The environmental cost of AI

A conversation with ChatGPT about consumption

Below, I have provided a thread with ChatGPT. The purpose of this was to see how much I could glean from the system itself on its own environmental impact. For the purpose of readability, my queries will be in bold.

---

what is an efficient way to measure the electricity consumption required to run a software application?

Measuring the electricity consumption of a software application efficiently depends on the level of precision required and the resources available. Here are some practical approaches:

1. Hardware-Based Power Monitoring

For the most accurate measurement, use external power meters or built-in hardware monitoring tools.

a) Smart Plug Power Meters (for entire device measurement)

• Devices like the Kill A Watt meter or Shelly Plug S can measure the power draw of a desktop, laptop, or server running the software.

• Measure idle power consumption vs. consumption while the software is running.

• Compute the difference to estimate the software’s energy usage.

b) Internal Power Sensors (for component-level measurement)

• Some high-end motherboards and power supplies provide real-time power usage metrics.

• Software like Intel Power Gadget (for Intel CPUs) or Ryzen Master (for AMD) can measure CPU-specific power usage.

c) Server-Level Power Monitoring

• Cloud platforms like AWS, Azure, and Google Cloud provide power usage insights.

• Enterprise tools like OpenStack Ceilometer or Dell iDRAC monitor data center power consumption.

2. Software-Based Power Estimation

If you don’t have access to hardware monitoring, software tools can estimate power usage.

a) OS-Level Power Monitoring

• Windows: Use the Windows Energy Estimation Engine (E3) via powercfg /energy or Windows Performance Toolkit.

• Mac: Use Intel Power Gadget or pmset -g.

• Linux: Use PowerTOP or read power data from /sys/class/power_supply/.

b) Application-Specific Power Profiling

• Joulemeter (Microsoft Research): Estimates power usage based on CPU, memory, and disk activity.

• PowerAPI: A power monitoring tool for Linux-based applications.

c) Programming Frameworks for Energy Profiling

• If you are a developer, you can integrate energy profiling within the application:

• Intel RAPL (Running Average Power Limit) – Estimates energy usage per CPU package.

• NVIDIA-SMI (for GPUs) – Reports real-time GPU power draw.

• Perf (Linux Performance Counter) – Measures energy usage per process.

3. Cloud-Based Estimations

If the software runs in a cloud environment, use:

• AWS Carbon Footprint Tool

• Google Cloud Sustainability Dashboard

• Azure Sustainability Calculator

These tools estimate energy consumption based on usage metrics.

4. Methodology for Measuring Software Power Consumption

1. Baseline Measurement: Record system power consumption while idle.

2. Active Measurement: Run the software under typical load conditions.

3. Calculate the Difference: Subtract baseline consumption from active consumption.

4. Repeat for Different Workloads: If possible, test under various loads to get an accurate range.

5. Estimate Energy Cost: Convert power (Watts) into energy (kWh) to estimate cost and environmental impact.

For quick estimates, use PowerTOP or Intel Power Gadget. For detailed analysis, use RAPL and external power meters.

Would you like help setting up a measurement plan for a specific software application?

can you provide statistics for your own electricity consumption?

I don’t have direct access to my own electricity consumption since I run on cloud infrastructure managed by OpenAI. However, OpenAI has shared insights on the energy consumption of large language models like me. Here’s what’s generally known:

Estimated Energy Usage of Large Language Models

1. Training Phase:

• Training large AI models like GPT-4 requires thousands of GPUs running for weeks or months.

• Estimates suggest that training GPT-3 (a predecessor) consumed around 1,287 MWh (megawatt-hours)—enough to power about 120 US homes for a year.

• GPT-4 is likely even more energy-intensive due to its increased size and complexity.

2. Inference (Everyday Usage) Phase:

• Running (inference) is more energy-efficient than training but still significant.

• Estimates suggest that a single query to a large AI model consumes 2-10 Wh (watt-hours).

• This means:

  • A Google search typically consumes 0.3 Wh.

  • A single AI-generated response might use 10x more energy than a search.

  • A person using ChatGPT for an hour could consume 50-100 Wh, equivalent to running a light bulb for 1-2 hours.

3. Cloud Computing Impact:

• OpenAI operates in data centers, which use energy-efficient GPUs like NVIDIA A100 and H100.

• Many AI data centers use renewable energy and optimise workloads to reduce carbon footprints.

Environmental Impact

• AI models like me contribute to carbon emissions, but cloud providers are moving toward carbon-neutral and renewable-powered data centers.

• Microsoft, OpenAI’s partner, aims to be carbon-negative by 2030, meaning their infrastructure will actively remove more carbon than it emits.

