AI answers 2.5 billion questions a day — and its power use will soon surpass South Korea's

"Now I even ask ChatGPT what to have for dinner"

Jang Jun-young, 28, an office worker living in Gwangjin-gu, Seoul, has been paying for a ChatGPT subscription for less than three months — but he already considers it indispensable. He relies on it not just for work, but for everyday decisions.

"I start the morning by asking about Saturday's weather and what to wear, then get meal recommendations, and sometimes talk through things that are on my mind," Jang said. "It never seems annoyed no matter how much I chat with it, so I use it without any hesitation."

AI has become what many now regard as an essential utility of modern life, occupying not just the workplace but the most routine corners of daily existence.

In fact, <style ref="s0">ChatGPT, the world's most widely used generative AI, processes roughly 2.5 billion queries a day.</style> The problem is the staggering resource consumption that comes with that scale.

Power consumption stands out above all other concerns. Data centers worldwide — driven largely by AI operations — already consume 448 terawatt-hours of electricity a year, a figure that would rank them 11th among the world's nations if counted as a single country.

<style ref="s1">By 2030, that figure is projected to double, surpassing South Korea's total national electricity consumption — reaching roughly 1.6 times the amount South Korea uses in a year.</style>

Put another way, generating that power will require ever-greater resource inputs. <style ref="s2">Voices are growing louder calling for more attention to the environmental side effects of AI's rapid advance.</style>

A report released June 3 by the United Nations University Institute for Water, Environment and Health (UNU-INWEH), titled "Environmental Costs of AI Energy Use: Carbon, Water and Land Footprints," found that global data center electricity consumption stood at approximately 448 terawatt-hours (TWh) as of 2025.

Treating the world's data centers as a single nation would make them the 11th-largest electricity consumer on earth. The more alarming issue is the pace of growth: <style ref="s3">if current trends continue, global data center power consumption will exceed 945 TWh by 2030</style> — enough to rank sixth in the world.

<style ref="s4">With South Korea's annual electricity consumption hovering around 600 TWh, data center power demand alone is on track to reach 1.6 times South Korea's total by 2030.</style> The primary driver is the surge in generative AI usage — researchers estimate that AI workloads alone will account for 378 TWh of that demand.

Until now, training AI models was widely understood to be the main source of energy consumption. But the report projects that as AI use expands, answering billions of queries every day — the so-called "inference" stage — could account for 80 to 90 percent of total AI energy use.

UNU estimates that ChatGPT processes roughly 2.5 billion prompts a day. Assuming an average of 0.42 watt-hours per text request, the report calculates that the chatbot alone requires approximately 383 gigawatt-hours of electricity a year. The energy demands grow sharply for more complex tasks: AI image generation consumes 60 times more power than a short text response, and 1,450 times more than basic text classification.

The deeper problem is not simply the volume of electricity consumed. <style ref="s5">The real concern is the breadth of resources required to generate that power — carbon emissions from fossil fuel-based electricity generation being the most obvious example.</style> Even a shift to renewable energy does not eliminate the burden, the UNU researchers warn, as it brings its own costs in water use and land occupation.

According to the report, <style ref="s6">switching from coal to bioenergy can cut the average carbon footprint by 72 percent — but it dramatically increases water and land burdens. Bioenergy's water footprint is on average more than 30 times larger than coal's, and its land footprint is 100 times greater.</style>

The finding illustrates how an energy source that appears clean can drive up other environmental costs once water use and land occupation are factored in. Brazil offers a telling example: the country generates a large share of its electricity from hydropower, which keeps its power-sector carbon emissions roughly 77 percent below the global average. Yet its water and land consumption per unit of electricity is three times the world average.

Current assessments of data center sustainability tend to focus primarily on renewable energy use and carbon emission reductions. Against that backdrop,<style ref="s7"> calls are growing for sustainability evaluations of AI data centers to be broadened to include water and land use as standard metrics.</style>

"Jointly assessing carbon, water and land footprints is essential," the report said. "Systems with a high share of renewable energy can reduce emissions but may generate a larger land footprint."

<style ref="s8">Regional inequality is also flagged as a side effect of data center expansion. The benefits of AI development are shared by users and companies around the world,</style> but the burden on power grids, water supplies and land falls disproportionately on the communities that host data centers. In some cases, local residents may end up bearing the infrastructure costs required to keep those facilities running.

Recent research has also found that beyond their enormous electricity consumption, data centers can trigger "heat island" effects, raising temperatures in surrounding areas by as much as 9.1 degrees Celsius — meaning that simply having a data center nearby can force local residents to live in a hotter environment.

Electronic waste is another growing concern. The AI arms race is accelerating the replacement cycle for high-performance servers and semiconductors: equipment is discarded to support larger models and faster inference, adding to the e-waste burden. <style ref="s9">The report projects that AI infrastructure could generate up to 2.5 million metric tons of electronic waste annually by 2030</style> — equivalent to scrapping roughly 250 Eiffel Towers every year. If that waste is not properly recycled, heavy metals could leach into soil and water supplies.

In response, the report called on AI developers and service providers to standardize and publicly disclose their resource consumption, including the electricity used for model training and inference. It also called for AI models to be designed for energy efficiency from the outset.

"Most current assessments focus solely on carbon emissions from the training process," the researchers said. "We need to require that environmental disclosures for AI cover the full training and inference pipeline and present carbon, water and land footprints in standardized units."