03 · What You Need to Know
Where the Environmental Footprint of AI-Assisted Research Comes From
Why a digital AI interaction has a physical footprint
AI systems operate on physical computing infrastructure. When researchers submit a request to a hosted language model, processors execute mathematical operations, memory systems store intermediate information, and supporting equipment enables the computation.
Electricity powers these operations. Depending on the deployment, additional resources support cooling, networking, storage, and infrastructure management.
Locally executed models also consume resources. Running inference on a laptop or institutional workstation does not eliminate electricity consumption or the environmental consequences associated with manufacturing the equipment.
The environmental footprint is therefore not limited to the electricity used during an individual response. A broader assessment may include infrastructure production, equipment operation, maintenance, and end-of-life treatment.
The OECD's report Measuring the Environmental Impacts of Artificial Intelligence Compute and Applications provides a framework for distinguishing these direct impacts from wider environmental consequences associated with AI applications.
What is the difference between training and inference?
Training is the computational process through which a model learns or adjusts its parameters. Inference occurs when a trained model processes an input and produces an output.
Training large models can require substantial computational resources. Inference may require much less computation for an individual request, but repeated use across many requests can accumulate significant consumption.
Researchers should not assume that training always accounts for most of an AI system's lifetime environmental impact. The balance depends on the model, the frequency and complexity of inference, and the period over which the system operates.
A researcher using an existing hosted model generally does not directly repeat its original training process. The immediate workload involves inference, although the model's development and infrastructure remain part of broader lifecycle considerations.
What environmental resources does AI use?
| Environmental dimension |
Source of impact |
What researchers should understand |
| Electricity |
Processors, memory, storage, networking, and supporting infrastructure |
Energy consumption depends on the actual workload and computing environment. |
| Greenhouse-gas emissions |
Electricity generation, equipment production, and other lifecycle processes |
Electricity use alone does not establish emissions without additional assumptions. |
| Water |
Some cooling systems, electricity generation, and manufacturing |
Water impacts depend on technology, geography, and local conditions. |
| Materials |
Production of chips, servers, networking equipment, and infrastructure |
Hardware has environmental burdens beyond its operating electricity. |
| Electronic waste |
Equipment replacement and disposal |
Equipment lifetime and end-of-life management affect lifecycle impacts. |
Not all dimensions are equally measurable for a particular study. A researcher operating a local workstation may be able to estimate its electricity consumption but have limited information about the environmental burdens associated with manufacturing its components.
A researcher using a commercial AI API may have token usage records without access to the provider's workload-specific energy measurements.
How much electricity do data centers consume?
The International Energy Agency's Energy and AI report, published in 2025, estimated that data centers consumed approximately 415 terawatt-hours of electricity worldwide in 2024, representing about 1.5% of global electricity consumption.
The report projected that data-center electricity demand could reach approximately 945 terawatt-hours by 2030 under its base-case assumptions.
These figures describe data centers collectively, including infrastructure supporting many digital services beyond AI. They should not be interpreted as the electricity consumed exclusively by AI research.
They also represent estimates and scenario projections rather than direct measurements of every facility or guaranteed future outcomes.
For individual researchers, global data-center statistics provide context but do not establish the environmental impact of a particular prompt, experiment, or study.
How does electricity consumption relate to carbon emissions?
Electricity consumption is measured in energy units such as kilowatt-hours. Greenhouse-gas emissions depend partly on the sources and characteristics of the electricity used.
Two AI workloads consuming the same amount of electricity may have different associated emissions when operated under different electricity systems or accounting conditions.
A basic operational estimate can be calculated when an appropriate electricity emissions factor is available.
The assumed factor is illustrative, not a verified value for a particular electricity grid or provider. The result excludes manufacturing and other lifecycle emissions unless they are separately included.
Researchers should not use this calculation when neither electricity consumption nor an appropriate emissions factor can be reasonably established.
Why water consumption is part of the environmental discussion
Some data centers use water-dependent cooling systems. Water may also be consumed indirectly through electricity generation and hardware manufacturing.
