“Because they can accomplish a wide range of tasks that were previously thought to be impossible for computers,” Bard said in response to my question about why massive language models are revolutionary. They can interpret languages, produce imaginative writing, and provide insightful answers to queries.”
When OpenAI’s ChatGPT reached 100 million users two months after its November 2022 launch, digital companies sprang into action to provide the public with additional “generative AI.” Commentators compared the influence of the new technology to that of the Industrial Revolution, the Internet, electricity, or the discovery of fire. Although hype will eventually give way to reality, one obvious effect of artificial intelligence’s rapid development is the technology’s significant and expanding environmental impact.
THE COST OF BEING A KNOW-IT-ALL
Although there are growing initiatives to incorporate renewable energy resources, the majority of data centres today function around the clock and are powered by fossil fuels. Data centres are responsible for 2.5% to 3.7% of worldwide greenhouse gas emissions due to their energy consumption, which is more than even the aviation sector.
Processors and chips consume the majority of the energy in a data centre. AI systems use zeros and ones to process information, much like any other computer system. A bit uses a little amount of power and produces heat each time it changes its state from one to zero, which is the smallest amount of data that computers can process. Large air conditioners use about 40% of the power used in data centres since servers need to be kept cold in order to operate. Servers would overheat and malfunction without them.
Data centres used between 0.9% and 1.3% of the world’s power demand in 2021. According to one analysis, by 2030, it may rise to 1.86%. Over the coming years, AI models’ processing and energy requirements will rise in tandem with their skills and complexity. According to one study firm, by 2028, computer performance will have grown fourfold, but processing workloads will have climbed fiftyfold as a result of more use, more complex queries, and sophisticated models with many more parameters. By 2030, data centres’ energy usage is predicted to increase by 28% on the European continent.
After models are put into use, inference—the process by which the AI predicts incoming data and answers questions—may use even more energy than training. According to Google, 40% of the energy needed in AI training and inference goes into training, while 60% goes towards inference. The carbon footprint of GPT-3 was calculated to be 50 pounds of CO2 per day, or 8.4 tonnes of CO2 annually. Because inference is used many times to serve millions of users, it consumes a lot of energy, whereas training is typically done several times to maintain models up to date and optimised. 100 million people were using ChatGPT two months after it launched. Many people are keen to utilise AI for everything instead of using the current online searches that rely on smaller AI models, but according to one tech expert, a single ChatGPT request can use 100 times as much energy as a single Google search.
Researchers from MIT and Northeastern University calculated that inference uses more energy than training; however, the question of whether inference uses more energy is still up for discussion. However, it is inevitable that the carbon footprints of OpenAI, Google, Microsoft, and the Chinese search engine Baidu will increase as they race to develop larger, more complex models and as more people use them. This may make the process of decarbonising our societies far more challenging.
The first infrastructure-aware benchmark, which shows the energy, water, and carbon costs of common AI tools—even for quick queries—was recently published. The businesses that create AI models are eager to provide detailed information on how well they perform on benchmarks. Information regarding their influence on the environment is something they are less willing to reveal. Since there is a lack of precise data, several estimations have been made. A more precise assessment of the environmental impact of AI use is provided by a recent research that was released on Cornell University’s preprint service arXiv.
Scientists from Providence College, the University of Rhode Island, and the University of Tunis in Tunisia made up the study team, which created what they call the first infrastructure-aware benchmark for AI inference, or usage. They computed the per-prompt environmental footprint for 30 popular AI models by integrating public API latency data with hints about the composition of regional power grids and the GPUs working in the background. Data on carbon, water, and energy were then combined to create an “eco-efficiency” score.
The results are striking. For a lengthy response, OpenAI’s o3 model and DeepSeek’s primary reasoning model need over 33 watt-hours (Wh), more than 70 times the energy needed by OpenAI’s more compact GPT-4.1 nano. The researchers assert that Anthropic’s Claude-3.7 Sonnet is the most environmentally friendly, pointing out that hardware has a significant influence on how AI models affect the environment: GPT-4o mini, which uses outdated A100 GPUs, uses more energy per query than the larger GPT-4o, which runs on more sophisticated H100 chips.
Additionally, the study discovered that the environmental cost increased with the length of the inquiry. Even quick searches use a significant amount of energy. One quick GPT-4o prompt consumes around 0.43 Wh. According to the researchers, OpenAI’s predicted 700 million GPT-4o calls each day might result in an annual energy consumption of 392–463 gigawatt hours (GWh), which is sufficient to power 35,000 houses in the United States.
MOVING TOWARDS GREENER GRASS
The energy and environmental costs of artificial intelligence are being considered by several experts and researchers who are working to make it more environmentally friendly. These are just a few examples of how AI may be made more environmentally friendly.
TRANSPARENCY
Enabling developers and businesses to understand how much power their computers are consuming and how that translates into carbon emissions is the first step towards making AI greener. After all, you can’t cure an issue if you can’t quantify it. For developers to assess the effects of various systems and solutions, the measurements of AI carbon footprints must also be standardised. Researchers from Stanford, Facebook, and McGill University have created a tracker to gauge carbon emissions and energy use during AI model training. Additionally, customers may determine the carbon impact of their cloud using Microsoft’s Emissions Impact Dashboard for Azure.
