Does the promise of artificial intelligence outweigh its environmental impact? That is the subject of a recent paper published in the IT Professional journal by Nir Kshetri at the University of North Carolina at Greensborough, which scrutinizes the energy and water consumption AI and generative AI (GAI) models require along with the carbon emissions they produce. But AI can do something other large power consumers (traditional data centers or crypto mining operations) can’t: AI can self-analyze and self-optimize. Can these efficiency gains offset AI’s negative environmental impacts? To find out, it’s necessary to assess the environmental impact of the three AI model development phases: model training, inference, and hardware and infrastructure production.
The Model Training Phase
During this phase, AI engineers conduct numerous iterations of mathematical operations across vast datasets. As a rule, the larger the AI model, the larger its environmental impact. Case in point, Meta’s training phase of its Large Language Model Meta AI (LLaMa) required 449 MWh of power use, but that was nothing compared to the 1287 MW of energy consumed and 552 t of CO2 emitted by Open AI’s Generative Pre-trained Transformer (GPT-3). But that was dwarfed in turn by GPT-4’s five-to-six-month training phase, which required 7200 MWh of energy.
The Inference Phase
Next is the Inference Phase, which focuses on predicting outcomes using new input data post-deployment of the machine learning (ML) model. GAI processes, by their nature, require more power than traditional consumer data center services, like search queries. For example, it’s estimated that Open AI’s ChatGPT consumes 2.9 Wh per request while a typical Google search uses just 0.3 Wh. And generating an image via an AI model requires the same amount of power necessary to fully charge a smartphone.
Production of Hardware and Infrastructure Phase
The production of computing hardware and infrastructure is the final dev phase, and implementing these models in the real world requires powerful CPUs and GPUs that consume significant amounts of water and energy. GPU technology is also evolving rapidly, requiring AI projects to refresh hardware more frequently than traditional data centers, which also contributes to global e-waste. And analysts are predicting hyperscale data center capacity could triple as soon as 2029 just to accommodate the demands of GAI.
Positive Environmental Effects
But it’s not all bad news because AI is already contributing to environmental sustainability. One data center in Germany used AI to manage cooling loads and energy use based on the weather, which resulted in a 9% increase in efficiency. Energy companies are also leveraging AI to optimize operations, including reducing carbon emissions, preventing costly cyber-attacks and anticipating mechanical issues. Shell has also pioneered the development of an AI tool to monitor and optimize methane emissions, reducing the amount of power required for generation.
A recent study by Google also revealed the startling way AI is poised to effect operational efficiency: by processing atmospheric data, AI can help commercial pilots choose flight paths that produce the fewest contrails—and the most intriguing part: if the entire aviation industry adopted this single AI use case, it could save more CO2 equivalent (CO2e) than the amount of CO2e produced by AI in an entire year.
In Conclusion
It’s no secret that AI and ML operations require significant energy and water; however, the pollution-reducing potential of AI has to be considered when assessing its broader environmental footprint.
For an in-depth look at AI’s environmental impact and potential to boost efficiency within the energy industry, download the full paper.
Download Paper
