How Much Energy Does Artificial Intelligence Really Consume?

Artificial intelligence has become part of our everyday lives, from simple queries in chats to complex analyses. But behind every answer lies a question that concerns experts: how much energy does it all take? According to Sam Altman, CEO of OpenAI, an average ChatGPT query consumes about 0.34 watt-hours of energy. That is roughly as much as an oven would consume in just over a second or an energy-efficient light bulb in a few minutes. But with 800 million weekly active users, this amounts to enormous cumulative consumption that is growing faster than many realize.

However, this figure has raised doubts. Sasha Luccioni, climate lead at Hugging Face, calls it unreliable because OpenAI does not provide details on how it arrived at the figure. What exactly does an "average query" mean? Does it include image generation? What about the energy used to train models or cool servers? Without this information, it is difficult to take the number seriously. Research from 2025 shows that the data centers where AI runs could consume as much as 20% of global electricity by 2030–2035, while in the US they already accounted for 4.4% of national consumption in 2023, and that figure could triple by 2028.

Why Are Companies Staying Silent?

Large companies such as OpenAI and Anthropic keep their models proprietary, which means external researchers cannot verify their energy consumption. An analysis by Sasha Luccioni and her colleagues, submitted for review in 2025, shows that in May 2025, 84% of large language model (LLM) usage involved models for which no information about environmental impact was available. Users are therefore choosing tools with unknown carbon footprints, which is alarming at a time of climate crisis.

Luccioni compares it to buying a car: you know how much fuel it consumes per 100 km, but no such metrics exist for AI. Instead, unverified estimates circulate, such as one from John Hennessy, chairman of Alphabet (Google's parent company), who said in 2023 that a ChatGPT query consumes 10 times more energy than a Google search. This figure is repeated in media and policy reports even though it comes from an unverified source. Research from 2025 confirms that complex queries can produce up to 50 times more CO₂ emissions than simple ones, and more accurate models require more energy.

What Do the Data Tell Us?

To arrive at more realistic figures, scientists are turning to open models whose consumption can be measured. A study published in Frontiers of Communication in 2025 examined 14 open large language models, including two Meta Llama models and three DeepSeek models. The results showed that some models consume up to 50% more energy than others when answering the same queries. The researchers submitted 1,000 benchmark questions from fields such as high school history and philosophy—half were multiple-choice questions with one-word answers, while half were open-ended questions requiring longer responses.

Models focused on complex reasoning generated more "reasoning tokens"—internal steps involved in processing an answer—which increases consumption. These models were more accurate on complex topics but struggled with brevity: even when instructed otherwise, they returned longer answers. The study's lead author, Maximilian Dauner, a doctoral student at Munich University of Applied Sciences, suggests routing simple queries to less demanding models that achieve good results with lower emissions. Companies such as Google and Microsoft already do this in some search features to provide faster, more energy-efficient answers.

Other Factors: Hardware, Infrastructure, and the Real World

Energy consumption is not just about the model itself—it depends on the hardware and operating conditions. Maximilian Dauner's study ran on Nvidia A100 GPUs, but the newer Nvidia H100, designed for AI, is even more energy-intensive. Data centers require cooling, lighting, and networking equipment, all of which add to consumption. These centers often operate in daily cycles, with lower activity at night, and are connected to different power grids—some reliant on fossil fuels, others on renewable energy.

Noman Bashir, a fellow in computing and climate impact at MIT, compares studies that fail to account for these factors to testing a car's fuel consumption with no load. Moreover, training and updating models consume enormous amounts of energy, figures that companies such as OpenAI keep secret. Research from 2025 shows that although efficiency measures such as AI-optimized cooling (Google DeepMind, for example, reduced consumption by 30%) or server virtualization lower costs, overall consumption is increasing due to the "rebound effect"—greater efficiency leads to greater use.

A Chance to Reduce Emissions?

Although AI increases emissions, it can also reduce them. New research from 2025 suggests that if AI is applied effectively in sectors such as energy, transportation, and food production, it could reduce global emissions by 3.2 to 5.4 billion metric tons of CO₂ equivalent annually by 2035. That would outweigh its own consumption. In the energy sector, AI improves the efficiency of renewable energy sources and grid management, with savings of 10–60% in some cases.

Sasha Luccioni is calling for mandatory reporting of carbon figures for all AI systems. Without transparency, we remain in the dark about this technology's true impact on the planet. It is time for companies to open their books and help us understand whether AI will be a drain on or a savior of our planet.