Comment

years to reflect the latest scientific projections

on sea-level rise and flood risks. Residents will
also benefit from real-time data monitoring,
green building certifications and disaster-risk
mitigation plans.
Developers should be mandated to include
affordable housing for low-income groups,
such as workers and teachers, displaced res-
idents and fishing communities who opt for
waterfront living.
No single climate-adaptation strategy will
work for all communities. An adaptive coastal
future requires flexibility and openness. Float-
ing cities or amphibious houses could be part
of climate adaptation for some riverine areas
and archipelagos; Indigenous communities
might make use of them to stop their heritage,

BRIAN LAWLESS/PA/ALAMY
identity and islands from vanishing. Other
communities might prefer to retreat from
the coast.
Ultimately, climatopias can become tech-
nical solutions to a complex global challenge
only if stricter planning processes, impact
assessments, ecosystem protection and jus- As more data centres crop up in rural communities, local opposition to them has grown.
tice for communities are at the heart of any

Light bulbs have energy

coastal future.

ratings — so why can’t

The authors

Idowu Ajibade is an associate professor of

AI chatbots?
environmental justice and climate adaptation
in the Department of Environmental Sciences,
Emory University, Atlanta, Georgia, USA.
Sameer H. Shah is an assistant professor
of climate adaptation at the School of
Environmental and Forest Sciences, University Sasha Luccioni, Boris Gamazaychikov, Sara Hooker, Regis Pierrard,
of Washington, Seattle, Washington, USA. Emma Strubell, Yacine Jernite & Carole-Jean Wu
e-mails: [email protected];
[email protected]

1. Horton, B. P. et al. npj Clim. Atmos. Sci. 3, 18 (2020). The rising energy and evolution. A report published in January1 by
the International Energy Agency estimated
2. Glavovic, B. C. et al. in Climate Change 2022: Impacts,
Adaptation and Vulnerability. Contribution of Working Group
environmental cost of the that the electricity consumption of data cen-
II to the Sixth Assessment Report of the Intergovernmental
Panel on Climate Change (eds. Pörtner, H.-O. et al.)
artificial-intelligence boom tres could double by 2026, and suggested that
improvements in efficiency will be crucial to
2163–2194 (Cambridge Univ. Press, 2022). is fuelling concern. Green moderate this expected surge.
3. Ude, C. ‘Climate Change and the Disastrous Rain and
Flooding in Nigeria’ New National Star (6 July 2024; policy mechanisms that Some tech-industry leaders have sought
available at go.nature.com/4cp2wrf).
4. Carrere, A., Broad, K. & Mach, K. J. Preprint at Research already exist offer a path to downplay the impact on the energy grid.
They suggest that AI could enable scientific
Square https://doi.org/10.21203/rs.3.rs-3170932/v1 (2023).
5. Chen, W. Heliyon 9, e12771 (2023).
towards a solution. advances that might result in a reduction in
6. Kent, M. J. Latin Am. Caribb. Anthropol. 13, 283–310 (2008).
7. Yogesh, S. WIT Trans. Built Environ. 210, 215–224 (2022).
planetary carbon emissions. Others have
8. Lawanson, O. I., Proverbs, D. & Ibrahim, R. L. J. Flood Risk thrown their weight behind yet-to-be-realized
Mgmt 16, e12838 (2023). energy sources such as nuclear fusion.
9. Ajibade, I. Int. J. Disaster Risk Reduct. 26, 85–92 (2017).
10. Sengupta, D. et al. Earths Future 11, e2022EF002927
However, as things stand, the energy
(2023). demands of AI are keeping ageing coal power

