The Environmental Impacts of AI
Cutting through the noise for more sustainable artificial intelligence
19 May 2026
Guidance
Table of Contents
- The environmental impact of AI
- Why AI footprint estimates vary
- Sustainable AI practices
- Practical steps for responsible and sustainable AI use
- How Anthesis can help
Share this article
In 2024, data centres consumed around 415 TWh of electricity globally, roughly 1.5% of total global demand. With the rise of Artificial Intelligence (AI), energy use from data centres is projected to more than double by 2030. In the U.S. alone, data centres accounted for approximately 4.4% of national electricity consumption in 2023, a share expected to triple by 2028. As that demand accelerates, organisations are pressing ahead with AI integration faster than Chief Sustainability Officers can get their arms around the implications for their environmental footprints.
AI’s environmental impact is not a peripheral technical question. Instead, it sits squarely within the scope of corporate sustainability governance, procurement decisions, and Scope 3 reporting. With incomplete data, highly variable claims, and a rapidly evolving AI landscape, however, it also presents a headwind for achieving carbon reduction targets.
The environmental impact of AI is real and multi-dimensional
While energy demand is the most visible dimension of AI’s environmental footprint, hardware and materials also tell an equally important story. Global chip production has increased fourfold since 2001, driven in part by AI acceleration. Meanwhile, only 22% of global e-waste is currently recycled, and many of the critical minerals required for AI hardware – cobalt, lithium, and rare earth elements among them – overlap almost entirely with those needed for the low-carbon energy transition.
Water demand adds another layer: global water consumption from AI is projected to reach between 4.2 and 6.6 billion cubic metres per year by 2027. Training a single large language model such as GPT-3 consumed around 5,440 cubic metres of freshwater, with nearly 90% of that demand coming indirectly from power generation rather than on-site cooling. Inference – the step after training where an AI executes actions – was estimated to demand 17 ml per request.
AI’s footprint is not hypothetical. It is growing, resource-intensive, and distributed across energy systems, water resources, and material supply chains.
Why AI footprint estimates vary so widely
Public estimates of the environmental impact of AI span orders of magnitude, with per-query carbon intensity figures ranging from fractions of a gram to several hundred grams of CO₂e. Methodology, not misinformation, explains most of the variation.
Four factors drive the spread in measurements of the environmental impact of AI:
- What is included in the system boundary: Some studies count only GPU electricity during inference; others include cooling, power infrastructure, network equipment, and embodied hardware emissions. Expanding the system boundary alone can increase estimated climate impact by nearly 70%.
- What the AI is asked to do: Task type is the single biggest driver of impact in determining energy consumption per query. Image generation requires around 60 times more energy than text-based tasks, with video generation another order of magnitude higher.
- Where computation takes place: The same model, running the same task on identical hardware, can produce three times the emissions depending on the local electricity grid. Low-carbon grids reduce climate impacts but can increase how much water is used, particularly where hydropower dominates.
- How quickly AI technology is becoming more efficient: NVIDIA reports that its GPU energy efficiency for large language models has improved 45,000 times over the last eight years, making older estimates unreliable reference points for newer, more energy efficient AI models. But this level of efficiency gain may not be sustainable over the long run. Additionally, efficiency improvements also reduce the cost and friction of using AI, which simultaneously drives higher usage volumes. Total environmental impact can still rise even as energy consumption per query falls .
Life cycle assessment (LCA) provides a framework for navigating this complexity. In an AI context, LCA can assess impacts across the full service chain; from data acquisition and model training, to inference, network transmission, end-user devices, and hardware end-of-life, rather than stopping at the boundary of a single data centre or task. LCA forces clarity on what is actually being compared, per query, per year, or per unit of computation, and moves assessment beyond carbon to capture water use, mineral depletion, and other impact categories a carbon-only lens misses. It is the foundation on which credible procurement standards and Scope 3 reporting for AI will ultimately need to be built.
Focus on strategic decisions that matter – sustainable AI practices
The difficulty of pinning down precise figures for AI’s environmental impact should not distract from a more useful question: which decisions have the biggest influence on environmental impact and sit within your control?
