The Energy Impact of LLMs — Where Are We in 2025?

Last year, I wrote a post on the carbon impact of LLMs, a rather humbling investigation. I found that whilst training an LLM is significant in terms of energy and carbon impact, inference can quickly overtake training due to the significant number of daily queries a chatbot such as ChatGPT receives.

Moreover, I found out that the task itself matters the most in terms of energy impact. Simple text-based queries could use as little as 0.047 kWh, but image generation used as much as 2.907 kWh, an increase of 60x! This led to the headline fact that the least efficient image generation model used as much energy as 522 smartphone charges. This means, per inference, it uses as much energy as half a smartphone charge [1]!

In this post, I highlight where we are currently with the carbon impact of LLMs. This mainly comes in the form of slightly more transparency from developers and additional information on newer tasks, such as video generation. However, as I will show, these figures that are published hold less weight than we would like. Mainly due to the lack of consistency across developers.

Updated Statistics

In my previous blog post, I did not note any specific statistics about the energy impact of ChatGPT. This was simply because none existed, other than estimates. Thankfully, progress has been made in this area, with official environmental estimates published by OpenAI and Google.

OpenAI states that the “average query” uses around 0.34 Wh of energy, and Google states the “median Gemini Apps text prompt” uses 0.24 Wh of energy [2]. However, the main issue with the reporting of these numbers is that they are not comparable. Each company uses different methods and assumptions to estimate the energy usage. For example, Google uses a custom methodology, but OpenAI does not specify the methodology. Moreover, we have no information about what “medium” or “average” query means. This makes the statistics published unspecific and, at worst, misleading, analogous to greenwashing. When AI companies control both the model and the energy reporting, the environmental impact can be as honest as they decide.

It is known that a longer query will be more energy-intensive. Feeding multiple book sources into Gemini will require significantly more energy than a simple sentence [3]. For example, a 1000-input prompt with 1000 output tokens uses approximately 3x more energy than a 100-input prompt with 300 output tokens [4].

Without allowing third-party validation, with consistent methodology and specific query details, we cannot reliably compare models. This means that currently, it is impossible to accurately compare these models. While this may seem like a non-issue, having the ability to compare the energy impact of models enables consumers to make the most environmentally conscious choice. It also encourages developers to compete to create the least energy-intensive and carbon-emitting model.

Regardless of the lack of comparison available, 0.3 Wh is a small amount of energy. This may appear to be that using LLMs has an insignificant impact on the environment. This is certainly true on an individual level, as Hannah Ritchie rightfully argues [4]. However, when we consider the number of queries ChatGPT receives, the cumulative energy impact becomes much more worrying. It is estimated that ChatGPT has 164 million monthly users [6] and approximately 1 billion queries per day [4]. This puts the estimated annual energy usage of GPT-4o at 391,509 MWh, exceeding the electricity consumption of 35,000 U.S. residential households [4]. This is an absolutely staggering figure. Whilst I agree that on a personal level, queries to ChatGPT are not significant, on a global level, it is an excess that cannot be ignored.

In addition, the figures reported by these AI companies are for text-based prompts only. However, as AI develops, models are becoming increasingly more multi-modal, which, as explained in my previous post, use a substantially larger amount of energy. Moreover, these advanced AI methods utilising images and video are constantly being advertised as a way for models to stand out compared to the competition, increasing the visibility of multi-modal prompts and potentially increasing their usage. In new research, Delavande et al estimate the energy usage of video generation, a task becoming more accessible and popular. They estimate that this task uses approximately 30x more energy than image generation and 2000x more energy than text generation [7]. In addition, the energy usage scales quadratically with the number of frames and the resolution of the video [7]. Energy-efficient design is desperately needed if these models are to be used by the public at a frequent rate.

Conclusions

Whilst it is clear that more transparency has been achieved in regards to energy usage of current LLMs, we are still a long way away from the transparency the public deserves. AI companies must report the energy usage and carbon emissions of their models consistently and thoroughly. In doing so, consumers can easily compare models to make the most informed choice. Furthermore, methods must be explored and developed to make multi-modal models more energy efficient, especially when they are being rolled out and introduced to the general public. Finally, the onus should be on the public to be considerate with the use of AI. As we have shown, text prompts are relatively energy-efficient and can offer tangible benefits in improving workflows and saving time and money. However, despite what OpenAI and Google would like us to believe, AI should not be a catch-all solution for all our troubles. Many of us (myself included) do not need to be generating videos and images using AI. Just because we can does not mean we should. Becoming a steward of AI usage and supporting artists is becoming incredibly important. Besides, simple queries can easily be a Google or, even better, an Ecosia search. This is estimated to save around 10x the energy compared to a query.

Thinking before using will be harder as AI evolves and becomes increasingly integrated into our lives and products. However, it is possible if awareness of the energy impact of AI is increased. Figures such as Dr Sasha Luccioni and Nathaniel Burola are doing a great job of this, and I would also implore you to check out GreenPT as an alternative to common models [8]. If we, as the public, become more aware and demanding of the environmental impact of AI, change can and will happen.

1. https://arxiv.org/abs/2311.16863
2. https://huggingface.co/blog/sasha/environmental-impact-disclosures
3. https://www-technologyreview-com.cdn.ampproject.org/c/s/www.technologyreview.com/2025/08/21/1122288/google-gemini-ai-energy/amp/
4. https://arxiv.org/abs/2505.09598
5. https://www.sustainabilitybynumbers.com/p/carbon-footprint-chatgpt?utm_source=substack&utm_campaign=post_embed&utm_medium=web
6. https://sustainability-news.net/net-zero/chatgpts-monthly-carbon-footprint-equivalent-to-260-transatlantic-flights/
7. https://arxiv.org/abs/2509.19222
8. https://greenpt.ai/