Can Al learn to consume with care?

Big Tech is pumping more than $60 billion into data centres in India to power a smart future. This will also put pressure on scarce power and water resources. Fortunately, there are solutions

- SHELLEY SINGH

The moment you type in a query and press enter on an artificial intelligence (AI) platform to get an image, infographic, text or video, the query goes to a football field-sized data centre where racks of power-hungry Al chips have a quick “think” before replying. This process, called “inference” — where a machine applies its knowledge (or learning, if you will) to answer the query —guzzles up power and water to generate an output.

Most technology companies are reluctant to share exactly how much electricity and water these facilities consume, even as millions make AI their daily assistant.

Sam Altman, chief executive officer (CEO), OpenAl, offered a rare glimpse in a blogpost titled The Gentle Singularity earlier this year, noting that an average ChatGPT query consumes about 0.34 watt-hours of electricity and 0.000085 gallons of water (roughly one-fifteenth of a teaspoon of water). Or, just a tiny drop of water, and about as much power as a microwave might consume in a little over one second — for a simple text query. Videos and images consume more.

That's insignificant at the level of a single prompt. But multiply that by billions of daily queries and tasks — from simple questions to complex tasks, like managing an intelligent robot factory or drug discovery—and you are talking about city-sized power and water consumption to make machines manufacture intelligence.

As it happens, power-hungry and water-deficit India is ChatGPT’s second largest market by user.