Tell me in 2 minutes
Generative AI (GenAI)’s rapid growth is generating increased attention on its environmental impact. Huge amounts of power are required for the training and development of GenAI models. As choice proliferates and adoption grows, the global appetite for GenAI is driving an insatiable hunger for electricity to power it.
The vast majority of AI systems are trained and deployed on power-hungry servers housed in large data centres that use enormous amounts of electricity and water, amplifying the effects of climate change that are already being felt around the world. This article explains the significance of these issues and calls for action to address the energy and climate effects of increasing our usage of and reliance on GenAI.
What is the environmental impact of GenAI?
Energy consumption
The development of generative AI such as OpenAI’s ChatGPT and DALL-E involves an initial training phase followed by an inference phase. Both phases can be extremely energy intensive.
In the training phase, the AI model is fed large datasets which ultimately guide the model’s behaviour. Training a large language model (LLM) like OpenAI’s GPT-3 requires nearly 1,300 megawatt-hours (MWh) of electricity, enough to power 130 U.S. homes annually.[1]
The inference phase sees the AI model generating outputs in the form of live responses based on new data input by users. There is varying data on the electrical costs of inference. Google reported that inference consumed 60% of AI-related energy;[2] whereas Hugging Face indicated the BigScience Large Open-Science Open-Access Multilingual (BLOOM) model consumed considerably less energy in the inference phase than in training.[3] This distribution is ultimately influenced by the differences in different AI models’ retraining frequency, user deployment and model efficiency.[4] While inference on a single entry requires much less computation than training an entire AI model, inference occurs much more frequently than training – as many as 100 billion times a day for a model like Google Translate.[5]
The chairman of Alphabet, Google’s parent company, revealed that interacting with an LLM cost 10 times more electricity than a single Google search, which uses 0.3 Wh of electricity. This is reflected in ChatGPT’s 2023 operating costs, which estimated an electricity consumption of 564 MWh for 195 million requests per day, averaging 2.9 Wh per request. The International Energy Agency (IEA) predicted that electricity demand would increase by 10 terawatt-hours annually if ChatGPT were incorporated into the 9 billion searches done a day, the consumption of 1.5 million European Union residents.[6]
Emissions
Perhaps more concerning is the significant carbon footprint trailing AI models. GPT-3 generated the same volume of CO2 emissions in its training phase as 550 round-trip flights between New York and San Francisco.[7] Training one AI model with neural architecture search is estimated to emit as much CO2 as five cars during their lifetime. This is unsurprising as BLOOM emitted a staggering 19 kg of CO2 emissions daily during its monitoring period.[8]
Given their surge in popularity, generative AI systems are experiencing the fastest overall growth in emissions. The CO2 emissions from 1,000 inferences in the most carbon-intensive image generation model is equivalent to driving 6.5 kms in a gasoline-powered vehicle.[9] This adds up quickly with growing complexity and size of datasets and when deployed by millions of users globally.
The energy costs associated with developing AI models are influenced by both internal and external factors. Internally, larger models demand more energy-intensive graphics processing units (GPUs), data centres and telecommunication networks.[10] Externally, the amount of CO2 emissions is determined by the energy efficiency of data centres and carbon intensity of the electricity grid. A crucial factor is whether high-emission energy or low-carbon sources are used.[11] Powering an electricity grid using non-renewable energy sources like coal, oil and natural gas can result in 60 times more CO2 emissions than using renewable energy sources such as solar, wind and hydroelectricity. Predictably, the primary energy source used to drive the hardware that train the majority of AI models are high-carbon energy sources.[12]
Water consumption
The water footprint of AI models has largely remained under the radar but is equally enormous.
Al uses water in three aspects:
(a) on-site water for server cooling;
(b) supply-chain water for server manufacturing; and
(c) off-site water for electricity generation.
