A data centre corridor lined with rows of black server racks

Myth vs Fact: Is AI Really Bad for the Environment?

Article 1 min read

A data centre corridor lined with rows of black server racks. Photo: Brett Sayles, Pexels License, cropped.

Download

Separating real AI environmental impacts from exaggerations and wishful thinking.

By ESG Astraa Advisory Team

In brief

Is AI wrecking the climate, or is its footprint too small to matter? Both claims circulate widely, and both are easy to support with a number from somewhere.

Data centres' share of global electricity, 2024 to a 2030 projection
1.5% → 3%
Rise in data centre electricity use in 2025 (IEA, April 2026)
+17%
Growth of AI-focused data centres in the same year
+50%
In this article
  1. 01 · BOUNDARIES
  2. 02 · OVERVIEW
  3. MYTH 1
  4. MYTH 2
  5. MYTH 3
  6. MYTH 4
  7. 03 · BROADER LESSONS
  8. CONCLUSION
  9. Frequently asked questions
  10. References

01 · BOUNDARIES

AI's footprint has several parts. Training builds a model, and inference runs it each time someone uses it. Impacts also fall in two places.

Exhibit 1

Diagram comparing direct data centre impacts such as computing and cooling with indirect impacts including grid emissions and chip manufacturing.
Direct and indirect impacts.

Exhibit 2

Horizontal bar chart comparing energy for a median Gemini text prompt: 0.10 Wh for active AI chips only versus 0.24 Wh for the full serving system.
View the data
BoundaryEnergy (Wh)
Active AI chips only0.10 Wh
Full serving system0.24 Wh
Same prompt, two boundaries. Source: Google, technical paper arXiv:2508.15734, August 2025. Counting the full system gives 2.4 times the chip-only figure.

02 · OVERVIEW

Each myth has a verdict. None of the four is simply true, and none is simply false.

MYTH 1

Why people believe it. Headlines about gigawatt campuses and rising power demand suggest AI is reshaping global energy use.

Exhibit 3

Bar chart of data centre electricity use showing 415 TWh in 2024, 485 TWh in 2025, and approximately 950 TWh outlook for 2030.
View the data
YearElectricity use (TWh)
2024415
2025485
2030 outlook≈950
Data centre electricity use, terawatt-hours (TWh). Source: IEA, Energy and AI (2025) and Key Questions on Energy and AI (April 2026). The 2025 base case put 2030 at about 945 TWh; the April 2026 update keeps it near 950 TWh.

MYTH 2

Why people believe it. Early academic estimates for large models were high. One widely cited study estimated that GPT-3 used a 500 mL bottle of water for roughly 10 to 50 medium-length responses, counting both on-site and off-site water.

Exhibit 4

Bar chart comparing estimated water consumption per prompt: 10 to 50 mL for GPT-3 (Li et al.) versus 0.26 mL for Google median Gemini text prompt.
View the data
EstimateWater per response (mL)
GPT-3 estimate (Li et al.), on-site and off-site10 to 50 mL
Google median Gemini text prompt, May 20250.26 mL
Water per response, millilitres (mL). Drawn to one scale, 0 to 50 mL. The two figures use different methods

MYTH 3

Why people believe it. Companies report large per-task gains, which suggests total impact must fall.

Exhibit 5

Side-by-side comparison of falling per-task energy (33x lower for Gemini over 12 months) against rising total data centre electricity demand (+17% in 2025).
Per-task efficiency versus total demand. Source: Google, August 2025; IEA, Key Questions on Energy and AI, April 2026.

MYTH 4

Why people believe it. AI is used in grid planning, building controls and industrial optimisation, so savings seem automatic.

Exhibit 6

Bar chart comparing current and future data centre emissions (180 Mt today, ~350 Mt in 2035) with conditional AI-enabled emission cuts of 1,400 Mt by 2035.
View the data
CategoryMillion tonnes of CO2 (Mt)
Data centre emissions today180 Mt
Data centre emissions, 2035≈350 Mt
Possible cuts from existing AI uses, 2035, if adoption widens1,400 Mt
Data centre emissions against potential AI-enabled cuts. Source: IEA, Energy and AI (2025), Widespread Adoption Case; IEA, April 2026 update. The 2025 report's base case for 2035 data centre emissions was 300 Mt. Hatched bar = conditional, not assured.

03 · BROADER LESSONS

  1. 101 Ask what is being measured
  2. 202 Match the model to the task
  3. 303 Check the power source and location
  4. 404 Treat savings as a hypothesis
  5. 505 Ask for disclosure

CONCLUSION

The evidence supports neither alarm nor dismissal. AI's electricity use is small globally but growing fast and concentrated locally, per-prompt figures depend on method, efficiency gains do not guarantee lower totals, and emission savings have to be earned.

Frequently asked questions

How much electricity does AI and data centre infrastructure consume?
Data centres used about 415 TWh in 2024 (roughly 1.5% of global electricity) and rose 17% to 485 TWh in 2025, with IEA projections reaching about 950 TWh or 3% of global electricity by 2030.
How much water does a typical AI prompt use?
Google reported approximately 0.26 mL of cooling water for a median Gemini text prompt, whereas earlier studies of larger models like GPT-3 estimated 10 to 50 mL when factoring in off-site power generation water.
Can AI reduce more carbon emissions than it produces?
While AI applications could theoretically enable up to 1,400 Mt of emission cuts by 2035 compared to 350 Mt of data centre emissions, the IEA notes these savings depend heavily on widespread deliberate adoption and are not guaranteed.

Where ESG Astraa can help

Carbon accounting softwareDecarbonisation planningESG training

About the author

ESG Astraa Advisory Team. ESG Astraa helps Indian companies collect ESG data, calculate emissions and file BRSR and other disclosures, with advisors alongside the software.

Talk to our teamDownload the PDF

References

  1. IEA, Energy and AI: Energy demand from AI iea.org
  2. IEA, Artificial intelligence topic page iea.org
  3. IEA, Energy and AI: AI and climate change iea.org
  4. Reglobal, Energy for AI: summary of the IEA report reglobal.org
  5. IEA, Key Questions on Energy and AI: Executive summary, 16 April 2026 iea.org
  6. Euronews, 11 April 2025, reporting the IEA's conclusions on AI and emissions euronews.com
  7. Google, How much energy does Google's AI use? We did the math, August 2025 (via techUK); technical paper arXiv:2508.15734 techuk.org
  8. DatacenterDynamics, Google median Gemini prompt uses 0.24 watt-hours of power and consumes 0.26 ml of water datacenterdynamics.com
  9. Li, Yang, Islam and Ren, Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models, arXiv:2304.03271 arxiv.org
  10. CBS News, What's the environmental cost of an AI text prompt? Google says it has an answer cbsnews.com
  11. Transition Economics Institute, IEA Says Data Centre Electricity Use Rose 17% in 2025 tei-energy.org

Related articles

New research, in your inbox.

White papers, case studies and regulatory explainers when they are published.

Work email only. Unsubscribe any time.