounded uncertainty range is 521-843 TWh, equal to 9.5% to 15.3% of U.S. electricity. These are planning cases, not promises. They depend on equipment shipments, device power, cooling, facility locations, utilization, and AI-chip operating life.

Lower bound521 TWh

About 9.5% of U.S. electricity.

Reference case649 TWh

About 11.8% of U.S. electricity.

Upper bound843 TWh

About 15.3% of U.S. electricity.

Why a local project can feel larger than the national percentage

Demand is concentrated near specific substations, transmission lines, and power plants. EPRI reports that data centers already use more than one-fifth of Virginia's electricity and could reach 39% to 57% by 2030. The U.S. Department of Energy also reports that synchronized AI training can create fast, repetitive load oscillations that require better grid monitoring.

Responsible planning therefore asks more than, "Is there enough annual generation?" It asks whether capacity exists in the right place and hour, whether the project can reduce load during grid stress, what new generation is added, and who pays for transmission, substations, backup power, and reliability.

From promise to test

Some AI compute can move when the grid is stressed.

EPRI's DCFlex program is running real-world demonstrations of workload modulation, demand response, and backup-power integration. A Phoenix preprint reports that software orchestration reduced power on a 256-GPU cluster by 25% for three hours while maintaining priority-service guarantees. A July 2026 University of Chicago working paper reaches a similar direction using four years of grid prices and 49.4 million production inference requests. These are bounded studies, not proof that every workload or facility is equally flexible.

  • 25% for 3 hoursReported cluster-power reduction in the Phoenix field demonstration.
  • 24.6-40.0%Estimated dependable reduction for single mixed-use and inference-dominant facilities in the working paper.
  • At least 10Field demonstrations planned by EPRI across different sites, grid structures, and technologies.
Flexible load is not automatically low-carbon load.

A peer-reviewed 2026 power-system study found that shifting data center demand lowered modeled system costs by up to 2% to 5% across three U.S. regions. Emissions fell where flexible demand aligned with abundant renewables, but rose by 3% in one Mid-Atlantic scenario because coal ran more steadily. The fix is to respond to grid stress and marginal emissions together, paired with genuinely additional clean supply.

Read the iScience study →

Water, chips, and materials

Water impact is local, while the supply chain is global.

There is no honest universal water number for an AI prompt. Water use changes with the facility, weather, cooling design, electricity source, model, hardware, and time of operation.

A 2026 PLOS Water analysis argues that data center water impacts should distinguish withdrawals from consumption and should be evaluated in the watershed where they occur. A liter used in a water-abundant region is not equivalent to a liter consumed during drought. Electricity can also carry an indirect water footprint depending on how it is generated.

At the facility

Cooling and operations

Evaporative cooling may lower electricity use but consume water. Dry or closed-loop systems can reduce onsite water while changing energy needs.

Beyond the facility

Power and manufacturing

Power plants, chip fabrication, equipment production, and construction can use water far from the final data center.

Across the lifecycle

Chips, steel, and concrete

Embodied emissions and resource depletion remain hidden when reporting covers only the electricity used after a facility opens.

Modeled U.S. AI-server footprint

Indirect impacts can be larger than the water used onsite.

A 2025 Nature Sustainability analysis modeled U.S. AI-server growth from 2024 through 2030. These are scenario results, not measured national totals, and the authors report substantial uncertainty from growth, efficiency, grid changes, and facility locations.

Read the analysis →
  • 71%of the modeled water footprint is indirect, mainly through electricity; direct facility water is 29%.
  • 24-44million metric tons CO2e per year across the study's expansion scenarios.
  • 731-1,125million cubic meters of water per year across the same modeled scenarios.

Recent research

Operational efficiency is only part of the answer.

A July 2026 review in Nature Reviews Clean Technology brings hardware, construction, inference, electricity, carbon, and water into one lifecycle view.

Read the review →
  • More than halfEmbodied emissions can exceed half of emissions from large AI data centers.
  • 40-60%Inference may account for this share of a model's lifetime climate emissions.
  • 10-20%Recycled or older components may reduce overall data center emissions by this range.
  • About 10%Grid-aligned workload management may reduce carbon on grids with high renewable generation.
A water-saving choice can raise electricity use, and an energy-saving choice can raise water use.

The best design evaluates both at once, uses local water-stress data, and reports what happens at the facility and in the electricity supply. Replenishment projects can help a watershed, but they are not the same thing as reducing a facility's withdrawal at the moment and place of use.

Communities and rules

Local concern is not anti-technology. It is a demand for a fair agreement.

Data centers can bring construction work, tax revenue, digital infrastructure, and new power investment. They can also bring land conversion, noise, water demand, backup generators, transmission lines, tax incentives, and utility costs that outlast the construction jobs.

