AI runs on real buildings, real power plants, and real supply chains. Data centres are expanding fast, and so are the resources behind them. The International Energy Agency (IEA) projects global electricity use by data centres to roughly double to ~945 TWh by 2030, with AI as a major driver. That’s about the annual electricity use of a large industrialized nation.
Public methods to track footprint are improving. In 2025, Google released a methodology to estimate energy, emissions, and water per AI prompt for its apps. Early figures suggest prompt-level impacts are small per use, yet they scale with volume and depend on when and where the workload runs.
This article maps the less visible environmental pressures linked to AI, then offers practical steps any team can adopt—procurement, siting, operations, and reporting that readers can put to work.
Table of Content
- What “hidden” means here
- Impact #1 — Electricity demand spikes and grid stress
- Impact #2 — Water use you don’t see (direct and indirect)
- Impact #3 — High-GWP gases in chip fabrication (NFâ, SFâ, NâO)
- Impact #4 — PFAS-based chemistries in semiconductor manufacturing
- Impact #5 — Backup diesel generation and neighborhood air
- Impact #6 — E-waste from rapid hardware refresh cycles
- Impact #7 — Land use, noise, and local biodiversity
- Impact #8 — Waste heat and local thermal effects
- Impact #9 — Embodied carbon in hardware (memory and storage matter)
- Impact #10 — Rebound effects (efficiency → more total use)
- Methods that help right now
- How to read the numbers (quick guide)
- Audience: who benefits from this guide
- Quick research overview (what recent work says)
- Practical checklist for teams
- Case examples you can cite
- Risks to watch over the next few years
- Conclusion
- FAQs
What “hidden” means here
Hidden impacts sit outside the headline “electricity use” story. They often appear:
-
Upstream: chip fabrication gases and PFAS-based chemistries.
-
Local: backup diesel testing, noise, and land-use pressure.
-
Indirect: water tied to the power that feeds AI.
Understanding these layers helps avoid blind spots and supports better choices.
Impact #1 — Electricity demand spikes and grid stress
What you see: annual TWh.
What matters: when peaks hit and where build-outs cluster.
IEA scenarios show data-centre electricity use more than doubling to ~945 TWh by 2030, with AI-optimized facilities more than quadrupling. Clustering in a few hubs creates tight local capacity and higher reliance on fossil peakers during evening peaks.
Practical steps
-
Adopt carbon-aware scheduling: shift batch and training jobs to hours and regions with cleaner power. Google publicly describes systems that move flexible compute across time and location to track hourly grid conditions.
-
Publish a peak management plan with utilities: reduce local grid strain during tight hours.
Impact #2 — Water use you don’t see (direct and indirect)
Cooling can use water on-site, yet a large share sits off-site in electricity generation. Thermoelectric plants withdraw and consume water for cooling; values vary widely by technology and cooling type, so the grid mix behind your data centre changes the real water story for the same AI task. Classic NREL and USGS work summarize ranges and trends; EIA maintains current thermoelectric water datasets.
Designs are shifting. Microsoft announced zero-water cooling for next-generation data centres, avoiding evaporative cooling for IT equipment and cutting total cooling withdrawals.
Google’s 2025 method estimates water per prompt by linking prompt energy to fleet water metrics. This moves teams toward budgeting water the same way they budget latency or accuracy.
Practical steps
-
Track both on-site cooling water and indirect water from the grid.
-
Prefer zero-water cooling in arid regions and publish watershed projects for each site.
Impact #3 — High-GWP gases in chip fabrication (NFâ, SFâ, NâO)
Advanced chips that run AI rely on gases with very high global-warming potentials. Atmospheric studies show NFâ emissions rose from ~1.93 Gg/yr in 2015 to ~3.38 Gg/yr in 2021. Fabs can abate these gases effectively, yet coverage and verification differ by site and region.
Practical steps
-
Require abatement and continuous monitoring from suppliers; seek third-party verification.
-
Include NFâ/SFâ control performance in contracts, not only energy metrics.
Impact #4 — PFAS-based chemistries in semiconductor manufacturing
PFAS help in photolithography and other fab steps; the EU is moving toward restrictions with time-limited exemptions or continued use under strict controls for sectors such as semiconductors. The updated REACH PFAS restriction proposal is under review, with reporting and substitution roadmaps in focus. Industry groups (SEMI) publish backgrounders on where PFAS appear and current alternatives.
