The Energy Dependence of AI

Envision a world where an algorithm can tackle food security challenges by enhancing crop yields, you can entrust a bot with planning an entire getaway, and poets can create visual art without needing a paintbrush. This vision is part of the exciting potential that generative AI holds.

A New Era in Technology

Since the introduction of the concept in the 1950s, artificial intelligence has made remarkable strides. Recently, generative AI—algorithms like ChatGPT capable of independently executing tasks and generating audio, code, images, tests, simulations, and videos—has surfaced as an exciting resource for both professional and personal applications.

Currently, generative AI is utilized for running customer service bots, facilitating monotonous tasks like data entry, adjusting document tones, generating medical leave notes for your employer, writing software, and aiding in diagnostic processes across fields like radiology, medical imaging, and oncology.

However, this production comes at a price that extends beyond just financial costs.

There is increasing evidence that AI significantly affects our energy systems and poses challenges to climate change mitigation, as evidenced by its energy demands, water usage, and greenhouse gas emissions.

“From the very start of the supply chain lifecycle, this impact is felt,” states Shaolei Ren, a researcher at the University of California, Riverside, focusing on the principles of responsible AI aimed at achieving a sustainable, resilient, and fair future. “The manufacturing phase carries a far greater environmental burden compared to the usage phase.”

Understanding the Burden of Development

The creation of interfaces like ChatGPT-3, Meta’s Llama models, and BERT (examples of large language multimodal models under the generative AI umbrella) requires that computers analyze extensive amounts of text produced by humans.

For instance, a large company referenced by the World Health Organization faced an estimated energy requirement of 3.4 GWh over a two-month period for this analysis. This amount corresponds to the annual energy consumption of 300 households in the U.S.

Additionally, the development of GPT-3 resulted in the emission of 552 tons of carbon dioxide—the equivalent produced by 123 passenger cars running for a year.

Equally concerning is the substantial quantity of water needed to manage the heat generated from these computational processes.

The Costs to Users

With each interaction with an AI system, energy consumption escalates. A study by Ren estimated that GPT-3 uses the equivalent of one 500 mL bottle of water for every 10 to 50 generated responses, and ChatGPT-4 is likely to require even more.

Ren explains that the water used by generative AI is not comparable to domestic use, such as during a shower, where water is often recycled back into sewage. Instead, this water consumption reflects the loss of water through evaporation into the atmosphere, which could exacerbate water inequities across regions.

The local environmental effects of AI are significant, adds Ren. For example, a single data center in one Oregon city accounted for over 25 percent of the city’s total water consumption.

Similarly, carbon emissions can present localized issues. Although some AI systems harness renewable energy, the majority still rely on fossil fuels for their power needs.

Envisioning a Sustainable Future

As the proliferation of data centers escalates alongside new iterations of AI models, forecasts suggest that this energy demand is set to increase.

This situation poses challenges for companies striving to develop technologies sustainably, but it also highlights the need for embedding AI’s environmental repercussions into our understanding of its future trajectory.

Consider the following questions:

With the energy needed for ongoing AI development, are we placing ourselves at a risk of resource shortages, or could we face challenges in other domains?
If we prioritize renewable energy for AI initiatives, are we prepared to increase fossil fuel usage in other sectors?
Should we restrict AI deployment to critical functions, limiting personal inquiries to ensure its effectiveness in research, data analysis, and healthcare?
How transparent should corporations be about their AI’s carbon footprint, water usage, and overall environmental impact? Is there a need for legislation mandating disclosure?