Artificial intelligence (AI) is taking the world by storm, growing not only in the number of available systems but also in complexity and ability. To fuel this expansion, a monumental amount of energy is needed to meet the increasing demand for computing power.
But at what environmental cost?
While it’s no secret that AI has a significant environmental footprint, especially considering the energy consumption of data centers that are continuously powered and require extensive cooling, pinpointing its true carbon impact has been challenging. The lack of a standardized method for measurement makes it difficult to quantify the exact impact of AI’s energy demands.
With AI expected to grow by 30-40% annually over the next decade, it is clear this technology is here to stay and will only get bigger. “A few studies have estimated the carbon footprint of individual AI systems, like GPT-3, especially after it became popular,” said Meng Zhang, a researcher at Zhejiang University in China. However, very few have attempted to calculate the combined emissions the world’s major AI systems to understand their collective footprint.
“[If we have this figure], researchers and the public can better understand how AI is [actually] affecting the environment. This would also get the attention of environmental experts, AI developers, and policy makers, helping us move towards using AI in a more sustainable way.”
Putting a number to AI
In their study, Zhang and his team calculated the carbon emissions emitted between 2020 and 2024 from 79 well-known AI systems, including Gemini Ultra, GPT-4, Mistral Large, and Inflection-2.
Central to their estimations was calculating the energy use of graphics processing units (GPU) — a computer chip originally designed to render graphics but now used in the training of AI systems, like deep neural networks, where large amounts of data need to be processed simultaneously. As such, GPUs are the main consumers of energy in these systems.
The team made estimations of carbon emissions based on single training runs and then extrapolated these to calculate the total emissions from both training and using the AI systems. What they found was that collectively, the top 20 AI systems they included in their study consumed enough energy to rival a small country, like Iceland or Democratic People’s Republic of Korea as examples. In fact, in 2022, the carbon emissions from these AI systems surpassed the emissions emitted by 137 individual countries.
The team also predicted that the projected total carbon footprint from the AI systems could reach up to 102.6 Mt of CO2 equivalent per year — similar to the emissions from 22 million people over the course of a year.
While alarming, Zhang says these estimations may just be the tip of the iceberg considering the rate at which AI is expanding. For example, ChatGPT-3.5 has roughly 175 billion parameters — internal variables and values that the model uses to make predictions or generate outputs. The more parameters a model has, the more complex and powerful it is, but it also requires more computational power to train.
ChatGPT-4 on the other hand may contain up to 1.8 trillion parameters and its emissions are estimated to be twelve times more than ChatGPT 3.5, according to Zhang.
Can we curb AI’s environmental impact?
It seems paradoxical to say we can use AI sustainably while it continues to grow so rapidly.
While Zhang points to renewable energy is a promising solution, it’s not yet scalable enough to meet the rapidly increasing energy needs of a growing population, let alone AI and other related industries. Renewable energy infrastructure still faces challenges in terms of global adoption and efficiency.
Zhang and his team have also framed their analysis in the context of carbon taxes, estimating that if they were to be applied to AI’s energy use, could incur costs of roughly $10 billion USD per year.
“To catch policymakers’ attention, we converted our carbon emission data into financial figures using carbon taxes,” he explained. “We believe this will really get their notice, especially given the pressing need for sustainable development [globally].”
This could potentially incentivize companies to reduce their carbon footprints by making polluting activities more expensive. However, the impact of carbon taxes would depend on their scope and how effectively they are implemented. For example, companies might be driven to use cleaner energy sources or make AI systems more energy-efficient to avoid penalties. But the effectiveness of carbon taxes could be limited if renewable energy sources remain insufficient or if companies can pass on the costs to consumers.
Additionally, without global cooperation, there’s a risk that emissions could simply shift to regions with laxer environmental regulations. Ultimately, while carbon taxes could help mitigate the environmental impact of AI, they are not a complete solution on their own and would need to be part of a broader set of policies and technological advancements to address the issue comprehensively.
Though these are pressing issues that need consideration, quantifying the impact of AI systems is an important first step. Zhang believes their method of calculation is a fair one but says more work is needed.
“Given the potential financial impacts of future policies, like carbon taxes, we need even more precise methods,” he said. “This includes deciding on clear boundaries for what we are calculating, like whether to include the production of AI hardware, understanding exactly how much energy is used during the AI’s inference process, and factoring in the carbon emissions from the type of electricity used.”
A double-edged sword
More transparency is needed regarding how such energy demands and carbon emissions are calculated, and the current study’s calculations may only be an approximation.
For example, the team have used a default value of carbon intensity of electricity production double that of the world average as estimated by Our World in Data. Together with their assumption that ChatGPT growth rate is representative of the entire AI sector and that carbon emissions from inference are 1000 times that of training-their predicted emissions may be overestimated.
It will take time to gather the correct data and may require cooperation from the companies running and developing such systems.
One thing is certain: AI is here to stay, and that’s not necessarily a bad thing given the benefits it brings to society. Like any technology, it has its pros and cons. For one, says Zhang, it could help in our efforts to become a more sustainable society.
“AI has the potential to significantly benefit the environment by enabling more efficient resource management, advancing renewable energy technologies, and improving climate change modelling,” said Zhang. “It is a double-edged sword.”
Reference: Yu, Y, et al., Revisit the environmental impact of artificial intelligence: the overlooked carbon emission source? Front. Environ. Sci. Eng (2024) DOI: 10.1007/s11783-024-1918-y
Feature image credit: Shubham Dhage on Unsplash
