O
Onder A.
Guest
Back in April of last year someone on X posted how much electricity OpenAI must spend when users add “please” and “thank you” to their ChatGPT prompts. Sam Altman responded with “Tens of millions of dollars well spent” then “you never know” [1].
Sounds like a dad joke. While doing some digging on the environmental impacts of AI, that quote started hitting different. My research led me down rabbit holes and the more I looked the joke stopped being funny because it wasn’t nearly close enough to the truth. “Please” equates to 1 or 2 tokens of the hundreds that could be part of any given prompt.
Where then is all that electricity going? To doing things we do mindlessly. Using the biggest model available just to polish up meeting minutes. Asking for a page long write-up when you really only needed two lines, or rephrasing a vague question over and over until you get the result you’re looking for.
Individually, these may not seem like problems. But at company scale doing this multiple times every working day for an entire year is the equivalent of running a small town. COP31 begins in Antalya on November 9th, so I thought I’d crunch some numbers.
Probably less then you think. Most numbers you see out there are from 2023 and used GPUs we dont really use anymore.
Google says their median prompt through gemini text is 0.24 Wh. This is good because they include the energy of machines sitting idle and cooling, not just the GPUs [2]. Open AI CEO Sam Altman thinks it's around .34 Wh per query [3]. I trust these Microsoft Researchers who measured Frontier models with live workloads and got a median of 0.31 Wh [4]. Because I like there methodology more I will use this number going forward.
We'll say .3 Wh. That's about as much energy as an led bulb (10 watts) would use if left on for just under 2 minutes. If you boiled a liter of water in your kettle it would take about 300 times more energy to make you a cup of tea.
A couple things to keep in mind before you get any ideas. One is how you define the boundaries of what you are measuring. If Google only looked at the AI accelerator they get .10 Wh. But when they include Host CPU's, Memory, Idle machines and building overhead they get .24 Wh [2]. Thats a factor of 2.4 before you even type a prompt. (Google says 58% of there energy is Accelerators 25% CPU and DRAM, 10% Machines sitting idle and 8% overhead).
The second thing is we are looking at the median, which won't tell us anything about the expensive tail end of prompts.
The table squashes a lot, so here’s the longer version.
If your use case requires an LLM to complete a task, then you're employing a lorry to drop off a postcard. As previously discussed there is going to be variance in energy usage per model and task. MIT Tech review estimated Meta Llama 3.1 405B at 1.86 Wh (Watt Hour) per completion and the 8B equivalent to only 0.03 Wh [12]. By using a smaller model built to complete specific tasks can save as much as 90% in energy usage according to UNESCO and UCL [14].
A reasoning mode will come in handy if your trying to solve a difficult planning problem but not if you want an email title. When Hugging Face calculated the energy score of AI models they discovered that using reasoning consumed 30x more energy on average and one model consumed almost 700x more [13].
An answer is generally going to cost you as many characters as your asking for. If your asking for an explanation when you could have just asked yes/no your going to pay for the long response. According to UNESCO and UCL by keeping your prompt and answer short you can save over half [14].
When you retry something your consuming more. You've created 5 answers when you only needed one and you've read the whole conversation 5 times. This probably makes up a large amount of the thank you money. Generating those words is virtually free but a simple "thanks!" will consume more energy because it has to read through your conversation to respond to something you didn't need to ask.
Copy and pasting a 100-page report into your chat to ask about one paragraph means your model will have to read that whole report. Copy the paragraph and paste it into your chat.
Using Agents will increase your numbers. Anthropic state that their agents use around 4x the amount of tokens than normal chatting and multi agent use around 15x [15].
Here's an example we can use. We know there are around 5,000 office workers in a typical UK company that may be using AI. Each worker may use an AI assistant 20 times a day, and we'll assume they work 250 days a year. This will equal 25 million tasks done by AI in a year. Using the figures from Microsoft on production [4] we can calculate how much energy is consumed per query. To put this in context we can look at the typical household usage which is 2,500 Kwh a year from Ofgem. This figure was updated in July 2026 [16]. Using the average UK grid emission for 2025 we get 126 gCO2/Kwh [10].
