Environmental Costs and Opportunities of AI (2/2)
Energy and Resource Consumption in AI Equipment
Due to AI's reliance on physical infrastructure and advanced semiconductors used in building GPUs and CPUs, the environmental cost of the resources needed by this industry cannot be ignored.
The sourcing of these materials presents significant geopolitical and environmental challenges. Countries like China, Australia, and the Democratic Republic of Congo lead in the production of many critical materials. Meanwhile, nations such as Taiwan, South Korea, and the United States dominate the design and manufacturing of high-end products, especially chips (processors) used for AI development.
 |
| Top global producers by mineral Source: UK Critical Minerals Intelligence Centre |
Key Materials
- Silicon: The primary material for semiconductor wafers.
- Cobalt: Used in the production of magnetic semiconductors and batteries.
- Gallium and Indium: Critical for making high-speed and high-frequency chips (processors).
- Rare Earth Elements: Such as neodymium, dysprosium, and praseodymium, used in various high-tech applications including chip manufacturing.
Extraction and processing of these materials often involve environmentally damaging practices. Mining operations can lead to habitat destruction, water pollution, and significant carbon emissions.
How is AI consuming energy?
Let's start with units of POWER!
 |
| 1000 light bulbs Source: ChatGPT |
Terawatt & Kilowatt-hours
One kilowatt-hour (1 kWh) is a common billing unit for electrical energy supplied by electric utilities, e.g., 1 kWh can power a typical vacuum for one hour.
One terawatt-hour ( 1 TWh) can power the entire state of California for about 1.5 weeks or 10 million light bulbs.Terawatts are often used to describe global or national energy consumption or production. To estimate the energy consumption of AI, let’s start with an example of AI-generated images.
Generating a single image using the DALL-E model consumes about 2.2 kWh (kilowatt-hours) of energy, while Stable Diffusion uses up to 2.9 kWh (Luccioni et al., 2024). On average, data centers use 1.8 liters of water per kWh of energy consumed. The most efficient data centers use as little as 0.2 liters per kWh, while the least efficient can use up to 12 liters per kWh.
Using the average figure, we can estimate the water consumption for generating a single AI image: it takes more energy than a vacuum cleaner running for around 2.5 hours, and it consumes almost 4 liters of water.
 |
| One image using AI tools requires around 5.4 liters of water and consumes more than a vacuum cleaner working around 3 hours. Source: INCO Group |
Although text-based tasks are typically less energy-intensive than image-based ones, they still demand more energy, water, and generate higher carbon emissions than non-generative digital solutions tailored to specific functions.
Journalists from The Washington Post, in collaboration with researchers at the University of California, Riverside, examined the energy and water usage of OpenAI’s ChatGPT, which operates on the GPT-4 language model. Their findings suggest that generating a typical 100-word email consumes approximately 0.14 kilowatt-hours (kWh) of electricity which is roughly equivalent to keeping 14 LED light bulbs lit for an hour.
Global demand for data center capacity is projected to grow by 19 to 22 percent annually between 2023 and 2030 (McKinsey & Company, 2024). This forecast aligns with revised estimates from the International Energy Agency (IEA), which has doubled its projections for data center electricity use, driven largely by the rapid expansion of AI and cryptocurrency from 2022 to 2024. In 2022, data centers worldwide consumed approximately 460 terawatt-hours (TWh) of electricity. By 2026, the IEA anticipates this figure could surpass 1,000 TWh, a level comparable to Japan’s total electricity consumption.
 |
| Global electricity demand from data centres, AI and cryptocurrencies, 2019-2026. Source: IEA |
AI Carbon Emissions
The generation of electricity for a data center, especially if generated through fossil fuel combustion, results in local air pollution, water pollution, and the production of solid wastes, including even hazardous materials. Using renewable energy sources will lower their CO2 (carbon dioxide, the main greenhouse gas) emissions, but not make it zero.
Many companies working on artificial intelligence, including OpenAI (the maker of ChatGPT), do not disclose their emissions. However, some companies like Google and Microsoft, which already report emissions due to regulations or their own environmental commitments, have provided data shedding light on AI's carbon footprint. For example, Google reported a 48% increase in greenhouse gas emissions since 2019, attributing this rise to higher energy consumption in its data centers and supply chain emissions.
Experts warn that incorporating generative AI into specialized services, such as search engines, will not only make these services less energy-efficient but also dramatically scale up their usage. Similarly, Microsoft’s sustainability report, released in May 2024, revealed a 29% increase in emissions since 2020, largely due to the construction of additional data centers designed and optimized for AI workloads.
