Upgrade to Pro — share decks privately, control downloads, hide ads and more …

apidays New York 2025 - Why I Built Another Car...

apidays New York 2025 - Why I Built Another Carbon Measurement Tool for LLMs by Pascal Joly (IT Climate Ed)

Why I Built Another Carbon Measurement Tool for LLMs (And What I Learned Along the Way)
Pascal Joly, Sustainability Consultant and Instructor at IT Climate Ed

apidays New York 2025
API Management for Surfing the Next Innovation Waves: GenAI and Open Banking
May 14 & 15, 2025

------

Check out our conferences at https://www.apidays.global/

Do you want to sponsor or talk at one of our conferences?
https://apidays.typeform.com/to/ILJeAaV8

Learn more on APIscene, the global media made by the community for the community:
https://www.apiscene.io

Explore the API ecosystem with the API Landscape:
https://apilandscape.apiscene.io/

Avatar for apidays

apidays

May 24, 2025
Tweet

More Decks by apidays

Other Decks in Programming

Transcript

  1. Why I Built a Carbon Measurement Tool for LLMs (And

    What I Learned Along the Way) Pascal Joly, IT Climate Ed, LLC
  2. “it's super fun seeing people love images in ChatGPT. but

    our GPUs are melting.” Sam Altman, March 2025
  3. The Ghibli image craze March 2025: ChatGPT image generator creates

    viral trend 5M images generated/hour Purely recreational = 1 AI image 1 phone charge Courtesy of X, @Zeneca- Ghibli-Style AI Art Misses What Made Miyazaki Matter
  4. Energy demand from data centers is set to more than

    double by 2030 to 945TW IEA report, April 2025
  5. Searching for answers Will AI benefits outweigh its environmental impact?

    How much of the data center expected growth can be attributed to AI? How models compare in terms of energy efficiency for training and inference? How much carbon emissions do I emit from my chatbot conversations? How much does LLM energy intensity vary by task type and complexity? Will gains in model energy efficiency solve the impending energy crisis?
  6. Energy impact of LLM conversations 400M Chatgpt Users 2.9 Wh

    Per Query (*) 33,000 US Homes (Daily Power Equivalent) (*) Source: IEA Energy report, 2025
  7. What if we could change how people use LLMs? Wasteful

    Prompt Hi There! I was wondering if you could please tell me about renewable energy sources. Thank you so much! Efficient Prompt List renewable energy sources Extra tokens: ~40% wasted energy Minimal tokens: optimal energy use
  8. Estimating AI emissions: tool ecosystem 👤 Role 🎯 Goal 🛠

    Recommended Tool(s) Product Manager Model selection (training phase) ML CO2 Calculator, Green Algorithm Product Manager Model selection (inference phase) EcologITs, AI Energy Score Sustainability Analyst (Enterprise) Carbon footprint of current projects Cloud Dashboards, Carbon Accounting Platforms Developer Evaluate/improve model performance CarbonTracker, CodeCarbon End User Estimate my inference emissions (chatGPT, claude…) ??
  9. Introducing AIWattch End User-focused ⏱ Real-time feedback on LLM interactions

    Open source and collaborative Private and Secure Educational Support most popular LLMs
  10. AI Emissions Estimator - System Flow ChatGPT Interface DOM Monitor

    Token Estimator Char/token: 4 Emissions Calculator Real-time Display Admin Dashboard PUE: 1.125 Grid: 383 gCO2e/kWh Input: 0.002 Wh/token Output: 0.01 Wh/token
  11. Calculation 1.Number of input and output tokens: #Char input x

    Token/Char = Ti #Char reply x Token/Char = Tr 2.Total Energy during conv exchange: PUE x ((Ti x (EnergyF Ti) + Tr x (EnergyF Tr)) = Conv Energy (kWh) 3.CO2 emissions: Conv Energy x CO2i = CO2 emitted (g) 2300.56
  12. My LLM-assisted Product Development Journey Frame the problem and brainstorm

    possible solutions. 󰣗 Prototype with Vibe coding…. Debugging (!) PRD and project proposal Hired a developer On boarded the project on Github for open source contributions
  13. Challenges and Lessons Learned 🔒 Transparency Gap Limited visibility into

    non-open source models' energy consumption 📍Data center Uncertainty Unknown carbon intensity and PUE ⏱ Real Time complexity Estimating query on the fly: response time Browser Integration DOM updates, dynamic web interfaces 🔄 Task Variability Different AI workloads have vastly different energy profiles 💡Key Learning Set the right level of expectations for accuracy. The goal is estimations.
  14. Impact: the Scale of Opportunity 🧠 Efficient Prompting If just

    1% (4M) adopt energy-efficient prompting (40%): → Saves ~1.6 GWh / year →Avoids 730 tons CO₂/year ➡ It’s Bigger Than ChatGPT Also used by millions: Claude · Gemini · LLaMA · Copilot · Mistral · DeepSeek… It’s Bigger Than Energy An average 100MW datacenter: 2M ltrs water/day (*) source Source: IEA AI and Energy report, April 2025
  15. Impact: transparency and standardization 📣 Awareness → Transparency • Creates

    user demand for AI energy labels • Pressure vendors for disclosure & accountability 🏢 From Habits to Standards • Teams: Best practices • Organization: Internal policies • Enterprise: Industry-wide standards
  16. Roadmap • Time based calculation • Actionable tips • Multimodal

    Feature backlog: https://github.com/AIWattch/browser-extension/issues Near Term Medium Term Long Term • API integration for token counting • Reasoning Model • Additional LLM and browser support • Teams • Summary dashboard • Gamified approach
  17. AIWach in the tool ecosystem 👤 Role 🎯 Goal 🛠

    Recommended Tool(s) Product Manager Model selection (training phase) ML CO2 Calculator, Green Algorithm Product Manager Model selection (inference phase) EcologITs, AI Energy Score Sustainability Analyst (Enterprise) Carbon footprint of current projects Cloud Dashboards, Carbon Accounting Platforms Developer Evaluate/improve model performance CarbonTracker, CodeCarbon End User Estimate inference emissions (ChatGPT, Claude…) AIWattch
  18. What can YOU do? Join the Movement Contribute: https://github.com/AIWattch/browser-extension Try

    AIWattch and provide feedback Take a course Bring changes from within
  19. Q&A