Practical steps
-
Map PFAS uses in your hardware supply chain and set phase-down targets where alternatives exist.
-
Align with EU data-centre and electronics guidance and track EPA disposal guidance for PFAS-bearing waste streams.
Impact #5 — Backup diesel generation and neighborhood air
Reliability standards demand fleets of emergency generators. Under U.S. rules, engines can run in emergencies without hour limits, and non-emergency operation is commonly capped at 100 hours per year with a subset of hours allowed for demand-response and maintenance/testing. Air permits and local rules add further constraints.
Irish planning files reveal how agencies model NOâ/PM from generator banks and set conditions on testing schedules, stack heights, and run-hour coordination.
Practical steps
-
Spread testing across days and units; avoid simultaneous starts. Publish testing calendars for local communities.
-
Where feasible, adopt cleaner fuels or hybrid backup designs, subject to code and safety limits.
Impact #6 — E-waste from rapid hardware refresh cycles
New models prompt frequent upgrades—GPUs, memory, storage—accelerating retirement. The Global E-waste Monitor 2024 projects 82 million tonnes of e-waste by 2030 and warns documented collection/recycling could drop toward 20% without stronger systems.
Practical steps
-
Set recovery targets for memory and storage modules; publish results.
-
Use certified refurbish/reuse pathways and enforce vendor take-back.
Impact #7 — Land use, noise, and local biodiversity
Large campuses reshape surrounding areas. Reviews in Northern Virginia note community concerns over noise, diesel tests, traffic, and viewsheds; local boards are updating zoning and siting standards to manage growth.
Practical steps
-
Apply noise and lighting limits, plant native buffers, and favour redevelopment of already-industrial sites.
-
Disclose cumulative impacts for multi-site clusters, not only single parcels.
Impact #8 — Waste heat and local thermal effects
Server racks release steady low-grade heat. In places with district-heating networks, that heat can warm homes. Odense, Denmark, pipes heat from a hyperscale campus to the city network through large heat pumps, supplying thousands of homes.
The EU Energy Efficiency Directive now requires data centres to report performance into a European database and includes duties that raise the profile of heat-reuse. Germany’s Energy Efficiency Act adds reuse targets for qualifying sites.
Practical steps
-
Evaluate heat-reuse at site selection; publish MWh of heat delivered and fuel avoided year by year.
Impact #9 — Embodied carbon in hardware (memory and storage matter)
Operational energy gets most attention, yet manufacturing the hardware also carries a large footprint. Peer-reviewed and industry studies find memory and storage can dominate the embodied emissions of servers and storage racks. One research track shows SSD racks carry ~10× the embodied emissions per TB compared with HDD racks; another shows low-carbon SKUs and reused DRAM/SSDs can shrink total footprint.
Practical steps
-
Right-size DRAM and SSD to real workloads; extend lifetimes; prioritize components with third-party embodied-carbon data.
-
Treat storage design as a carbon decision, not only a performance decision.
Impact #10 — Rebound effects (efficiency → more total use)
Efficiency per prompt keeps improving. Google reports large reductions in prompt-level energy and carbon for a major app in 2024–2025. Yet studies warn about rebound effects: lower cost per task can drive more total usage, erasing gains at fleet scale unless teams add guardrails.
Practical steps
-
Pair efficiency with budgets (energy, carbon, water) and caps for high-volume features.
-
Publish per-task metrics along with totals, so growth patterns stay visible.
Methods that help right now
Run jobs when the grid is cleaner
Shift flexible training and batch jobs across hours and regions with higher hourly carbon-free energy. Industry examples show this is feasible at global scale.
Pick cooling that fits the climate and the grid
Cold-plate and other liquid cooling can lower energy use and cut water 30–50% across the life cycle in some configurations; Microsoft’s LCA work informs design choices for new sites.
Budget water like you budget latency
Track direct cooling water and grid-embedded water; adopt zero-water cooling where climate allows; include watershed accounting in site reports.
Push upstream controls
Ask foundries for NFâ/SFâ abatement performance and PFAS substitution plans; track upcoming EU rules and disposal guidance.