One of the major reasons data centres demand so much energy is to keep the servers cool. Most data centres utilise cooling towers to prevent overheating, which require water to operate. Manufacturing the servers also requires ultrapure water for wafer fabrication, subsequently discharging water that contains toxic chemicals and hazardous wastes.[13]
Following the rising demand for AI, Microsoft’s water consumption jumped by 34% between 2021 and 2022.[14] In fact, there is growing social tension over the competing water consumption between human needs and data centres. In 2023, Uruguayans protested Google’s planned data centre construction in Uruguay amidst its worst drought in 74 years. This is against the backdrop that Google data’s centres consumed 25 billion litres of water for on-site cooling in 2022, the bulk of which was drinking water.[15]
It doesn’t stop there. Not only does electricity generation emit CO2, thermoelectric power plants such as coal and natural gas also require water for each kWh generated. 1 MWh of electricity is expected to exhaust 7,100 litres of water.[16] On this trajectory, the water consumption by AI may reach 6.6 billion litres by 2027, equivalent to 4 to 6 times the annual water consumption of Denmark.[17]
Patterson, D., Gonzalez, J., Holzle, U., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D.R., Texier, M., and Dean, J. (2022). The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. Computer 55, 18–28. https://doi.org/10.1109/MC.2022.3148714.
Luccioni, A.S., Viguier, S., and Ligozat, A.-L. (2022). Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. Preprint at arXiv. https://doi.org/10.48550/arXiv.2211.02001.
Gabriel, I., et al, ‘The Ethics of Advanced AI Assistants’, 2024, p 180. https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/ethics-of-advanced-ai-assistants/the-ethics-of-advanced-ai-assistants-2024-i.pdf
S. Luccioni, S. Viguier, and A.-L. Ligozat. Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model, Nov. 2022. http://arxiv.org/abs/2211.02001. arXiv:2211.02001 [cs].
Luccioni, S. Viguier, and A.-L. Ligozat. Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model, Nov. 2022. http://arxiv.org/abs/2211.02001. arXiv:2211.02001 [cs].
Luccioni, A., Jernite, Y., and Strubell, E. ‘Power Hungry Processing: Watts Driving the Cost of AI Deployment?’ 28 Nov 2023. Page 5. https://arxiv.org/pdf/2311.16863
S. Luccioni and A. Hernandez-Garcia. Counting Carbon: A Survey of Factors Influencing the Emissions of Machine Learning, Feb. 2023. URL http://arxiv.org/abs/2302.08476. arXiv:2302.08476 [cs].
Dodge, T. Prewitt, R. T. D. Combes, E. Odmark, R. Schwartz, E. Strubell, A. S. Luccioni, N. A. Smith, N. DeCario, and W. Buchanan. Measuring the Carbon Intensity of AI in Cloud Instances, June 2022. URL http://arxiv.org/abs/2206.05229. arXiv:2206.05229 [cs].
Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren, Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, https://arxiv.org/abs/2304.03271
Microsoft. Environmental sustainability report, 2022, https://www.microsoft.com/en-us/corporate-responsibility/sustainability/report.
Google. Environmental report, 2023, https://sustainability.google/reports/.
Siddik, M. A. B., et al. (2021). The environmental footprint of data centers in the United States. Environmental Research Letters, 16(6), 064017. doi.org/10.1088/1748-9326/abfba1
What does this mean for us?