This is just a matter of habits. You can see above that with the same work being done scenario C uses 38 times as much energy as scenario A. This would equate to around 3 million kettles boiled. As providers will charge you based on how many tokens you use your bill will increase by that factor. Being green is also cheaper in this case.
Please note this is just an example. There will be variances depending on what model and hardware is used to answer the question. I have made some assumptions about the amount of retries and length of answers. All figures can be found online and small changes to these numbers won’t affect the big picture.
The primary hurdle is that most organizations lack visibility into these figures. A study by Sasha Luccioni and colleagues, examining token traffic on OpenRouter in May 2025, found that 84% of it went to models that provided no information about their energy consumption. Only 2% went to models with directly measured figures [21].
However, this situation is slowly changing:
As stated before a single prompt is no different then having a light bulb on for 2 mins. But we don’t only prompt once. By using too large of models and over-reasoning. Answering with too much information. Or retrying our tasks can cause a single task to use tens of times more energy. If we took one company that could be the energy usage of over 100 houses. Across the UK we could be using as much energy as most of London’s transport for 2 weeks.
We don't need more tech to help with this. We just need to be more mindful of how we use these tools and track our consumption. With COP31 coming up we should all be making small changes to our lives. That's what engineering teams should be doing.
Ask your chatbot for "please" but maybe leave out the thank you.
1. Futurism. “Sam Altman Admits That Saying ‘Please’ and ‘Thank You’ to ChatGPT Is Wasting Millions of Dollars in Computing Power.” April 2025. https://futurism.com/altman-please-thanks-chatgpt
2. Google Cloud. “Measuring the environmental impact of AI inference.” August 2025. https://cloud.google.com/blog/produ...ring-the-environmental-impact-of-ai-inference
3. Altman, S. “The Gentle Singularity.” June 2025. https://blog.samaltman.com/the-gentle-singularity
4. Oviedo, F. et al. “Energy use of AI inference: Efficiency pathways and test-time scaling.” Joule, 2026. https://www.sciencedirect.com/science/article/pii/S2542435126001145
5. Adamska, M. et al. “Green Prompting: Characterizing Prompt-driven Energy Costs of LLM Inference.” arXiv:2503.10666, 2025. https://arxiv.org/html/2503.10666v4
6. Epoch AI. “How much energy does ChatGPT use?” February 2025. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
7. Anthropic. “Prompt caching.” Claude Platform documentation. https://platform.claude.com/docs/en/build-with-claude/prompt-caching
8. International Energy Agency. “Key Questions on Energy and AI.” April 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
9. Shacknews. “Google CEO Sundar Pichai says the company is processing over 3.2 quadrillion tokens/month.” May 2026. https://www.shacknews.com/article/149205/google-3-2-quadrillion-monthly-ai-tokens
10. Carbon Brief. “Analysis: UK renewables enjoy record year in 2025, but gas power still rises.” January 2026. https://www.carbonbrief.org/analysi...record-year-in-2025-but-gas-power-still-rises
11. National Grid Electricity Distribution. “Data Centre Impact Study.” November 2025. https://dso.nationalgrid.co.uk/downloads/15083/data-centre-impact-study2.pdf
12. O’Donnell, J. and Crownhart, C. “We did the math on AI’s energy footprint.” MIT Technology Review, May 2025. https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/
13. Hugging Face. “AI Energy Score v2: Refreshed Leaderboard, now with Reasoning.” December 2025. https://huggingface.co/blog/sasha/ai-energy-score-v2
14. UNESCO and University College London. “Smarter, Smaller, Stronger: Resource-Efficient Generative AI and the Future of Digital Transformation.” July 2025. https://www.unesco.org/en/articles/...-shows-small-changes-can-reduce-energy-use-90