A
11. Al-Amadi, D. M., Major, M. D., Tannous, H. O. & plants in service and significantly increasing
AlKandari, A. Habitat Int. 138, 102856 (2023).
the emissions of companies that provide the
12. Penning-Rowsell, E. Landsc. Res. 45, 395–411 (2020).
13. de Lima, R. L. P., de Graff-van Dinther, R. E. & s millions of people increasingly use computing power for this technology. Given
Boogaard, F. C. J. Water Clim. Chang. 13, 1185–1203 (2022). generative artificial intelligence (AI) that the clear consensus among climate sci-
14. Qiu, L. et al. Ecol. Indic. 125, 107477 (2021).
models for tasks ranging from search- entists is that the world faces a ‘now or never’
15. Li, J. et al. Geoderma 226–227, 130–139 (2014).
16. Nunn, P. D., Klöck, C. & Duvat, V. Ocean Coast. Mgmt 205, ing the Web to creating music videos, moment to avoid irreversible planetary
105554 (2021). there is a growing urgency about change2, regulators, policymakers and AI firms
17. Flikkema, M. M. B., Lin, F.-Y., van der Plank, P. P. J.,
minimizing the technology’s energy footprint. must address the problem immediately.
Koning, J. & Waals, O. Front. Mar. Sci. 8, 619462 (2021).
The worrying environmental cost of AI For a start, policy frameworks that encour-
The authors declare no competing interests. is obvious even at this nascent stage of its age energy or fuel efficiency in other economic

736 | Nature | Vol 632 | 22 August 2024

 sectors can be modified and applied to of our test set, we hope that private firms will thousands of times larger than those involving
AI-powered applications. Efforts to monitor participate in benchmarking their proprietary only text (see ‘AI’s energy footprint’). Creative
and benchmark AI’s energy requirements — models as consumer interest in the topic grows. industries considering large-scale adoption
and the associated carbon emissions — should of AI, such as film-making, should take note.
be extended beyond the research commu- The evaluation Within our sample set, the most efficient
nity. Giving the public a simple way to make A single AI model can be used for a variety question-answering model used approxi-
informed decisions would bridge the divide of tasks — ranging from summarization to mately 0.1 watt-hours (roughly the energy
that now exists between the developers and speech recognition — so we curated a data set needed to power a 25W incandescent light bulb
the users of AI models, and could eventually to reflect those diverse use cases. For instance, for 5 minutes) to process 1,000 questions. The
prove to be a game changer. for object detection, we turned to COCO 2017 least efficient image-generation model, by con-
This is the aim of an initiative called the and Visual Genome — both established evalu- trast, required as much as 1,600 Wh to create
AI Energy Star project, which we describe ation data sets used for research and develop­ 1,000 high-definition images — that’s the power
here and recommend as a template that gov- ment of AI models — as well as the Plastic in necessary to fully charge a smartphone approx-
ernments and the open-source community River data set, composed of annotated exam- imately 70 times, amounting to a 16,000-fold
can adopt. The project is inspired by the US ples of floating plastic objects in waterways. difference. As millions of people integrate AI
Environmental Protection Agency’s Energy We settled on ten popular ways in which models into their workflow, what tasks they
Star ratings. These provide consumers with a most consumers use AI models, for example, deploy them on will increasingly matter.
transparent, straightforward measure of the as a question-answering chatbot or for image In general, supervised tasks such as ques-
energy consumption associated with products generation. We then drew a representative tion answering or text classification — in which
ranging from washing machines to cars. The sample from the task-specific evaluation models are provided with a set of options to
programme has helped to achieve more than data set. Our objective was to measure the choose from or a document that contains
4 billion tonnes of greenhouse-gas reductions the answer — are much more energy effi-
over the past 30 years, the equivalent of taking “Greater transparency will cient than are generative tasks that rely on
almost 30 million petrol-powered cars off the the patterns learnt from the training data to
road per year. encourage developers of AI produce a response from scratch3. Moreover,
The goal of the AI Energy Star project is models to treat energy use as summarization and text-classification tasks
similar: to help developers and users of AI an important parameter.” use relatively little power, although it must be
models to take energy consumption into noted that nearly all use cases involving large
account. By testing a sufficiently diverse array language models are more energy intensive
of AI models for a set of popular use cases, we amount of energy consumed in responding to than a Google search (querying an AI chat-
can establish an expected range of energy con- 1,000 queries. The open-source CodeCarbon bot once uses up about ten times the energy
sumption, and then rate models depending on package was used to track the energy required required to process a web search request).
where they lie on this range, with those that to compute the responses. The experiments Such rankings can be used by developers to
consume the least energy being given the high- were carried out by running the code on state- choose more-efficient model architectures to
est rating. This simple system can help users to of-the-art NVIDIA graphics processing units, optimize for energy use. This is already pos-
choose the most appropriate models for their reflecting cloud-based deployment settings sible, as shown by our as-yet-unpublished
use case quickly. Greater transparency will, using specialized hardware, as well as on the tests on models of similar sizes (determined
hopefully, also encourage model developers central processing units of commercially on the basis of the number of connections in
to consider energy use as an important param- available computers. the neural network). For a specific task such
eter, resulting in an industry-wide reduction In our initial set of experiments, we eval- as text generation, a language model called
in greenhouse-gas emissions. uated more than 200 open-source models OLMo-7B, created by the Allen Institute in
Our initial benchmarking focuses on a suite from the Hugging Face platform, choosing the Seattle, Washington, drew 43 Wh to gener-
of open-source models hosted on Hugging 20 most popular (by number of downloads) ate 1,000 text responses, whereas Google’s
Face, a leading repository for AI models. for each task. Our initial findings show that Gemma-7B and one called Yi-6B LLM, from the
Although some of the widely used chatbots tasks involving image classification and gen- Beijing-based company 01.AI, used 53 Wh and
released by Google and OpenAI are not yet part eration generally result in carbon emissions 147 Wh, respectively.
With a range of options already in exist-
AI’S ENERGY FOOTPRINT ence, star ratings based on rankings such as
SOURCE: UNPUBLISHED ANALYSIS BY S. LUCCIONI ET AL./ AI ENERGY STAR PROJECT