However, water withdrawal and water consumption are not equivalent. Withdrawn water may be returned to its source, whereas consumed water is not immediately returned to the same local water system.
The significance of water use also depends on local availability. A particular volume of consumption may have different implications in water-stressed and water-abundant locations.
Researchers should therefore be cautious about universal claims that every AI prompt consumes a fixed quantity of water.
Without workload-specific information about infrastructure, cooling arrangements, and geographic conditions, such estimates may create a misleading impression of precision.
What are operational and embodied environmental impacts?
Operational impacts arise while computing infrastructure is being used. Embodied impacts arise from activities associated with producing and providing the equipment and infrastructure, including material extraction, manufacturing, and transportation.
Operational impacts
Environmental burdens associated with running the AI workload, including electricity consumption and relevant supporting infrastructure.
Embodied impacts
Environmental burdens associated with producing the hardware and infrastructure used to support computation.
Embodied impacts are part of a broader lifecycle assessment. Researchers should not describe measured GPU electricity consumption as the complete environmental footprint of an AI system.
Allocating manufacturing impacts to one research workload can be difficult because hardware may support many projects over several years.
Where such allocation is attempted, the assumptions and boundaries should be made explicit.
Does using a larger AI model necessarily cause greater environmental impact?
Not in every comparison.
Larger models often require greater memory and computational resources under comparable execution conditions. However, actual consumption also depends on architecture, hardware, numerical precision, software optimization, context length, and the amount of processing needed to complete the task.
A smaller model may be more efficient per request but require additional attempts or extensive correction. A larger model may complete the task with fewer operations, although that advantage must be demonstrated.
Researchers should therefore consider the resources required to obtain a scientifically acceptable result, not model size alone.
The scientific sufficiency of smaller AI models is relevant when evaluating whether more computationally demanding alternatives are necessary.
Does every AI query have the same environmental footprint?
No. Requests can differ in input length, generated output length, model architecture, reasoning demands, and whether additional tools or processing stages are involved.
Infrastructure conditions also matter. Hardware utilization, batching, cooling requirements, and electricity sources may influence the resources associated with processing.
A short request to classify a sentence is not directly comparable to generating a lengthy analysis of a large document collection.
The environmental impact of individual AI queries should therefore be treated as variable rather than assigned a universal numerical value.
Can AI reduce environmental impacts elsewhere?
Potentially. AI can support applications involving energy systems, environmental monitoring, resource management, and other activities that may produce environmental benefits.
However, potential benefits should not be confused with demonstrated net environmental improvements.
An AI system intended to improve efficiency may also increase demand for a service or require substantial computation. Such indirect effects can complicate the overall assessment.
Researchers should distinguish the direct footprint of operating an AI system from the environmental consequences of the activities it enables.
Claims that an AI application offsets its own environmental footprint require a defensible comparison and evidence about actual outcomes.
Should researchers avoid AI because it consumes energy?
Not automatically. Scientific research routinely uses physical resources, including laboratory equipment, computing infrastructure, travel, and specialized instruments.
The relevant question is whether AI provides sufficient scientific value relative to its costs and available alternatives.
For some tasks, AI may reduce the need for other resource-intensive activities. For others, a conventional computational method or human analysis may be more appropriate.
The comparison should consider scientific validity and the complete workflow rather than assuming that either AI or non-AI methods are universally more sustainable.
When processing occurs at substantial scale, environmental and resource costs in large-scale AI research may become an important part of methodological planning.
What can researchers realistically measure?
Researchers operating their own computing infrastructure may be able to measure system power and processing duration using appropriate monitoring equipment or software.
Those using hosted services may have access to request counts, token usage, processing duration, or provider-reported sustainability information.
These quantities are not interchangeable. API expenditure is not a direct measure of electricity consumption, and token counts alone do not establish greenhouse-gas emissions.
Researchers should report the measurements actually available and identify any estimation assumptions.
The principles of reporting AI computational costs can help establish a transparent record of resource consumption even when environmental impacts cannot be calculated precisely.