For the world
It is clear that the development and maintenance of AI systems is likely to come at a steep cost to the environment. Global data centre electricity consumption has skyrocketed by 20-40% annually in recent years, constituting 2% of the 2022 global energy demand. This number is estimated to double by 2026, reaching the amount of electricity consumption by the whole of Japan.[18]
Recent figures put the contribution from data centres, devices and networks deployed for training and maintaining AI models at 2-6% of global greenhouse gas emissions, although the precise numbers are still debated.[19] In fact, there is considerable uncertainty around how much energy we feed into AI and its carbon footprint due to a lack of reporting.[20]
Human-caused climate change is impacting communities worldwide. The World Health Organisation predicts that climate change could result in an additional 250,000 deaths per year between 2030 and 2050.[21] It is now a matter of urgency that the world reduces global carbon emissions by 45% by 2030 to achieve the net zero goals set out in the Paris Agreement, but we are not on track to meet these targets.[22]
For Australia
The widespread use of AI has seen increasing numbers of data centres being built in Australia to host data. This introduces new environmental stresses in additional locations, primarily on the urban fringe of major cities. If global warming causes a 2°C climb across Australia, this could potentially lead to 50°C days in Melbourne and Sydney and create competition for cooling demands between households and data centres.[23]
The Federal Government has legislated to reduce emissions by 43% by 2030 to reach net zero emissions by 2050 in the Climate Change Act 2022. This obligation will undoubtedly be passed on to the private sector, including through disclosure requirements around emissions minimisation measures and sustainability clauses in government contracts. For example, most businesses across the EU have already made ESG commitments in keeping with the EU’s Corporate Sustainability Reporting Directive (CSRD). Google, Meta and Microsoft have also committed to being ‘water positive’ by 2030.[24] Conversely, being ‘an importer of technology rather than an innovator’, Australia has lagged behind in understanding the impact of AI on energy and water consumption, leading to a lack of action.[25]
What can we do better?
Given the multifaceted impacts of AI, a multidisciplinary approach is required to address its environmental and social impacts.
Green AI
Utilising more sustainable hardware and infrastructure is crucial to minimising AI’s environmental footprint. This involves adopting energy-efficient cloud services and processors optimised for machine learning training.[26]
The public sector can lead by example by prioritising sustainable cloud infrastructure for public projects. Technology companies can also reduce energy consumption without compromising model performance through energy-efficient AI techniques such as model compression, algorithmic optimisations and distributed training.[27]
Importantly, sourcing carbon-free energy to power data centres ensures AI progresses responsibly and aligns with a greener future. This is reflected in the much lower carbon footprint from AI in Norway, where the power grid is largely sustained by renewable energy, compared to the US where the grid is heavily reliant on fossil fuels.[28] We can look at this as a call-to-action for policymakers to enhance access to carbon-free energy for model deployment. A helpful first step may be funding research and development for carbon-free energy technologies that can accelerate grid decarbonation.
Elsewhere in the world, the warm climate in Singapore posed an additional challenge as cooling constitutes 40% of energy usage in data centres. The Singapore Government combatted its water footprint by launching a new standard for optimising energy efficiency, which involves gradually raising data centres’ operating temperatures from 22°C to 26°C. Singapore’s efforts show that achieving ‘green AI’ in Australia may mean allowing servers to operate at higher temperatures so they can dynamically respond to heat events.
Sustainability Reporting
In January 2024, the Treasury released a new draft Bill to implement a mandatory sustainability reporting framework in Australia. In line with this, accountability to environmentally conscious practices can be achieved in many ways.
Model developers can incorporate AI models’ computational efficiency into benchmarking evaluations to support better transparency around the environmental costs of AI and its infrastructure.[29] Workforce training programs can play a crucial role in fostering responsible AI deployment within the AI community. At a national level, policymakers can mitigate uncertainty by supporting the development of better measuring systems for AI’s environmental impacts.
Conclusion
Despite the potential for AI to transform the way we live, work and play, its immense environmental footprint might also make our lives much harder in the long run. We must take steps to address these issues early rather than wait until it’s too late.
The positive news is that recognising this imperative creates opportunity – to invest in and build data campuses capable of sustainably powering AI’s continued development.
International Telecommunication Union. 2020. Greenhouse gas emissions trajectories for the information and communication technology sector compatible with the UNFCCC Paris agreement: L. 1470. http://handle.itu.int/11.1002/1000/14084; Copenhagen Centre on Energy Efficiency. 2020. Greenhouse gas emissions in the ICT sector: Trends and methodologies [Internet]. https://c2e2.unepdtu.org/wp-content/uploads/sites/3/2020/03/greenhouse-gas-emissions-in-the-ict-sector.pdf
Heikkilä, M. (2022, November 14). We’re getting a better idea of AI’s true carbon footprint. MIT Technology Review. technologyreview.com/2022/11/14/1063192/were-getting-a-better-idea-of-ais-true-carbon-footprint/
World Health Organization. COP24 special report: health and climate change. Technical report, World Health Organization, 2018. URL https://www.who.int/publications-detail-redirect/9789241514972.
