As artificial intelligence moves from research labs into the daily lives of billions, its physical footprint is becoming impossible to ignore. AI is not “just code”; it is a transformative force built on concrete, copper, silicon, and vast amounts of electricity. To understand why AI uses so much energy, we must look at the two distinct phases of its lifecycle: training and inference.
1. The “Hidden Engine”: Energy Intensity of AI Training
Training a frontier AI model is a massive, months-long industrial process. It involves processing billions of data points—words, images, or code—using high-performance accelerators like GPUs and TPUs that draw significantly more power than standard computer chips.
- Scaling Up: OpenAI’s GPT-4 likely consumed between 50 and 70 GWh of electricity during its training phase.
- The Comparison: This is roughly 40–55 times more energy than its predecessor, GPT-3.
- Human Impact: The energy used to train GPT-4 alone is equivalent to the annual residential electricity consumption of 460,000 people in Sub-Saharan Africa.
2. The “Daily Load”: The Inference Phase
While training makes headlines, the inference phase—the process of AI generating responses to user prompts—is the dominant energy consumer. Inference is estimated to account for 80% to 90% of total AI energy use.
As of 2025, ChatGPT alone processes an estimated 2.5 billion prompts per day. While a single text query uses a relatively small amount of power (roughly 0.42 Wh), the sheer global scale translates into an infrastructure-level load of approximately 383 GWh per year.
3. Modality Matters: Why Video and Images Cost More
The type of task you ask an AI to perform dramatically shifts its energy appetite. As users move from text to more complex media, the energy demand explodes:
- Text vs. Image: Generating a single AI image requires roughly 1,450 times more energy than simple text classification (like a spam filter).
- The Video Frontier: Video generation is the most energy-intensive frontier. A single high-complexity AI video can draw as much electricity as 200,000 spam filter operations.
- Real-World Equivalence: Generating one typical AI image uses enough energy to power a 10-watt LED bulb for 17 minutes; a complex video can power that same bulb for 42 hours.
4. The Data Center Backbone
Data centers are the physical engines of AI. In 2025, global data centers consumed an estimated 448 TWh of electricity. If data centers were a country, they would have ranked 11th globally in electricity consumption, ahead of Saudi Arabia. By 2030, this demand is projected to double, potentially accounting for 3% of all global electricity use.
5. Beyond Carbon: The Triple Footprint
The cost of AI energy isn’t just measured in kilowatts. The report from UNU-INWEH highlights that every unit of energy carries a water and land footprint:
- Water Consumption: Data centers require billions of liters for cooling. In 2025, their water footprint was 4.5 trillion liters—enough to fill 1.8 million Olympic-sized pools.
- Electronic Waste: By 2030, AI infrastructure could generate 2.5 million metric tons of e-waste annually, the equivalent of discarding 250 Eiffel Towers every year.
The Way Forward: Efficiency-by-Design
Technological gains in efficiency are happening, but they often trigger the Rebound Effect (Jevons Paradox): as AI becomes more efficient and cheaper, people use it more, which can increase total energy consumption.
To ensure a sustainable future, the international community is calling for transparency, efficiency-by-design, and “fit-for-purpose” AI use—choosing the lightest model possible for the task at hand. Responsibility must extend across the entire lifecycle, from the minerals used in chips to the recycling of obsolete hardware.
Reference: UNU
