A common statement about artificial intelligence says that “every search in ChatGPT consumes the same amount of water as a bottle.” It’s a common statement that circulates widely on the internet and can be found in various debates regarding artificial intelligence and its impact on the environment. However, the statement may have some truth in it, yet it’s not entirely accurate.
A major investigation conducted by MIT Technology Review aimed at finding out how much energy artificial intelligence systems consume. The investigation found the power consumption of a single answer from a large language model used in ChatGPT.
The investigation found that “one answer from a large language model might use anywhere from 114 joules of electricity to 6,706 joules of electricity. The lower amount of electricity is equivalent to one-tenth of a second of microwave time. The higher amount of electricity is equivalent to eight seconds of microwave time.”
This wide gap arises because not all AI models work in the same manner. Small models have a smaller number of parameters. Parameters are variables used in the process of information processing. Small models use less power in computation since they need less computation.
However, the disadvantage of using these models lies in their inaccuracy. Small models give less accurate answers compared to large models that use heavy computation resources.
The High-Energy Cost of Creativity, Mapping the AI Power Surge
The energy requirement increases dramatically when the AI system is used for generating images or videos. This is because generating images is much more computationally intensive for the system, considering the processing required for images, colors, and other visual representations. When generating videos, the system has to produce several images and ensure smooth transitions between them.
The MIT Technology Review research on energy consumption by AI systems revealed that generating a five-second video by an AI system requires 3.4 million joules of energy. This is over 700 times the energy required to produce a high-quality image by an AI system. This energy consumption is equivalent to running a microwave oven for over an hour.
The MIT researchers used a simple example to demonstrate the energy consumption by an AI system over a period. Suppose a person uses an AI chatbot to ask fifteen questions, generate ten images, and then create three videos, each five seconds long.
The total energy cost of that activity reaches about 2.9 kilowatt-hours of electricity. That amount equals running a microwave for more than three and a half hours.
While these numbers may seem large, the bigger issue lies in the infrastructure that supports AI systems. Large language models run inside massive data centers filled with specialized chips and cooling equipment. These facilities operate around the clock to handle requests from millions of users.
Before the rapid rise of artificial intelligence, the electricity demand from data centers had remained mostly stable. Advances in server efficiency helped offset the growing demand for cloud computing and digital services.
The situation changed once companies began deploying large-scale AI models. The report shows that electricity consumption from U.S. data centers has doubled since 2017. Artificial intelligence plays a central role in that increase.
Tracking the Growing Carbon Footprint of Generative AI
Government data cited in the investigation suggests that by 2028, roughly half of all electricity used by data centers in the United States will power AI systems. That shift reflects the intense computing requirements behind modern AI models.
The timing of this report also matters. Tech companies now embed generative AI across many digital services. At events like Google I/O, companies highlight how AI will shape the next generation of products.
For example, Google has introduced AI features across Google Search, Gmail, Google Docs, and Google Meet.
At the same time, everyday users rely on generative AI for many tasks. People use it to prepare job interviews, create digital content, write essays, and generate synthetic images or videos.
Each of those actions carries an energy cost. One request may use a small amount of power. Millions of requests each day create a much larger impact.
The MIT Technology Review report does not argue that AI should stop. Instead, it highlights the need to understand the real energy demands behind the technology. As artificial intelligence spreads across digital tools, its environmental footprint will become an important issue for researchers, companies, and policymakers.