15. Anthropic. “How we built our multi-agent research system.” June 2025. https://www.anthropic.com/engineering/multi-agent-research-system
16. Ofgem. “Summary of changes to the energy price cap, 1 July to 30 September 2026: revised typical domestic consumption values.” June 2026. https://www.ofgem.gov.uk/sites/defa...-1-July-to-30-September-2026-revised-TDCV.pdf
17. Office for National Statistics. “Families and households in the UK: 2024.” July 2025. https://www.ons.gov.uk/peoplepopula...families/bulletins/familiesandhouseholds/2024
18. Transport for London. Finance Committee paper on power purchase agreements. November 2024. https://board.tfl.gov.uk/documents/s22990/fc-20241119-item04a-PPA-part1.pdf
19. Ember. “Global Electricity Review 2026: Major countries and regions.” 2026. https://ember-energy.org/latest-insights/global-electricity-review-2026/major-countries-and-regions/
20. Jegham, N. et al. “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference.” arXiv:2505.09598, 2025. https://arxiv.org/abs/2505.09598
21. Luccioni, S. et al. “Misinformation by Omission: The Need for More Environmental Transparency in AI.” arXiv:2506.15572, 2025. https://arxiv.org/abs/2506.15572
22. ITU-T. “Recommendation L.1801: Guidelines for assessing the environmental impact of artificial intelligence systems.” February 2026. https://www.itu.int/rec/T-REC-L.1801-202602-P
23. ISO/IEC. “TR 20226:2025, Information technology, Artificial intelligence, Environmental sustainability aspects of AI systems.” 2025. https://www.iso.org/standard/86177.html
24. Green Software Foundation. “SCI for AI specification ratified.” December 2025. https://greensoftware.foundation/ar...andard-for-measuring-ai-emissions-across-the/
25. IEEE Standards Association. “P7100: Standard for Measurement of Environmental Impacts of Artificial Intelligence Systems.” https://standards.ieee.org/ieee/7100/11671/
26. European Commission. “Targeted consultation on measuring energy consumption and emissions of AI models and systems.” 2026. https://digital-strategy.ec.europa....nsumption-and-emissions-ai-models-and-systems
27. European Commission. “Making data centres energy efficient thanks to new EU rating system.” September 2026. https://commission.europa.eu/news-a...ent-thanks-new-eu-rating-system-2026-09-21_en
28. National Energy Administration of China. Action plan on AI and energy. April 2026. https://www.nea.gov.cn/20260508/4dae97ca01d348e4871bb8654be34b3a/c.html
29. CMS. “UK Government publishes final sustainability reporting standards.” 2026. https://cms.law/en/gbr/legal-update...shes-final-sustainability-reporting-standards
30.Greenhouse Gas Protocol. “Corporate Value Chain (Scope 3) Accounting and Reporting Standard.” 2011. https://ghgprotocol.org/sites/defau...ain-Accounting-Reporing-Standard_041613_2.pdf
Sounds like a dad joke. While doing some digging on the environmental impacts of AI, that quote started hitting different. My research led me down rabbit holes and the more I looked the joke stopped being funny because it wasn’t nearly close enough to the truth. “Please” equates to 1 or 2 tokens of the hundreds that could be part of any given prompt.
Where then is all that electricity going? To doing things we do mindlessly. Using the biggest model available just to polish up meeting minutes. Asking for a page long write-up when you really only needed two lines, or rephrasing a vague question over and over until you get the result you’re looking for.
Individually, these may not seem like problems. But at company scale doing this multiple times every working day for an entire year is the equivalent of running a small town. COP31 begins in Antalya on November 9th, so I thought I’d crunch some numbers.
How much energy does one prompt take?
Probably less then you think. Most numbers you see out there are from 2023 and used GPUs we dont really use anymore.