Why some communities support projects
  • Construction jobs and local contracts
  • Property and business tax revenue
  • Grid and broadband investment
  • Potential clean-energy procurement
Why some communities oppose projects
  • Water, power, land, and air concerns
  • Noise, traffic, and quality-of-life effects
  • Fear that households will absorb grid costs
  • Few permanent jobs relative to facility scale

Gallup reported in May 2026 that 70% of Americans would oppose an AI data center in their area, including 48% who would strongly oppose it. The most common objections concerned resource use, pollution, bills, and quality of life. Supporters most often pointed to jobs, economic activity, and tax revenue. That split makes public evidence more important, not less.

Rules are beginning to catch up

Lifecycle measurement

ITU-T L.1801 calls for assessment across raw materials, production, training, inference, use, and end-of-life, with uncertainty stated clearly.

Energy and water reporting

The European Union's database collects and publishes energy-performance and water-footprint information from facilities with significant energy use.

Ratings and minimum standards

The European Commission is preparing a rating scheme and beginning work on minimum performance standards.

U.S. federal grid rules are targeting cost shifting and inflexible connections.

In June 2026, the Federal Energy Regulatory Commission directed all six regional grid operators under its jurisdiction to justify or reform their large-load rules. The proceedings call for transparent studies, protections against shifting transmission costs to other customers, and new service options for loads able to reduce grid withdrawals. These are active regulatory proceedings, not finished nationwide tariffs.

Review the FERC action →
U.S. siting rules are also becoming more specific.

Virginia's 2026 law requires a sound assessment before local approval of a new facility expected to need at least 100 megawatts. Localities may also require review of water, agriculture, parks, historic sites, and forestland. That does not settle whether a project should be built, but it moves concrete evidence earlier in the decision.

Read the Virginia law →

The fair position is neither automatic approval nor automatic rejection. It is informed consent: public data before approval, enforceable operating limits, fair cost allocation, meaningful community benefits, and annual reporting after a facility opens.

What companies disclose

Transparency and footprint are two different questions.

A large company may publish more environmental data while still operating a much larger footprint. A smaller company may use less infrastructure but reveal too little to evaluate it. The comparison below measures the kind of public evidence available, not whether a company is "good" or "bad."

How to read this snapshot

Model-level, company-level, and project-level claims are not directly interchangeable. All company figures below are self-reported unless a source says otherwise. Replenishment, renewable matching, avoided emissions, and operational efficiency each describe different things.

Mistral AI

Model-level lifecycle study

Published estimates for training and inference, including carbon, water, and resource depletion. For Mistral Large 2 after 18 months of use, it reported 20.4 ktCO2e and 281,000 cubic meters of water. The detail is unusual, but it covers one company's methods and is not a universal per-prompt benchmark.

Google

Broad company and data center reporting

Reports energy, emissions, water, waste, and clean-energy procurement. Its 2026 report says it replenished 78% of 2025 freshwater consumption and contracted 12 GW of clean energy in 2025. Replenishment does not mean the same water remained available at every facility and hour.

Amazon / AWS

Infrastructure water and efficiency reporting

Reports that global leased, owned, and shared data centers withdrew 9.4 billion liters in 2025. Separately, it says replenishment projects returned 9.4 billion liters to communities. The report gives a WUE of 0.12 L/kWh, says 22% of leased and owned withdrawals occurred in high or extremely high water-stress regions, and reports reclaimed-water use at 26 operating data centers. Equal global totals do not mean the same water was restored in the same watershed or at the same time.

Microsoft

Company reporting shows growth pressure

Microsoft's July 2026 update says total Scope 1, 2, and 3 emissions rose 25% year over year, driven mainly by data center expansion and a change in renewable-certificate accounting. Separate water reporting says average WUE across its owned fleet fell to 0.27 L/kWh in 2025 and water-withdrawal intensity was 25% below its 2022 baseline. The two trends show why efficiency and total growth must be reported together.

Meta

Company and facility reporting

Publishes data center information on electricity, water, clean-energy matching, and efficiency. As with other hyperscalers, company-wide environmental reporting is more complete than AI-model-specific lifecycle reporting.

NVIDIA

Hardware lifecycle reporting

Published cradle-to-gate product carbon footprints for two AI baseboards using ISO 14067 methods and third-party verification. NVIDIA reports 24% lower embodied carbon across large AI workloads when comparing HGX B200 with HGX H100. This is a useful hardware-level step, but it covers two products and not the full operating footprint of a data center.

OpenAI

Specific project commitments; not operating results

For its planned 3.2 GW Project Camellia in Georgia, OpenAI says it will pay project-specific power and infrastructure costs, proactively reduce demand before residential customers are affected during high-demand periods, use closed-loop cooling, and provide $80 million in community benefits. These are pre-operation commitments and should be checked against permits, contracts, and measured performance after the facility opens.

Anthropic and xAI

Selected commitments; limited comparability

Public materials describe selected electricity-cost, cooling, power, or community commitments. This ACC review did not locate a comparable company-wide AI lifecycle total for either company, so direct ranking would create more certainty than the evidence supports.

The most useful next step is standardized disclosure.