Would you like me to estimate the energy impact of a specific AI usage scenario?

what was the energy impact of you responding the previous question from me and this one?

To estimate the energy impact of my response to your last two questions, let’s break it down using approximate figures.

1. Energy per AI Query

• A single AI-generated response is estimated to consume 2-10 Wh (Watt-hours), depending on the complexity and length of the answer.

• Simpler queries (a few sentences) are on the lower end (~2 Wh).

• Long, detailed answers (like my previous one) lean toward the higher end (~8-10 Wh).

3. What Does This Energy Usage Compare To?

• 18 Wh is equivalent to:

• Running an LED light bulb (10W) for 1.8 hours.

• Charging a smartphone (5 Wh per charge) nearly 4 times.

• Running a laptop (50 Wh per hour) for about 22 minutes.

• Watching Netflix in HD for 3-4 minutes.

4. CO₂ Emissions Estimate

• Assuming a global electricity carbon intensity of 400g CO₂ per kWh:

  • 18 Wh = 0.018 kWh

  • CO₂ Emissions = 0.018 × 400 = 7.2 grams of CO₂

  • That’s about the same CO₂ impact as sending 5-10 emails.

Final Thought

Your two questions had a tiny energy impact compared to everyday digital activities, but at a large scale (millions of users), AI inference does contribute to overall electricity demand.

Would you like a similar estimate for other AI-related activities, like using ChatGPT for an hour or processing large datasets?

based on your calculations for this thread, and your understanding of the current average daily users globally for ChatGPT, provide an estimate of the energy used. In addition, estimate the total water consumption required to cool the data centres hosting OpenAI servers

Estimating the energy and water consumption associated with ChatGPT’s operations involves several considerations. Here’s a detailed breakdown:

1. ChatGPT’s User Activity

• Weekly Active Users: As of February 2025, ChatGPT has approximately 400 million weekly active users.

• Daily Active Users: This translates to an estimated 57 million daily active users (assuming consistent usage throughout the week).

• Daily Queries: ChatGPT processes over 1 billion queries per day.

2. Energy Consumption per Query

• Per Query Energy Use: Estimates suggest that each ChatGPT query consumes between 2 to 10 watt-hours (Wh), depending on complexity.

• Daily Energy Consumption: With 1 billion queries daily, the total energy consumption ranges from 2,000 megawatt-hours (MWh) to 10,000 MWh.

3. Annual Energy Consumption

• Annual Usage: Extrapolating daily usage, ChatGPT’s annual energy consumption is between 730,000 MWh and 3,650,000 MWh.

4. Data Center Water Consumption

Data centers use water primarily for cooling purposes. The efficiency of this usage is measured by Water Usage Effectiveness (WUE), defined as liters of water consumed per kilowatt-hour (kWh) of energy used.

• Average WUE: The industry average WUE is approximately 1.8 liters per kWh.

• Efficient Data Centers: Leading operators like AWS report a lower WUE of 0.19 liters per kWh.

5. Estimating ChatGPT’s Water Consumption

Using the energy consumption figures and WUE metrics:

• With Average WUE (1.8 L/kWh):

• Daily Water Use: 2,000 MWh to 10,000 MWh equates to 3.6 million to 18 million liters.

• Annual Water Use: This totals 1.3 billion to 6.6 billion liters annually.

• With Efficient WUE (0.19 L/kWh):

• Daily Water Use: 2,000 MWh to 10,000 MWh equates to 380,000 to 1.9 million liters.

• Annual Water Use: This totals 138 million to 693 million liters annually.

6. Contextualizing the Water Usage

• Per Query Water Use: With an efficient WUE of 0.19 L/kWh and per-query energy consumption of 2 to 10 Wh, each query uses approximately 0.00038 to 0.0019 liters (or 0.38 to 1.9 milliliters).

• Comparison: For perspective, the average data center uses about 300,000 gallons (1.14 million liters) of water daily, comparable to the water usage of 100,000 homes.

7. Factors Influencing These Estimates

• Data Center Efficiency: Variations in WUE across different data centers can significantly impact water usage.

• Geographical Location: Data centers in cooler climates may require less water for cooling.

• Technological Advances: Innovations in cooling technologies, such as liquid cooling systems, can reduce water consumption.

• 8. Environmental Considerations

• Resource Management: The substantial water usage by AI operations like ChatGPT underscores the importance of sustainable practices in data center management.

• Regional Impact: In areas facing water scarcity, the water consumption of large data centers can strain local resources, necessitating careful site selection and resource planning.

In summary, while ChatGPT’s operations do consume significant energy and water resources, efforts toward improving data center efficiency and adopting sustainable practices can help mitigate the environmental impact.