Most people don’t think twice before asking ChatGPT to summarize an email or generate an idea. It’s quick, it’s easy, and it feels weightless. But the generative AI environmental impact behind that single query is anything but.

Researchers have estimated that one ChatGPT query consumes about five times more electricity than a simple web search. Multiply that by millions of daily users, and the numbers start to look very different from the clean, frictionless experience on your screen.

This isn’t a reason to stop using AI. But it is a reason to understand what’s actually happening behind it.

The Buildings Powering Your Prompts

Every time someone uses a generative AI tool, the request travels to a data center. These are temperature-controlled buildings packed with servers, storage drives, and networking equipment. They have been around since the 1940s. But the rise of generative AI has pushed them into a completely different league.

A typical computing workload is already energy-intensive. A generative AI training cluster, however, consumes seven or eight times more energy than that. Power requirements of data centers in North America nearly doubled in just one year. They jumped from 2,688 megawatts at the end of 2022 to 5,341 megawatts at the end of 2023. By 2026, global data center electricity consumption is expected to approach 1,050 terawatt-hours. That would make data centers the fifth largest electricity consumer in the world, sitting between Japan and Russia.

To put that in perspective, training GPT-3 alone consumed enough electricity to power around 120 average US homes for a full year. It also generated approximately 552 tons of carbon dioxide in the process.

“The demand for new data centers cannot be met in a sustainable way,” says Noman Bashir, a Computing and Climate Impact Fellow at MIT. “The pace at which companies are building new data centers means the bulk of the electricity to power them must come from fossil fuel-based power plants.”

It’s Not Just Electricity

The generative AI environmental impact doesn’t stop at power bills. Data centers also consume enormous amounts of water, and that side of the story gets far less attention.

Water cools the hardware inside data centers by absorbing heat from servers and other equipment. For every kilowatt hour of energy a data center consumes, it needs roughly two liters of water for cooling. That adds up fast across thousands of facilities running around the clock.

“Just because this is called cloud computing doesn’t mean the hardware lives in the cloud,” says Bashir. “Data centers are present in our physical world, and because of their water usage, they have direct and indirect implications for biodiversity.”

Then there is the hardware itself. GPUs, the powerful processors that handle generative AI workloads, carry their own environmental cost. They require more complex manufacturing processes than standard chips. Their production involves toxic chemicals, energy-intensive fabrication, and raw materials that often come from environmentally damaging mining operations. Shipping them around the world adds more emissions on top of that.

In 2023, the three major GPU producers, NVIDIA, AMD, and Intel, shipped 3.85 million units to data centers. That number has grown every year since.

So What Do We Do With This?

None of this means generative AI is purely destructive. The technology has real benefits, from improving productivity to accelerating scientific research. But the generative AI environmental impact is real too, and right now the industry is moving faster than its ability to measure or manage it.

Part of the problem is invisibility. AI interfaces are designed to feel effortless. There are no emissions counters, no energy meters, no reminders that each query has a cost. “The ease-of-use of generative AI interfaces and the lack of information about the environmental impacts mean that, as a user, I don’t have much incentive to cut back,” says Bashir.

Companies release new AI models every few weeks. Each new model is typically larger than the last, which means more energy to train. The previous version becomes obsolete, and all the energy spent training it goes to waste. It is a cycle that currently has no clear end point.

Elsa Olivetti, professor at MIT’s Department of Materials Science and Engineering, puts it plainly. “We need a more contextual way of systematically and comprehensively understanding the implications of new developments in this space. Due to the speed at which there have been improvements, we haven’t had a chance to catch up with our abilities to measure and understand the tradeoffs.”

The industry is not beyond fixing. But fixing it starts with acknowledging the cost. And right now, most people using AI every day have no idea what that cost looks like.