- Article Summary
-
Introduction
Companies measure the environmental footprint of their AI use by calculating three interconnected metrics — carbon, water, and land — across both training and inference phases, applied against actual query volumes and the grid intensity of the data centers processing them. The core calculation is: Environmental Impact = (energy per query) × (number of queries) × (grid footprint intensity). Without standardized measurement infrastructure, however, most organizations currently have zero visibility into these figures, leaving a material gap in their GHG inventories and ESG disclosures.
Key Takeaways
- AI inference — not training — accounts for an estimated 80–90% of total AI energy use, making everyday enterprise AI usage the primary measurement priority.
- The energy cost of AI varies by orders of magnitude: a typical AI image query draws 2.9 Wh, roughly 1,450 times the energy of a basic text classification task.
- Three footprints must be measured simultaneously — carbon, water, and land — as reducing one can amplify another depending on the data center’s energy mix.
- No industry-wide standard for AI environmental measurement currently exists, leaving most companies with fragmented or zero data on their AI-linked emissions.
Why Is Measuring AI’s Environmental Footprint So Difficult — and Why Does It Matter Now?
AI’s environmental footprint is hard to measure for three reasons: it spans three impact dimensions (carbon, water, land), two operational phases (training and inference), and fluctuates dramatically by task type, model choice, and data center location.
The governance gap is real. Unlike traditional industries with established frameworks such as the Greenhouse Gas Protocol, AI still lacks standardized methods to track energy use, carbon emissions, water consumption, and e-waste across its lifecycle. Existing tools can log some data automatically, but they often trade accuracy for ease of use, leaving organizations with incomplete and fragmented assessments.
The pace problem. The speed of AI improvement has outpaced organizations’ ability to measure and understand the tradeoffs. Without visibility into relative resource intensity, neither efficiency gains nor behavioral change can meaningfully reduce AI’s environmental footprint.
Why it matters now. AI demand is accelerating at a pace that makes this a present-day operational issue, not a future planning item. The global AI market is projected to grow from USD 189 billion in 2023 to nearly USD 5 trillion by 2033 — a roughly 25-fold increase in a decade. Global AI expenditure is projected to exceed USD 2.5 trillion in 2026 alone. As enterprise AI adoption scales, the associated energy load scales with it: by 2028, more than half of all data center electricity could be devoted to AI workloads. Companies that cannot quantify their AI footprint today are building a measurement gap that will only grow harder to close.
Training vs. Inference: Where Does Your Company’s AI Footprint Actually Come From?
Most media coverage focuses on the energy cost of training large models. For enterprise sustainability officers, this is the wrong place to look.
Training benchmarks
| Model | Electricity Consumed | Duration | Carbon Footprint |
|---|---|---|---|
| GPT-3 | 1.287 GWh | 34 days | ~552 tonnes CO₂e |
| GPT-4 | 50–70 GWh | 100 days | ~25,000 tonnes CO₂e |
| GPT-5 (projected) | ~100 GWh | N/A | ~42,000 tonnes CO₂e |
Source: Aczel et al. (2026). Environmental Cost of AI’s Energy Use. UNU-INWEH. doi: 10.53328/INR26RMA002
Training is a discrete, one-off event carried out by the model developer — not the enterprise user. The footprint that belongs to your organization is inference.
Inference is the dominant phase:
- Inference accounts for an estimated 80–90% of total AI energy use.
- ChatGPT alone processes approximately 2.5 billion prompts per day.
- At a conservative 0.42 Wh per text prompt, that translates to roughly 383 GWh of electricity per year.
- As models grow more complex and more embedded in enterprise workflows, inference energy demands will increase further.
Every time your team uses a SaaS AI tool, drafting an email, generating a report image, or running an AI-assisted search, you are generating inference emissions continuously, not in discrete training events. This is your measurement target.
What Is the Formula for Calculating AI’s Environmental Impact?
A practical three-variable formula:
- Environmental Impact = (energy per query for the task/model) × (number of queries) × (grid footprint intensities)
Each variable requires a different data source.