Training vs Inference
AI models incur two major types of energy costs: training and inference. While training happens once per model and is becoming more resource-intensive due to larger datasets and expanding capabilities, inference, the energy used to run models for user requests, is now the dominant concern as AI adoption grows. Companies manage these costs through usage restrictions, such as limiting access to models like OpenAI’s Sora. Meanwhile, regulatory bodies like the European Union (EU) and International Telecommunication Union (ITU) are advocating for increased transparency and stricter energy efficiency standards in the AI sector.
AI Model Training
Training is the process of teaching an AI model to recognize patterns and make decisions. It involves trial and error but once completed, the model can be deployed. The training process includes:
- Feeding a large dataset to the model
- Iteratively updating model parameters to minimize errors using techniques like backpropagation
- Running complex mathematical computations multiple times across the entire dataset
Energy and Environmental Costs of Training
- Training requires immense computational power, often using large GPU or TPU clusters for weeks or months.
- The electricity consumed can come from fossil fuel sources, resulting in high carbon dioxide (CO2) emissions.
- Training often involves high-end hardware, requiring significant cooling and maintenance.
For instance, training large language models, such as the now outdated GPT-3 model, can use up 1,300 megawatt hours (MWh) of electricity, which is comparable to the annual energy consumption of 130 U.S. households.
AI Model Inference
Inference is the process of using a trained AI model to make predictions or generate outputs based on new input data. For example, it occurs each time we ask a generative AI chatbot a question. Inference involves:
- Forward-passing input data through the model to get results
- Much less computational complexity compared to the training process
Energy and Environmental Costs of Inference
- Inference typically requires fewer resources because it doesn't involve parameter updates, only forward computation
- Processing a single input is relatively low-cost, but large-scale applications like chatbots or recommendation systems used by millions can lead to substantial cumulative energy consumption.
For example, running GPT-3-like models in real-time across millions of queries daily can consume tens of megawatt-hours resulting in tens times the carbon emissions for training GPT-3 such model once.
Localized Environmental Costs of AI
Whether training AI models or hosting them for clients, the computation has to happen somewhere, usually in a data center. This is why the environmental impact of AI tends to be highly localized.
If a data center is located in a hotter region and relies on air conditioning for cooling, it will use much electricity to keep the servers at a low temperature. This leads to higher carbon emission of such data center, if it is not using renewable energy sources. If a data center is relying on liquid cooling and is located in drought-prone areas or close to only source of freshwater, they risk depleting the area of a crucial natural resource.
Energy Grid
The rapid expansion of data centers supporting AI applications is putting significant strain on energy grid infrastructure, particularly in the USA. This is affecting the quality of power supplied to millions of consumers, especially in major data center hubs like Northern Virginia (Bloomberg, 2024).
How to mitigate AI energy demand ?
Data centers always require massive amounts of energy to support other activities, such as cloud computing. Tech companies operating them are leaders in energy optimization and are the biggest buyers of renewable energy, often building their own solar and wind farms close to data centers. Artificial intelligence has only increased their energy consumption, typically outpacing those companies plans to become fully powered by renewables. Some, such as Microsoft and Google, started investing in nuclear power or decided to buy non-renewable energy to fulfil the rapid growth of demand for AI services.
While AI usage might grow, its power consumption might be optimized. Updated regulations and technological improvements, including on efficiency, will be key to moderate currently observed surge in energy consumption from data centers. Many companies are working on new hardware and software optimization for AI applications to use less energy. While the biggest AI companies like OpenAI, Google and Microsoft invest in private power plants, from solar to nuclear, others like Meta and Chinese DeepSeek, build open source models, that can be trained, fine-tuned (modified) and run on less resources, even personal computers.
Key approaches to optimize AI applications from optimizing AI models, to leveraging renewable energy and increasing transparency, are:
- Awareness
- Renewable Energy
- Optimization of AI models
- Hardware Optimization
- Transparency
AI Carbon Calculators
By using these tools, you can estimate the energy consumption of some AI models or tools you are using. Test them out for yourself:
CarbonScaleDown: Browser extension to monitor websites and AI chatbots digital footprint directly in Chrome.
Deloitte AI Carbon Footprint Calculator: This detailed calculator is aimed at businesses to make more informed decisions about AI use. It will help estimate AI carbon footprint on factors such as hardware, models, location, use case and types of tasks.
References
Comments
Post a Comment