ai image of a classroom in front of a data centre
|

Generative AI Power & Water Use: Promise or Peril

Randy LaBonte

Image generated using Gemini

The Problem

Many students, in my University of Victoria edtech course (EDCI 336), expressed a high degree of concern about the power consumption of GenAI server farms. They cited publications such as the 2024 paper, The Climate and Sustainability Implications of Generative AI, indicating that GenAI clusters consume seven or eight times more energy than a typical computing cluster’s workload. While server farms are not exclusive to managing GenAI computations, they also support both reactive (e.g. FaceID, voice commands) and predictive (e.g. video/audio algorithms) AI transactions. In addition, server farms host video and audio streaming, cloud computing, and data storage, among other data processes.

I wondered, given the digital boom and the fears that GenAI is gobbling up increasing amounts of power and water that is needed to cool server farms, what else consumes that power and water? Since we were studying the use of GenAI, I set out to find out what our Internet scrapers could find out about power and water use. I used three different AI platforms to see if there were differences.

The Prompt

I selected three platforms to use: ChatGPT 5.2, Gemini 3, and Claude Sonnet 4.5. I used the following prompt:

  • Please calculate the power cost and water use for the average person’s use for one hour of the following items.  Please put them in a chart format with columns for the item, power used (kilowatt hours), water used (in litres), and environmental impact. Items: Generative Artificial Intelligence, Predictive Artificial Intelligence, Video streaming (including YouTube), Audio streaming (including music and podcasts), Gas-powered car, Electric-powered car, electric hot water tank, dryer, and a computer. Please adjust for a Canadian context and add a “relative impact” rating of low, medium, or high.

The outputs and data results provided can be summarized as follows.

Relative Power and Water Impact Ratings:

  • High: Gas car, electric hot water tank
  • Medium-High: Dryer
  • Medium: Generative AI (Gemini rated it as Low)
  • Low: Predictive AI, streaming services, computer (ChatGPT rated streaming as Medium)

Note: Each GenAI rated the power and water consumption of electric cars differently. As such, it was difficult to compare differences, so the data was not included in the summary.

Key Insights (from Gemini):

  1. AI vs. Daily Life: While Generative AI is significantly more intensive than a standard search or Predictive AI, one hour of “chatting” uses less energy than five minutes of driving a car or one load of laundry. The environmental concern with AI is its scale (billions of users), not individual hourly use.
  2. The AI Divide: Generative AI (like ChatGPT) is roughly 10x to 30x more energy-intensive than traditional Predictive AI (like a Google Search or a weather forecast) because of the massive GPU clusters required to “generate” new content.
  3. The “Invisible” Water: Even if you aren’t “using” water, your electricity does. Running a 4000W clothes dryer for an hour effectively “consumes” dozens of litres of water at the power plant for cooling and steam generation.
  1. Cost Context:
    • Generative AI: Costs roughly $0.01 in electricity per hour of active use.
    • Gas Car: Costs roughly $15.00 – $18.00 in fuel per hour of highway driving.
    • Electric Car: Costs roughly $3.00 of electricity per hour of highway driving.
  2. Streaming: Choosing Audio (Podcasts) over Video (YouTube) reduces your digital carbon and water footprint by nearly 85%. (Gemini rated video streaming & GenAI medium)
  3. Reducing hot water use and driving less yields much larger environmental benefits than optimizing digital behaviour alone.

To use AI or not: What is the question?

Use of artificial Intelligence began in earnest eight decades ago with the advent of the mainframe computers for data processing, followed soon after by the personal computer and, at the turn of the century, the Internet. Generative Artificial Intelligence (GenAI) became available to the public just three years ago and is being integrated into other AI types and education apps we have used for years. Finally, GenAI is quickly being rivalled by the development of Artificial General Intelligence.

Some questions to ponder:

  1. Do we have any choice about using GenAI?
  2. Can we avoid its use?
  3. Can the environmental impacts be mitigated?
  4. Will the tech giants solve the power and water overconsumption (see US “Ratepayer Protection Pledge“)?
  5. Will Elon Musk save us with an orbital data centre solution?
  6. Does whether I use GenAI or not really make a difference?