Companies should publish facility electricity and water, local water stress, time- and location-based power data, training and inference impacts, chip and construction impacts, and uncertainty using comparable boundaries. Without that, a polished sustainability claim can still be impossible to compare.

Climate uses and solutions

AI can be useful, but usefulness should be demonstrated.

The strongest case for AI is not a promise that it will solve climate change. It is a growing set of specific applications that can be tested against their own cost.

One concrete example arrived in May 2026, when the European Centre for Medium-Range Weather Forecasts made AIFS v2 operational. The system added data-driven wave and snow-cover forecasting to its operational weather suite. AI is also being applied to grid forecasting, renewable integration, fault detection, methane detection, building controls, materials research, and industrial optimization.

Operational example

AI weather forecasting moved from experiment to public infrastructure.

ECMWF's AIFS v2 now runs alongside physics-based forecasting. That is a stronger claim than "AI may help someday" because the system has a defined public purpose, measurable forecast performance, and an operational setting.

Review the ECMWF update →
  • Weather forecasting
  • Grid and renewable forecasting
  • Fault and methane detection
  • Efficient buildings and industry
Public weather services are expanding the operational pathway.

On July 27, 2026, NOAA announced that its Weather and Climate Operational Supercomputing System would move to commercial cloud infrastructure and identified AIGFS as its first machine-learning weather-model suite. This is a meaningful public-infrastructure commitment, but it is not yet evidence that every model in the suite will improve forecasts or reduce total resource use; those outcomes still need published evaluation.

Review the NOAA announcement →

A simple net-value test

PurposeWhat real problem does it solve?

The benefit should be specific enough to measure.

Additional impactWhat energy, water, hardware, and land does it add?

Count the infrastructure, not only the software interface.

EvidenceDoes it improve outcomes in practice?

Measure actual performance, not only a demonstration.

AlternativesCould a simpler tool solve it with less impact?

Use AI where it adds value, not by default.

What responsible AI infrastructure requires

Comparable lifecycle data

Report training, inference, facilities, hardware, construction, water, emissions, and uncertainty using shared boundaries.

New clean supply

Add low-emission generation, storage, and transmission instead of simply claiming clean power already serving someone else.

Carbon-aware flexibility

Shift suitable workloads away from grid stress and high-emission hours. Price alone can move demand toward fossil generation, so grid and emissions signals must be considered together.

Water-smart siting

Use watershed conditions, reclaimed water, closed-loop systems, and cooling tradeoffs to avoid worsening local water insecurity.

Longer hardware life

Improve utilization, reuse components, reduce premature replacement, and account for chips, steel, and concrete.

Enforceable community safeguards

Protect ratepayers, publish operating data, limit noise and pollution, and make companies pay for required infrastructure.

Efficient cooling

Use higher-temperature liquid loops and climate-appropriate dry cooling where they reduce chiller energy and water without shifting hidden impacts elsewhere.

Useful heat and water partnerships

Evaluate treated-wastewater cooling and nearby heat users before siting, then report measured savings instead of claiming modeled potential as an operating result.

Promising engineering, with boundaries

Cooling and infrastructure partnerships can reduce resource use.

A 2026 peer-reviewed study, validated against onsite measurements, estimates that converting from air-to-chip to liquid-to-chip cooling could reduce annual energy use and PUE by 4% to 13% per unit of compute, depending on climate and controls. Another peer-reviewed global model mapped 4,775 data centers against 57,547 wastewater plants and found large potential for treated-effluent cooling and useful heat recovery. The first mixes measurement and modeling; the second is a modeled opportunity. Both need project-level verification.

  • 4-13%Modeled reduction in annual energy use, carbon, and PUE per unit of compute after liquid-to-chip conversion.
  • 6-14%Modeled reduction in peak power per unit of compute in the same cooling study.
  • Verify locallyClimate, water stress, heat users, control quality, and the electricity mix determine the real result.

Questions to ask locally

A serious proposal should answer these questions before approval.

Power

How much electricity and peak capacity will the project request, what new generation will serve it, and who pays for grid upgrades?

Water

How much will be withdrawn and consumed, from what source, during what seasons, and under what drought restrictions?

Climate and air

What are the operational and embodied emissions, backup-generator impacts, and time- and location-based power sources?

Community

What are the noise, traffic, land, tax, housing, and ratepayer effects, and what commitments are legally enforceable?

Accountability

What data will be published every year, who verifies it, and what happens if the facility exceeds approved limits?

Prepare a local response

Turn a proposal into specific questions, evidence requests, and messages.

Open the Take Action Directory →

ACC takeaway

Judge AI by its whole footprint and its demonstrated usefulness.

AI is neither weightless nor automatically harmful. The responsible path is to measure the full system, publish comparable evidence, protect host communities, add genuinely cleaner infrastructure, and use AI where the real-world benefit justifies the cost.

Sources and further reading

Review the evidence behind this guide.

Current demand, grids, and communities

Lifecycle, water, and measurement

Company disclosures and useful applications

Foundational context