Training
GPT-4 consumed an estimated 50 gigawatt-hours of electricity, roughly
equivalent to the annual energy consumption of 5,000 average UK homes,
according to estimates published by research groups including the AI Now
Institute and the University of Massachusetts. This figure is contested and
uncertain because major AI laboratories do not publish the energy consumption
data needed to calculate it precisely, but the order of magnitude is not
seriously disputed. The inference workload that follows training, running the
trained model to answer queries from millions of users, adds further energy
demand that grows with adoption. As AI models become larger, more capable,
and more widely deployed, the energy and carbon implications of the industry
are becoming a significant and increasingly discussed environmental concern
that sits awkwardly alongside the industry’s own sustainability
commitments.
The framing of AI energy consumption as a hidden cost is accurate
in a specific sense: unlike the carbon footprint of air travel or beef
consumption, which has received sustained public attention, the carbon cost
of AI use is largely invisible to the individuals and organisations
generating it. When a user asks a question of ChatGPT or generates an image
with Midjourney, they receive no information about the energy consumed in
producing the response. When a business deploys an AI model in production,
the energy cost is typically buried in cloud computing bills that obscure the
specific contribution of AI workloads. This opacity is not accidental, and
addressing it requires both technical transparency from AI providers and
policy frameworks that treat AI energy consumption as a reportable
environmental impact.
The Numbers and Their Uncertainty
The difficulty of calculating AI’s carbon footprint precisely
reflects genuine technical complexity alongside deliberate opacity from major
AI providers. Training costs depend on model size, training duration,
hardware efficiency, and the carbon intensity of the electricity grid where
training is performed. A model trained on Nvidia H100 GPUs powered by
Norwegian hydroelectricity has a fraction of the carbon footprint of the same
model trained on older hardware powered by coal-heavy grid electricity. The
same model generates different inference emissions depending on whether
queries are served from data centres in Iceland or from facilities in regions
with high-carbon electricity grids.
Research from Stanford University’s Human-Centered Artificial
Intelligence Institute has found that the carbon footprint of AI
training grew by an order of magnitude between 2018 and 2023, driven by
increasing model sizes and the democratisation of AI development to a larger
number of organisations. The emissions from AI inference, which is the
operational deployment of trained models rather than the training itself, now
exceeds training emissions in aggregate because of the scale of deployment. A
model that took weeks and significant energy to train may run billions of
inference queries over its operational lifetime, each consuming a small but
non-trivial amount of energy.
Water Consumption: The Less-Discussed Impact
Alongside electricity consumption, AI data centres consume
significant quantities of fresh water for cooling. A 2023 study published in
Nature by researchers at UC Riverside estimated that GPT-3’s training
consumed approximately 700,000 litres of fresh water for cooling, and that
each conversation with ChatGPT consumes roughly 500ml of water. Microsoft,
Google, and Amazon have all reported increasing water consumption in their
data centre operations alongside increasing AI workloads, though the precise
attribution to AI specifically versus other cloud computing workloads is
difficult to disaggregate.
The water consumption concern is most acute in regions where data
centres are sited in water-stressed environments. Several major AI data
centre developments have been proposed or are operating in areas with below-average
water availability, where industrial-scale water consumption for cooling
creates direct competition with agricultural and domestic water needs. The
environmental justice dimensions of data centre water use, which tends to
concentrate consumption in communities that host data centre infrastructure
while distributing the benefits globally, deserve attention from planners and
policymakers that they have not consistently received.
The Industry Response
Major AI companies have made significant commitments to renewable
energy procurement, and several are genuinely investing in making their
infrastructure greener. Microsoft has committed to being carbon negative by
2030 and water positive by the same date. Google has published detailed
sustainability reports documenting its renewable energy procurement. Amazon
Web Services operates large renewable energy programmes that include the UK
facilities discussed in its recent investment commitments. These commitments
are genuine, but they need to be evaluated against the rate of growth in AI
workloads, which is outpacing renewable energy procurement in ways that the
overall trend in data centre carbon emissions reflects. The International
Energy Agency has noted that data centre electricity consumption is
growing faster than the renewable capacity being added to serve it in most
major markets.
What This Means for You
The carbon cost of AI is not equally distributed. AI training and
inference is performed in data centres; the communities hosting those data
centres bear the local environmental impacts, including noise, water
consumption, and land use, while the benefits of AI services are consumed
globally. This is not unique to AI, but it is a form of environmental
injustice that deserves acknowledgment alongside the genuine economic and
social benefits that AI infrastructure provides to host communities. As an
individual user, the carbon cost of your AI use is small relative to other
consumption choices, but it is not zero, and transparency from AI providers
about the environmental cost of their services is a reasonable expectation
that the industry has not yet consistently met. The positive feedback loop
between AI capability and AI energy consumption deserves explicit
acknowledgment. More capable AI models require more compute to train and to
run; more compute requires more energy; more energy requires more data centre
infrastructure; more data centre infrastructure enables more AI capability.
Breaking this loop requires either significant efficiency improvements in AI
hardware and algorithms that reduce energy consumption per unit of AI
capability, or governance frameworks that impose energy consumption
constraints on AI development and deployment. Both are technically possible;
neither is currently happening at the scale or speed that the trajectory of
AI energy consumption demands. The research community working on AI
efficiency, including groups at Google DeepMind, MIT, and the UK’s own Turing
Institute, is making genuine progress on reducing energy per parameter of AI
models, but efficiency gains are currently being absorbed by model scaling
rather than reducing absolute energy consumption. For related analysis, see
our coverage of green
AI infrastructure investment and the
renewable energy transition.
The policy responses available to address AI’s carbon footprint
range from voluntary disclosure to mandatory regulation, and the evidence
from other sectors suggests that voluntary approaches are unlikely to produce
the behavioural change needed at the pace required. The UK’s approach to data
centre environmental regulation has been primarily through planning
requirements and voluntary commitments, with limited mandatory reporting on
energy and water consumption. The EU’s Energy Efficiency Directive imposes
mandatory reporting requirements on data centres above a certain size
threshold, providing a precedent that UK post-Brexit policy has not yet
matched. Mandatory carbon disclosure for AI workloads, equivalent to the
scope 3 emissions reporting requirements that apply to other large energy
consumers, would create accountability for AI providers that currently face
no regulatory incentive to reduce the carbon intensity of their operations.
The Climate Change
Committee has identified data centre emissions as a growing
category in UK carbon accounts that current policy is not adequately
addressing, and its recommendations for mandatory reporting and efficiency
standards deserve implementation ahead of the next AI infrastructure
investment cycle.
About the Author
Stuart Kerr is a technology correspondent at LiveAIWire, covering
artificial intelligence, digital innovation, and the social impact of
emerging technologies. Follow LiveAIWire for daily analysis at liveaiwire.com.