For many common business tasks, using a smaller AI model instead of a very large one can reduce energy use by orders of magnitude (e.g., by ~60 times when comparing an 8-billion parameter model to a 400-billion parameter model) while still delivering an equally valuable output. These impacts are shaped by strategic decisions focused on:
- Which models are procured or approved
- How employees use AI in practice
Managing AI’s sustainability impact therefore depends as much on governance and usage norms as it does on technical efficiency.
Practical steps for responsible and sustainable AI use
Sustainability leaders who are being asked how their organisation is addressing AI’s environmental impacts will find the most credible answers in clear governance frameworks and deliberate design choices, not in chasing precise emissions figures the underlying data cannot yet support.
In most cases, the actions most likely to reduce AI’s environmental impact are the same ones that reduce cost and improve output quality.
1. Establish usage governance before optimising efficiency
Total AI impact is a product of efficiency and volume – with volume being the variable most organisations underestimate. Efficiency gains can be entirely offset by increased usage.
Many organisations can build on existing ethics and data protection governance to define appropriate-use guidelines that specify when AI should and should not be used and can monitor usage patterns at the organisational level. Without governance over volume, you risk optimising the wrong variable. Once volume is known and tracked, you’re in a strong position to measure impact as the precision of environmental metrics increases over time.
2. Match models to tasks
Smaller, lighter models should be the starting point for day-to-day workflows. Larger, deeper-thinking models are the escalation option, not the default.
Task-specific models represent a more efficient choice at scale, but only when viable options exist. The reality of today’s AI market is that most organisations find the only viable models are general-purpose options designed for breadth and volume. The economics of large-scale AI deployment favour these energy-intensive, jack-of-all-trades systems. To move the market, business customers should let these organisations know that there is demand for more energy-efficient, task-specific options.
3. Avoid duplicate computation
Many enterprise workflows involve repeated or near-identical queries, driving unnecessary energy use and cost. Establish processes to identify where outputs can be reused rather than regenerated and build shared repositories for standard analyses. This is often higher impact than prompt-level optimisation.
4. Be intentional about task type
Text should be the default modality. Reserve image and video generation for tasks where visual output is genuinely necessary, not merely convenient.
5. Set expectations for concise outputs
Longer outputs drive disproportionate computational and energy costs. Encourage employees to use their prompts or custom instructions to specify concise and direct outputs as the default. Longer responses should be the exception, not the default.
It is also worth noting that general sustainability best practices likely already implemented should be expanded to explicitly cover AI. On procurement, raise the bar for what disclosure looks like. Beyond carbon certificates and renewable energy claims, request location-based emissions data, water efficiency metrics, and transparency on hardware lifecycles. Data centre siting, cooling strategies, and power sourcing sit outside most organisations’ direct control, but not outside the scope of procurement expectations. To shift provider behaviour, make these questions a standard part of vendor evaluation, not an afterthought.
Begin incorporating AI-related energy and emissions into your Scope 3 reporting framework, even where estimates remain directional. Standardised methodologies for AI carbon accounting are still emerging, but starting now – with vendor disclosures, energy consumption data, and location-based grid carbon intensity – positions your organisation to evolve with the methodology rather than scrambling to catch up later.
How Anthesis can help
AI usage is a material and growing part of most organisations’ environmental footprint, and it’s an emerging area for the sustainability function to drive impact and add value.
We support organisations navigating sustainability questions where data, expectations, and regulation are developing quickly. Our relevant services include life cycle assessment of AI-enabled products and digital operations, Scope 3 accounting and measuring AI impact against targets, evaluation of AI decarbonisation opportunities and development of roadmaps, procurement guidance for cloud and AI service providers, tech stack supplier engagement, development of AI use guidelines grounded in environmental materiality, and identifying and implementing responsible AI applications that amplify impact and drive efficiency.
Explore our Life Cycle Assessment Solutions
Anthesis is a global leader in delivering actionable Life Cycle Assessments (LCAs) for organisations across sectors with a team of industry-leading experts.
We are the world’s leading purpose driven, digitally enabled, science-based activator. And always welcome inquiries and partnerships to drive positive change together.