Treat storage and memory as carbon levers
Choose capacities carefully, extend lifetimes, and target reuse. Publish embodied-carbon assumptions for major components.
Plan for heat as a product
Co-site with heat networks or future heat loads; set up metering and disclose deliveries each year.
How to read the numbers (quick guide)
-
Totals hide peaks: annual MWh can’t show evening spikes that raise grid emissions; time-shifting matters.
-
Direct vs. indirect water: per-prompt water depends on cooling and the grid behind the server.
-
Training vs. inference: training is intense and episodic; inference runs daily at scale, so prompt-level metrics help teams steer growth.
-
Manufacturing counts: DRAM/SSDs can dominate embodied emissions; right-sizing and reuse help.
Audience: who benefits from this guide
-
Students and educators building curricula in sustainability, computing, or public policy.
-
IT, facilities, and procurement teams deciding where to site, what to buy, and how to operate.
-
Civic planners and regulators reviewing permits for new campuses and backup systems.
Quick research overview (what recent work says)
-
Electricity: growth remains steep through 2030; AI demand is a key driver.
-
Water: direct cooling shrinks with newer designs; indirect water linked to grid mix still matters.
-
Manufacturing: gases with high GWP and PFAS controls require stronger disclosure and abatement.
-
Operations: carbon-aware workload shifting cuts emissions without new hardware.
-
End-of-life: e-waste volumes rise faster than formal recycling rates.
Practical checklist for teams
Procurement
-
Request embodied-carbon data for servers, DRAM, and SSDs; prioritise reuse and longer lifetimes.
-
Ask suppliers for NFâ/SFâ abatement records and PFAS use maps.
Siting
-
Favour regions with low-carbon grids and district-heat options; check water stress indexes.
-
Publish community impact notes on noise and backup testing.
Operations
-
Turn on carbon-aware scheduling; report per-task metrics not only totals.
-
Track water per MWh and water per prompt using a clear method.
Reporting
-
Align with EU data-centre reporting timelines where relevant; share heat-reuse and water data.
Case examples you can cite
-
Heat reuse in Odense, Denmark: surplus server heat warms thousands of homes via large heat pumps.
-
Zero-water cooling designs: vendor disclosures point to chip-level cooling that removes evaporative cooling for IT equipment.
-
Prompt-level impact accounting: energy, carbon, and water per prompt now measurable for some services.
Risks to watch over the next few years
-
Local grid strain if clusters grow faster than transmission upgrades.
-
PFAS rulemaking in the EU and state actions in the U.S., which may alter fab chemistries, disclosure, and disposal.
-
E-waste volumes outpacing collection unless recovery systems catch up.
-
Rebound effects if efficiency gains lower costs and usage surges with no guardrails.
Conclusion
AI’s footprint is not a single number. It spreads across when and where workloads run, how hardware is made and cooled, what happens to components at end-of-life, and which chemicals sit upstream. The good news: teams can act now—shift flexible jobs to cleaner hours, publish water and per-task metrics, push suppliers on abatement and PFAS controls, right-size memory and storage, and treat waste heat as a product. These steps reduce impact without waiting for the next hardware cycle or a distant policy deadline.
FAQs
1) Is training or inference the bigger issue for emissions and water?
Training draws intense power for limited periods. Inference runs every day at large scale. Prompt-level methods now estimate energy, carbon, and water per request, which helps teams steer the daily load.
2) How much water does a single prompt use?
Google’s recent study translates prompt energy into an estimated water per prompt using fleet metrics. Media summaries describe figures near 0.26 mL for a median text prompt, though values vary by grid and method. Treat these as order-of-magnitude and use supplier methods for your stack.
3) Can waste heat from data centres do real work?
Yes. Odense demonstrates city-scale heat reuse with large heat pumps and metered deliveries to the district network. New EU rules put more attention on this path.
4) Why call out memory and storage when talking about embodied carbon?
Studies show DRAM and SSDs can dominate a server’s manufacturing footprint and SSD racks carry far higher embodied emissions per TB than HDD racks. Extending component life and right-sizing capacities reduce total impact.
5) Are diesel generators a major daily pollution source?
They are engineered for emergencies and periodic testing. U.S. guidance places annual limits on non-emergency operation and sets conditions for demand-response use. Local permits add extra rules.