Google says their median prompt through gemini text is 0.24 Wh. This is good because they include the energy of machines sitting idle and cooling, not just the GPUs [2]. Open AI CEO Sam Altman thinks it's around .34 Wh per query [3]. I trust these Microsoft Researchers who measured Frontier models with live workloads and got a median of 0.31 Wh [4]. Because I like there methodology more I will use this number going forward.
We'll say .3 Wh. That's about as much energy as an led bulb (10 watts) would use if left on for just under 2 minutes. If you boiled a liter of water in your kettle it would take about 300 times more energy to make you a cup of tea.
A couple things to keep in mind before you get any ideas. One is how you define the boundaries of what you are measuring. If Google only looked at the AI accelerator they get .10 Wh. But when they include Host CPU's, Memory, Idle machines and building overhead they get .24 Wh [2]. Thats a factor of 2.4 before you even type a prompt. (Google says 58% of there energy is Accelerators 25% CPU and DRAM, 10% Machines sitting idle and 8% overhead).
The second thing is we are looking at the median, which won't tell us anything about the expensive tail end of prompts.
If your use case requires an LLM to complete a task, then you're employing a lorry to drop off a postcard. As previously discussed there is going to be variance in energy usage per model and task. MIT Tech review estimated Meta Llama 3.1 405B at 1.86 Wh (Watt Hour) per completion and the 8B equivalent to only 0.03 Wh [12]. By using a smaller model built to complete specific tasks can save as much as 90% in energy usage according to UNESCO and UCL [14].
A reasoning mode will come in handy if your trying to solve a difficult planning problem but not if you want an email title. When Hugging Face calculated the energy score of AI models they discovered that using reasoning consumed 30x more energy on average and one model consumed almost 700x more [13].
An answer is generally going to cost you as many characters as your asking for. If your asking for an explanation when you could have just asked yes/no your going to pay for the long response. According to UNESCO and UCL by keeping your prompt and answer short you can save over half [14].
When you retry something your consuming more. You've created 5 answers when you only needed one and you've read the whole conversation 5 times. This probably makes up a large amount of the thank you money. Generating those words is virtually free but a simple "thanks!" will consume more energy because it has to read through your conversation to respond to something you didn't need to ask.
Copy and pasting a 100-page report into your chat to ask about one paragraph means your model will have to read that whole report. Copy the paragraph and paste it into your chat.
Using Agents will increase your numbers. Anthropic state that their agents use around 4x the amount of tokens than normal chatting and multi agent use around 15x [15].
Doing the maths on one company
Here's an example we can use. We know there are around 5,000 office workers in a typical UK company that may be using AI. Each worker may use an AI assistant 20 times a day, and we'll assume they work 250 days a year. This will equal 25 million tasks done by AI in a year. Using the figures from Microsoft on production [4] we can calculate how much energy is consumed per query. To put this in context we can look at the typical household usage which is 2,500 Kwh a year from Ofgem. This figure was updated in July 2026 [16]. Using the average UK grid emission for 2025 we get 126 gCO2/Kwh [10].
This is just a matter of habits. You can see above that with the same work being done scenario C uses 38 times as much energy as scenario A. This would equate to around 3 million kettles boiled. As providers will charge you based on how many tokens you use your bill will increase by that factor. Being green is also cheaper in this case.
Please note this is just an example. There will be variances depending on what model and hardware is used to answer the question. I have made some assumptions about the amount of retries and length of answers. All figures can be found online and small changes to these numbers won’t affect the big picture.
The Challenge of Measurement
The primary hurdle is that most organizations lack visibility into these figures. A study by Sasha Luccioni and colleagues, examining token traffic on OpenRouter in May 2025, found that 84% of it went to models that provided no information about their energy consumption. Only 2% went to models with directly measured figures [21].
However, this situation is slowly changing:
- The ITU approved L.1801 in February 2026, providing guidelines for assessing the environmental impact of AI systems over their life cycle, with separate considerations for training and inference [22].
- ISO/IEC TR 20226:2025 outlines the environmental sustainability aspects of AI systems [23].