The power consumed by artificial intelligence (AI) tools varies greatly depending on the task. An AI model that ours could nudge model developers towards
provides answers to queries is much less energy-intensive than one that generates images from text prompts,
lowering their energy footprint. On our part,
for example. And the data show that even AI models of the same type can vary widely in energy consumption.
we will be launching an AI Energy Star leader­
= AI model* = Mean 22 Wh 1,000 Wh
The energy required A battery with this much energy
board website, along with a centralized test-
to fully charge a smartphone could run a laptop for 20 hours ing platform that can be used to compare and
Type of task benchmark models as they come out. The
477
Image generation energy thresholds for each star rating will shift
109 if industry moves in the right direction. That is
Image captioning
why we intend to update the ratings routinely
25
Automatic speech recognition and offer users and organizations a useful met-
1 ric, other than performance, to evaluate which
Question answering
AI models are the most suitable.
0.5
Text classification
The recommendations
0 0.01 0.1 1 10 100 1,000 10,000 To achieve meaningful progress, it is essen-
Total energy consumed (in Wh†) to perform a task 1,000 times tial that all stakeholders take proactive steps
*Tests conducted on 20 popular open-source models. Each dot represents one model; †1 Watt-hour represents power consumption of 1 W extended over 1 hour. to ensure the sustainable growth of AI. The

Nature | Vol 632 | 22 August 2024 | 737

Comment
following recommendations provide some
specific guidance to the variety of players
involved.

Get developers involved. AI researchers and

developers are at the core of innovation in this
field. By considering sustainability throughout
the development and deployment cycle, they
can significantly reduce AI’s environmental
impact from the outset. To make it standard
practice to measure and publicly share the
energy use of models (for example, in a ‘model
card’ setting out information such as training
data, evaluations of performance and meta-
data), it’s essential to get developers on board.

Drive the market towards sustainability.