Variable 1 — Energy per query (Wh). This depends entirely on task type and model choice:
| AI Task | Energy per Query | Relative to Text Classification |
|---|---|---|
| Text classification (spam filter) | Baseline | 1× |
| Short text generation | ~0.047 Wh | ~10–25× |
| Typical ChatGPT-style query | ~0.42 Wh | ~200× |
| Long-form GPT response | ~1.9 Wh | ~1,000× |
| Typical AI image generation | 2.9 Wh | ~1,450× |
| High-resolution AI image | 4.08 Wh | ~2,000× |
| High-resolution AI video clip | >415 Wh | ~200,000× |
Source: Aczel et al. (2026). Environmental Cost of AI’s Energy Use. UNU-INWEH. doi: 10.53328/INR26RMA002
Variable 2 — Number of queries. Sourced from:
- SaaS platform billing dashboards
- API call logs from IT or engineering teams
- Software procurement records
Variable 3 — Grid footprint intensities. When vendor-specific data center location is unavailable, use these global averages as a conservative baseline:
| Footprint Type | Global Average Intensity |
|---|---|
| Carbon | 422 g CO₂e per kWh |
| Water | 9.9 L per kWh |
| Land | 154 cm² per kWh |
Source: Aczel et al. (2026). Environmental Cost of AI’s Energy Use. UNU-INWEH. doi: 10.53328/INR26RMA002
AI Task Types and Their Energy Costs: A Reference for Sustainability Officers
Understanding the energy spectrum of your company’s AI use cases is the first step toward setting meaningful reduction targets.
Low-energy use cases (text-based, retrieval-style):
- Email drafting assistants
- Document summarization
- AI-assisted search and Q&A
- Sentiment analysis and classification
Medium-energy use cases (conversational, generative text):
- Extended ChatGPT-style interactions
- Code generation with large context windows
- Long-form content drafting with advanced models
High-energy use cases (image and video generation):
- AI-generated marketing images
- Product visualization tools
- AI video generation for social or advertising content
A company running daily AI image generation at scale carries a categorically different Scope 3 footprint than one using only text-based AI assistants. Enterprises need to know which use cases are driving their footprint before they can act on reducing it.
How Does Data Center Location Affect Your AI Carbon Footprint Calculation?
The same AI workload produces dramatically different footprints depending on which country’s electricity grid powers the data center processing it. This variable is routinely overlooked in enterprise AI assessments.
Carbon intensity varies up to 18-fold across the top 20 data center hubs:
| Country | Carbon Intensity (g CO₂e/kWh) | vs. Global Average |
|---|---|---|
| Indonesia | 682 | +62% |
| India | 635 | +51% |
| China | 510 | +21% |
| United States | 345 | −18% |
| United Kingdom | 218 | −48% |
| France | 51 | −88% |
| Switzerland | 37 | −91% |
Source: Aczel et al. (2026). Environmental Cost of AI’s Energy Use. UNU-INWEH. doi: 10.53328/INR26RMA002
Low-carbon is not automatically low-water or low-land:
- Brazil, Canada, Switzerland, and Sweden carry water footprints more than double the global average due to hydropower dominance.
- The United Kingdom’s electricity grid carries a land intensity more than four times the global average due to bioenergy and onshore wind.
- Renewable-heavy systems can cut emissions but may result in larger land footprints, directly challenging the assumption that data centers powered by renewables are automatically sustainable.
Request data center location information from AI vendors as part of your Scope 3 supplier assessment. Apply location-based grid intensity factors where available and use global averages as a fallback.
Conclusion
Measuring AI’s environmental footprint is no longer an academic exercise. The measurement framework is established: the three-footprint model, the task-level energy benchmarks, and the location-based grid intensity data are all grounded in authoritative institutional research.
What most organizations lack is the platform infrastructure to operationalize that framework at scale, normalizing query volumes, applying grid intensity factors, and producing audit-ready Scope 3 figures that hold up to third-party verification.
Frequently Asked Questions
Sources
- Aczel M., Chamanara S., Matin M., Farsi A., Marwala T., Madani K. Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints. United Nations University Institute for Water, Environment and Health (UNU-INWEH), 2026. View source
- Costa, D. Making AI sustainable by design is key to a better future. World Economic Forum, February 5, 2026. View source
- Zewe, A. Explained: Generative AI’s environmental impact. MIT News, January 17, 2025. View source
Why Work with ASUENE Inc.?
ASUENE is a key player in carbon accounting, offering a comprehensive platform that measures, reduces, and reports emissions, including Scope 1-3. ASUENE serves over 50,000 clients worldwide, providing an all-in-one solution that integrates GHG accounting, ESG supply chain management, a Carbon Credit exchange platform, and third-party verification.
ASUENE supports companies in achieving net-zero goals through advanced technology, consulting services, and an extensive network.