- The Green Software Foundation's SCI for AI, building on ISO/IEC 21031, enables reporting of emissions per token or per inference [24].
- IEEE has a working group developing P7100, a standard for measuring the environmental impact of AI systems [25].
- The EU AI Act already requires providers of general-purpose models to document their known or estimated energy consumption. The European Commission consulted on measurement methods in 2026 [26] and is introducing an A to G energy rating for data centers [27].
- Meanwhile, China has Power Usage Effectiveness (PUE) targets for data centers. In April 2026, China further mandated that computing power be measured alongside electricity consumption and carbon emissions [28].
- In the UK, the government published the final UK Sustainability Reporting Standards in February 2026 [29]. While currently voluntary, the FCA has proposed making them mandatory for listed companies from January 1, 2027. These standards include Scope 3 emissions, and the GHG Protocol generally categorizes purchased services, such as cloud and AI, as Scope 3 [30]. Companies will receive transitional relief initially, and I anticipate some adjustments to the details. Nevertheless, if your team begins logging tokens now, you will be significantly better prepared.
Things I would do:
- Educate people on how to prompt. You will always have people that don’t know how to properly use tokens. People need to learn how to write a task and what output they want. In how much detail they want. At the beginning of their conversation. Not something we offer to teach them but taught at onboarding. Even better if we taught them in schools and universities. Which UNESCO suggests we do[14].
- Use lower models. Use the lowest model that passes your testing and scale up when they don’t. Make people think about when to use advanced reasoning.
- Limit how many tokens are used. Set max_tokens and ask for output in JSON or X amount of bullet points. Don’t send a chat api "thanks".
- Reduce your context. Only get what you need from your pdf's. Shorten conversations when they get long. Start a new chat when you start talking about something different. Keep your prompt prefix the same to allow caching.
- Count how many tokens you used per task not request. Make sure you include reasoning tokens in this number. Ask your provider how many energy per token there is. Use L.1801 or SCI standard for AI. And put this in your scope 3.
Conclusion:
As stated before a single prompt is no different then having a light bulb on for 2 mins. But we don’t only prompt once. By using too large of models and over-reasoning. Answering with too much information. Or retrying our tasks can cause a single task to use tens of times more energy. If we took one company that could be the energy usage of over 100 houses. Across the UK we could be using as much energy as most of London’s transport for 2 weeks.
We don't need more tech to help with this. We just need to be more mindful of how we use these tools and track our consumption. With COP31 coming up we should all be making small changes to our lives. That's what engineering teams should be doing.
Ask your chatbot for "please" but maybe leave out the thank you.
References
1. Futurism. “Sam Altman Admits That Saying ‘Please’ and ‘Thank You’ to ChatGPT Is Wasting Millions of Dollars in Computing Power.” April 2025. https://futurism.com/altman-please-thanks-chatgpt
2. Google Cloud. “Measuring the environmental impact of AI inference.” August 2025. https://cloud.google.com/blog/produ...ring-the-environmental-impact-of-ai-inference
3. Altman, S. “The Gentle Singularity.” June 2025. https://blog.samaltman.com/the-gentle-singularity
4. Oviedo, F. et al. “Energy use of AI inference: Efficiency pathways and test-time scaling.” Joule, 2026. https://www.sciencedirect.com/science/article/pii/S2542435126001145
5. Adamska, M. et al. “Green Prompting: Characterizing Prompt-driven Energy Costs of LLM Inference.” arXiv:2503.10666, 2025. https://arxiv.org/html/2503.10666v4
6. Epoch AI. “How much energy does ChatGPT use?” February 2025. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