USING RENEWABLE ENERGY
Microsoft claims that by 2030, all of the major cloud providers want to—and some already do—run their cloud data centres entirely on carbon-free electricity. Microsoft has long-term contracts for green energy for several of its campuses, buildings, and data centres, and it is dedicated to operating entirely on renewable energy by 2025. Already, all of the energy used in Google’s data facilities comes from renewable sources.
It also makes a significant impact to move major workloads to data centres where the energy may be obtained from a sustainable energy grid. For instance, 433 MWh of power and 25 metric tonnes of CO2 equivalent were used to train the massive language model BLOOM, which has 176 billion parameters, for the AI firm Hugging Face. It was trained on a nuclear-powered supercomputer in France. In contrast, the 175 billion parameters used to train GPT-3 resulted in 502 metric tonnes of carbon dioxide equivalent emissions and 1287 MWh of power consumption.
EFFICIENT HARDWARE
Some GPUS are slower and use more energy than chips made specifically for training massive language models, like Google’s tensor processing units. Google says that by utilising gear that produces less heat and hence requires less energy for cooling, its data centres have drastically reduced their energy use. Numerous other businesses and researchers are also working to create more effective hardware designed especially for artificial intelligence.
USING APPROPRIATE MODELS
Not every type of activity requires large language models. One option to conserve energy is to employ a smaller AI model for basic tasks; more concentrated models are more effective than models that can do everything. To attempt to discover new medicines, for example, it could be worth the power they need to use huge models, but not to create limericks.
Using data sets that are 1/10,000 of the size of the huge language models, several researchers are attempting to build language models. The goal of the BabyLM Challenge is to use a dataset of words that kids are exposed to to teach a language model how to acquire linguistic subtleties from scratch, much like a human learns. Young children are exposed to 2,000–7,000 words annually; the BabyLM Challenge dataset has a maximum of 100,000 words, which is the same quantity of words a 13-year-old would have encountered. A smaller model uses less energy since it requires less time and resources to train.
MODIFICATION OF CURRENT MODELS
AI might be made more effective and energy-efficient by improving current models rather than attempting to create even larger new models. There are “overparameterised” AI models. Reducing computational expenses and storage requirements may be achieved by pruning the network to eliminate unnecessary parameters that do not affect a model’s performance. Finding methods to lower the number of parameters without compromising accuracy is the aim for AI engineers. Another technique for reducing the size of AI models is knowledge distillation, which involves moving the knowledge the model has acquired from a large network to a smaller one.
Federated learning, which brings the model training to the data rather than the data to a central server, is being studied by Virginia Tech and Amazon. Instead of using a centralised or cloud server, this approach uses data saved on several devices in various places to train certain components of the model. Federated learning saves energy by cutting down on training time as well as the quantity of data that has to be moved and stored. Additionally, data security is ensured because the data on each device remains in its original location. The updated models are transmitted back to a central server and combined into a new, more complete model after being trained on the local devices.
NEW METHODS OF COOLING
A unique fluid designed to boil 90 degrees below the boiling point of water is being used by Microsoft researchers to cool computer chips since conventional cooling techniques, like air conditioning, can’t always keep data centres cold enough. The fluid, which does not damage electronic equipment, is submerged in servers, allowing them to continue running by removing heat from the heated chips. Server power consumption may be reduced by 5% to 15% with liquid immersion cooling, which is more energy-efficient than air conditioners.
Additionally, Microsoft is experimenting with the usage of underwater data centres that employ local wind turbines, ocean currents, and the ocean’s natural cooling to provide sustainable energy. Computers are immersed underwater in a cylindrical container. On land, oxygen, air moisture, and temperature changes may all affect computer performance. An oxygen-free, steady atmosphere is provided by the underwater cylinder. According to researchers, the failure rate of computers underwater is one-eighth that of computers on land.
Data centres could eventually relocate from the cloud to space. A few years ago, a business named Lonestar funded $5 million to construct mini data centres on the moon. Lunar data centres would benefit from plentiful solar energy and be less vulnerable to sabotage and natural calamities. The feasibility of constructing solar-powered data centres in orbit is being investigated by Thales Alenia Space. The goal of the project is to ascertain if building and launching data centres in space would produce lower carbon emissions than those on the ground.
It is up to us all—researchers, tech firms, legislators, and citizens—to promote more environmentally friendly AI practices and capitalise on AI’s potential for sustainability as we traverse this complicated terrain. How will we contribute to the development of a future in which the advantages of AI surpass its negative environmental effects? Although the road ahead won’t be simple, we must rise to the challenge. We can lessen AI’s negative environmental effects by making investments in efficiency improvements, renewable energy, and structural policy reforms. And we can make this instrument an ally of the Earth by using AI to speed up climate solutions.
