How much energy does a ChatGPT query consume? – NewsBytes
As generative AI becomes more common, questions about its energy consumption and carbon emissions are becoming increasingly important. OpenAI CEO Sam Altman recently claimed that an “average ChatGPT query” consumes energy equivalent to what an oven uses in just over a second. This estimate is within the realm of reason, as AI research firm Epoch AI previously calculated a similar estimate. However, experts believe the statement lacks context, particularly regarding what constitutes an “average” query.
Emission disparity
A study published on June 19 in Frontiers in Communication examined 14 open-source large language models (LLMs), including those from Meta and DeepSeek. The findings revealed that some models emitted up to 50 times more CO2 than others. This data highlights the variability of carbon emissions across different AI models, further complicating the task of estimating an average energy consumption for a single prompt.
Energy consumption
Large language models (LLMs) are known for their high energy consumption due to their numerous parameters. These “internal knobs” are adjusted during training to improve model performance. The more parameters a model has, the more capable it is to learn patterns and relationships in data. For instance, GPT-4 is estimated to have over a trillion parameters and runs on powerful chips called graphics processing units (GPUs) in massive data centers worldwide.
Emission estimation
Training an LLM requires processing vast datasets and adjusting internal parameters, which takes weeks and thousands of GPUs. However, companies rarely reveal their training methods, making emissions from this process largely unknown. Inference, or the model’s response to user prompts, is expected to account for most of a model’s emissions over time. But like training, estimating the environmental impact of a single query is complicated by factors such as data center location and energy grid powering it.
Research findings
While proprietary models from companies like OpenAI and Anthropic are not publicly available, others such as Meta and DeepSeek have released open-source versions of their AI products. Researchers can run these models locally to measure the energy consumed by their GPU as a proxy for inference energy consumption. The June study tested 14 open-source AI models on the NVIDIA A100 GPU and found reasoning models consumed more energy during inference than standard ones due to processing more tokens per question.
Power consumption
On average, reasoning models consume 543.5 tokens per query—far exceeding the 37.7 tokens used by standard models. At scale, the emissions add up. Answering 600,000 questions with DeepSeek R1, a 70-billion-parameter reasoning model, generates as much CO₂ as a round-trip flight between London and New York. The real impact is likely even higher. Currently, data centers account for 4.4% of total US electricity use, including AI needs. That share could rise to as much as 12% by 2028.
Emission factors
The carbon footprint of an AI model is determined by several factors, including the number of parameters it has, its architecture, and how it was trained. However, these details are often not disclosed by companies. The environmental impact of a single query can vary widely depending on which data center it goes to and other variables like the time of day or season.
Action
Choosing the right model for each task can lower environmental impact as most queries don’t need a power-hungry reasoning model. Running AI tasks at night or during cooler months can ease pressure on the grid since cooling needs at data centers are lower. Phrases like “please” and “thank you” must be avoided as they cost millions of extra dollars. Users can also check Hugging Face’s AI Energy Score leaderboard to find a model that balances performance, accuracy, and energy efficiency.