7. Anthropic. “Prompt caching.” Claude Platform documentation. https://platform.claude.com/docs/en/build-with-claude/prompt-caching
8. International Energy Agency. “Key Questions on Energy and AI.” April 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
9. Shacknews. “Google CEO Sundar Pichai says the company is processing over 3.2 quadrillion tokens/month.” May 2026. https://www.shacknews.com/article/149205/google-3-2-quadrillion-monthly-ai-tokens
10. Carbon Brief. “Analysis: UK renewables enjoy record year in 2025, but gas power still rises.” January 2026. https://www.carbonbrief.org/analysi...record-year-in-2025-but-gas-power-still-rises
11. National Grid Electricity Distribution. “Data Centre Impact Study.” November 2025. https://dso.nationalgrid.co.uk/downloads/15083/data-centre-impact-study2.pdf
12. O’Donnell, J. and Crownhart, C. “We did the math on AI’s energy footprint.” MIT Technology Review, May 2025. https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/
13. Hugging Face. “AI Energy Score v2: Refreshed Leaderboard, now with Reasoning.” December 2025. https://huggingface.co/blog/sasha/ai-energy-score-v2
14. UNESCO and University College London. “Smarter, Smaller, Stronger: Resource-Efficient Generative AI and the Future of Digital Transformation.” July 2025. https://www.unesco.org/en/articles/...-shows-small-changes-can-reduce-energy-use-90
15. Anthropic. “How we built our multi-agent research system.” June 2025. https://www.anthropic.com/engineering/multi-agent-research-system
16. Ofgem. “Summary of changes to the energy price cap, 1 July to 30 September 2026: revised typical domestic consumption values.” June 2026. https://www.ofgem.gov.uk/sites/defa...-1-July-to-30-September-2026-revised-TDCV.pdf
17. Office for National Statistics. “Families and households in the UK: 2024.” July 2025. https://www.ons.gov.uk/peoplepopula...families/bulletins/familiesandhouseholds/2024
18. Transport for London. Finance Committee paper on power purchase agreements. November 2024. https://board.tfl.gov.uk/documents/s22990/fc-20241119-item04a-PPA-part1.pdf
19. Ember. “Global Electricity Review 2026: Major countries and regions.” 2026. https://ember-energy.org/latest-insights/global-electricity-review-2026/major-countries-and-regions/
20. Jegham, N. et al. “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference.” arXiv:2505.09598, 2025. https://arxiv.org/abs/2505.09598
21. Luccioni, S. et al. “Misinformation by Omission: The Need for More Environmental Transparency in AI.” arXiv:2506.15572, 2025. https://arxiv.org/abs/2506.15572
22. ITU-T. “Recommendation L.1801: Guidelines for assessing the environmental impact of artificial intelligence systems.” February 2026. https://www.itu.int/rec/T-REC-L.1801-202602-P
23. ISO/IEC. “TR 20226:2025, Information technology, Artificial intelligence, Environmental sustainability aspects of AI systems.” 2025. https://www.iso.org/standard/86177.html
24. Green Software Foundation. “SCI for AI specification ratified.” December 2025. https://greensoftware.foundation/ar...andard-for-measuring-ai-emissions-across-the/
25. IEEE Standards Association. “P7100: Standard for Measurement of Environmental Impacts of Artificial Intelligence Systems.” https://standards.ieee.org/ieee/7100/11671/
26. European Commission. “Targeted consultation on measuring energy consumption and emissions of AI models and systems.” 2026. https://digital-strategy.ec.europa....nsumption-and-emissions-ai-models-and-systems
27. European Commission. “Making data centres energy efficient thanks to new EU rating system.” September 2026. https://commission.europa.eu/news-a...ent-thanks-new-eu-rating-system-2026-09-21_en
28. National Energy Administration of China. Action plan on AI and energy. April 2026. https://www.nea.gov.cn/20260508/4dae97ca01d348e4871bb8654be34b3a/c.html
29. CMS. “UK Government publishes final sustainability reporting standards.” 2026. https://cms.law/en/gbr/legal-update...shes-final-sustainability-reporting-standards
30.Greenhouse Gas Protocol. “Corporate Value Chain (Scope 3) Accounting and Reporting Standard.” 2011. https://ghgprotocol.org/sites/defau...ain-Accounting-Reporing-Standard_041613_2.pdf