UNFCCC. Secretariat. Technical dialogue of the first global stocktake. synthesis report by the co-facilitators on the technical dialogue. Sept. 2023.
Lewis, S.C., King, A.D., Mitchell, D.M., 2017. Australia’s Unprecedented Future Temperature Extremes Under Paris Limits to Warming. Geophys. Res. Lett. 44, 9947–9956. https://doi.org/10.1002/2017GL074612
Urs H¨olzle. Our commitment to climate-conscious data center cooling, 2022, https://blog.google/outreach-initiatives/sustainability/our-commitment-to-climate-conscious-data-center-cooling/; Melanie Nakagawa. The journey to water positive, 2023, https://blogs.microsoft.com/on-the-issues/2023/03/22/water-positive-climate-resilience-open-call/; Meta. Sustainability— water, https://sustainability.fb.com/water/.
Gordon Noble, ‘ARTIFICIAL INTELLIGENCE’S SUSTAINABILITY CHALLENGE’ 5 Feb 2024 https://rsv.org.au/ai-sustainability-challenge/
Patterson, J. Gonzalez, U. Hölzle, Q. Le, C. Liang, L.-M. Munguia, D. Rothchild, D. So, M. Texier, and J. Dean. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink, Apr. 2022. URL http://arxiv.org/abs/2204.05149. arXiv:2204.05149
Zhuk A. Artificial Intelligence Impact on the Environment: Hidden Ecological Costs and Ethical-Legal Issues. Journal of Digital Technologies and Law. 2023;1(4):932-954. https://doi.org/10.21202/jdtl.2023.40. EDN: ffvrya
Jesse Dodge, Taylor Prewitt, Remi Tachet Des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A. Smith, Nicole DeCario, and Will Buchanan. 2022. Measuring the Carbon Intensity of AI in Cloud Instances. In 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22), June 21–24, 2022, Seoul, Republic of Korea. ACM, New York, NY, USA, 25 pages. https://doi.org/10.1145/3531146.3533234; Hannah Ritchie, ‘Which countries get the most electricity from low-carbon sources?’. URL https://ourworldindata.org/low-carbon-electricity-by-country; Shannon Williams, ‘AQ Compute's AI-ready data centre launches in Norway’. URL https://datacentrenews.uk/story/aq-compute-s-ai-ready-data-centre-launches-in-norway.
Patterson, J. Gonzalez, U. Hölzle, Q. Le, C. Liang, L.-M. Munguia, D. Rothchild, D. So, M. Texier, and J. Dean. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink, Apr. 2022. URL http://arxiv.org/abs/2204.05149. arXiv:2204.05149
KWM'S TECH REG TOOLS
Australia’s regulatory framework is constantly evolving, as lawmakers try to keep up with rapidly developing technologies that are transforming our economy. Navigate this dynamic tech landscape and stay ahead of developments with KWM’s AI Regulatory map and technology regulatory tracker.
AI REGULATORY MAP
Our map of AI regulation highlights key players who have a voice in the regulation of AI in Australia, with tiering based on our assessment of participant’s interest in AI, the scope of their mandate and assessment of public statements made by them on AI and the potential impact of their initiatives.
TECH REG TRACKER
Our easy-to-use and frequently updated tech reg tracker helps you stay on top of important developments across key areas of tech-related regulation, including AI, with links through to KWM insights or public resources explaining the significance of each development.
If you would like to talk you through regulatory developments as it relates to GenAI, or data and tech more broadly, then please reach out to the experts listed at the bottom of our regulatory map and tracker.
From risk management and mitigation, to decarbonisation and adaptation, our experts from around the world share their knowledge and insights.