Enterprises and product developers play a
crucial part in the deployment and commer-
cial use of AI technologies. Whether creating
a standalone product, enhancing existing
software or adopting AI for internal business
processes, these groups are often key decision
makers in the AI value chain. By demanding
energy-efficient models and setting procure-

GETTY
ment standards, they can drive the market
towards sustainable solutions. For instance, Tools to quantify AI’s energy use can improve efficiency and sustainability.
they could set baseline expectations (such
as requiring that models achieve at least two go.nature.com/4dfp1wb). The AI Energy Star The authors
stars according to the AI Energy Star scheme) project is a small beginning and could be
or support sustainable-AI legislation. refined further. Currently, we do not account Sasha Luccioni is the AI & climate lead at
for energy overheads expended on model Hugging Face in Montreal, Canada. Boris
Disclose energy consumption. AI users are storage and networking, as well as data-centre Gamazaychikov is the senior manager of
on the front lines, interacting with AI prod- cooling, which can be measured only with emissions reduction at Salesforce in San
ucts in various applications. A preference for direct access to cloud facilities. This means Francisco, California, USA. Sara Hooker leads
energy-efficient solutions could send a power- that our results represent the lower bound of the research laboratory Cohere For AI and
ful market signal, encouraging developers and the AI models’ overall energy consumption, is based in San Francisco, California, USA.
enterprises to prioritize sustainability. Users which is likely to double4 if the associated over- Régis Pierrard is a machine-learning engineer
can nudge the industry in the right direction head is taken into account. at Hugging Face, Paris, France. Emma
by opting for models that publicly disclose How energy use translates into carbon emis- Strubell is the Raj Reddy assistant professor
energy consumption. They can also use AI of computer science in the Language
products more conscientiously, avoiding “Users can nudge the Technologies Institute at Carnegie Mellon
wasteful and unnecessary use. University, Pittsburgh, Pennsylvania, USA.
industry by opting for Yacine Jernite is the machine-learning and
Strengthen regulation and governance. models that publicly disclose society lead at Hugging Face, and is based
Policymakers have the authority to treat energy consumption.” in New York, New York, USA. Carole-Jean Wu
sustainability as a mandatory criterion in AI is a director of AI research at Meta, based in
development and deployment. With recent Boston, Massachusetts, USA.
examples of legislation calling for AI impact sions will also depend on where the models are e-mail: [email protected]
transparency in the European Union and the ultimately deployed, and the energy mix avail-
United States, policymakers are already moving able in that city or town. The biggest challenge, 1. International Energy Agency. Electricity 2024 (IEA, 2024).
2. IPCC. Climate Change 2022: Mitigation of Climate
towards greater accountability. This can ini- however, will remain the impenetrability of Change. Contribution of Working Group III to the Sixth
tially be voluntary, but eventually governments what is happening in the proprietary-model Assessment Report of the Intergovernmental Panel on
could regulate AI system deployment on the ecosystem. Government regulators are start- Climate Change (eds Shukla, P. R. et al.) (Cambridge Univ.
Press, 2023).
basis of the efficiency of the underlying models. ing to demand access to AI models, especially 3. Luccioni, A. S., Jernite, Y. & Strubell, E. in Proc. 2024 ACM
Regulators can adopt a bird’s-eye view, and to ensure safety. Greater transparency is Conf. Fairness, Account. Transpar. 85–99 (ACM, 2024).
their input will be crucial for creating global urgently needed because proprietary models 4. Luccioni, A. S., Viguier, S. & Ligozat, A. R. J. Machin. Learn.
Res. 24, 253 (2023).
standards. It might also be important to estab- are widely deployed in user-facing settings.
lish independent authorities to track changes The world is now at a key inflection point. The
in AI energy consumption over time. decisions being made today will reverberate
for decades as AI technology evolves alongside The authors declare competing interests; see go.nature.
Taking stock an increasingly unstable planetary climate. We com/3m9smk2 for details.
Clearly, a lot more needs to be done to put hope that the Energy Star project serves as a
Disclaimer: All experimentation and data processing was
a suitable regulatory regime in place before valuable starting point to send a strong sustain- carried out by Hugging Face, and Hugging Face hosts all
mass AI adoption becomes a reality (see ability demand throughout the AI value chain. data related to this research.

738 | Nature | Vol 632 | 22 August 2024