Energy, emissions and water for one text prompt or response from a large language model. Each entry records the claim as published: what it counts, who made it and who paid,
and any dispute. Figures with different boundaries are not comparable without reading what each counts.
ChatGPT request in early 2023, from SemiAnalysis via de Vries
at most 2.9 Wh per ChatGPT request
This figure aligns with SemiAnalysis’ assessment of ChatGPT’s operating costs in early 2023, which estimated that ChatGPT responds to 195 million requests per day, requiring an estimated average electricity consumption of 564 MWh per day, or, at most, 2.9 Wh per request.
An average ChatGPT request in early 2023 (564 MWh per day over 195 million requests per day)
Scope
not stated
Counts
not stated
Leaves out
not stated
Model or version
ChatGPT (early 2023)
Place
not stated
Period measured
early 2023
Method
Taken from SemiAnalysis' early-2023 estimate of ChatGPT's daily electricity use (3,617 Nvidia HGX A100 servers, 564 MWh per day) divided by 195 million requests per day; the paper states it as an upper bound ("at most").
Claimed by
Independent: Alex de Vries (VU Amsterdam, Digiconomist)
Paid for by
not stated
Dispute
Epoch AI (February 2025) says this estimate, the origin of the widely cited 3 Wh figure, rests on outdated hardware and an unrepresentative assumption of about 4,000 input and 2,000 output tokens per query, and estimates roughly 0.3 Wh for a typical GPT-4o query. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
History
Added.
ChatGPT request, IEA Electricity 2024
2.9 Wh per ChatGPT request
When comparing the average electricity demand of a typical Google search (0.3 Wh of electricity) to OpenAI’s ChatGPT (2.9 Wh per request), and considering 9 billion searches daily, this would require almost 10 TWh of additional electricity in a year.
Epoch AI (February 2025) argues that the widely cited figure of about 3 Wh per ChatGPT query, which traces to de Vries (2023), is likely an overestimate, and puts a typical GPT-4o query at roughly 0.3 Wh. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
History
Added.
ChatGPT request, EPRI Powering Intelligence
2.9 Wh per ChatGPT request
At 2.9 watt-hours per ChatGPT request, AI queries are estimated to require 10x the electricity of traditional Google queries, which use about 0.3 watt-hours each
Industry-funded: Electric Power Research Institute (EPRI)
Paid for by
EPRI's members, chiefly electric utilities (no funder named for this report)
Dispute
Epoch AI (February 2025) argues that the widely cited figure of about 3 Wh per ChatGPT query, which traces to de Vries (2023), is likely an overestimate, and puts a typical GPT-4o query at roughly 0.3 Wh. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
History
Added.
Text generation inference, Power Hungry Processing
a mean of 0.047 Wh per text generation inference (published as: 0.047 kWh per 1,000 inferences (0.047 kWh / 1,000 = 0.047 Wh per inference))
image generation uses, on average, over 60 times more energy than text generation (0.047 vs. 2.9 kWh).
One text-generation inference generating 10 new tokens, unbatched, on benchmark datasets
Scope
servers
Counts
GPU, CPU and RAM energy measured with the Code Carbon package
Leaves out
not stated
Model or version
The 8 most popular text-generation models on the Hugging Face Hub at the time of the study
Place
Amazon Web Services us-west-2 (Oregon), one node of 8 Nvidia A100-SXM4-80GB GPUs
Period measured
not stated
Method
Each model ran 1,000 inferences on each of 3 datasets, repeated 10 times without batching, with energy measured by Code Carbon; Table 2 gives the mean (0.047 kWh) and standard deviation (0.03) of energy per 1,000 queries for text generation.
Claimed by
Company: Alexandra Sasha Luccioni, Yacine Jernite (Hugging Face) and Emma Strubell (Carnegie Mellon University)
Paid for by
not stated
Dispute
none
History
Added.
Typical ChatGPT query on GPT-4o, Epoch AI estimate
0.3 Wh per ChatGPT query
We find that typical ChatGPT queries using GPT-4o likely consume roughly 0.3 watt-hours, which is ten times less than the older estimate.
A typical ChatGPT query with about 500 output tokens
Scope
data centre
Counts
H100 GPU power; server and data centre overhead (about 1,500 W per GPU at peak, at 70% average power)
Leaves out
model training and other upstream costs (discussed separately, not in the figure); extra cost of long inputs (discussed separately)
Model or version
GPT-4o (assumed about 100 billion active parameters)
Place
not stated
Period measured
not stated
Method
Model-based estimate from assumed FLOP per token, 500 output tokens, Nvidia H100 throughput at 10% utilisation, and about 1,500 W per GPU including server and data centre overhead at 70% average power.
Claimed by
Independent: Epoch AI (Josh You)
Paid for by
not stated
Dispute
none
History
Added.
GPT-3 water for 10 to 50 medium-length responses, Making AI Less Thirsty
500 mL of water (consumption) per 10 to 50 medium-length responses
Additionally, GPT-3 needs to “drink” (i.e., consume) a 500ml bottle of water for roughly 10 – 50 medium-length responses, depending on when and where it is deployed.
10 to 50 medium-length GPT-3 responses (each about 800 words or fewer of input and 150 to 300 words of output), depending on when and where the model is deployed
Scope
data centre
Counts
on-site (Scope 1) water consumed for data centre cooling; off-site (Scope 2) water consumed in generating the electricity
Leaves out
not stated
Model or version
GPT-3 (175 billion parameters)
Place
Microsoft's U.S. data centres (U.S. average and six state locations)
Period measured
not stated
Method
Per-request server energy of 0.004 kWh (from OpenAI's estimate for GPT-3) multiplied by on-site water usage effectiveness plus PUE times the off-site electricity water intensity, using Microsoft-reported and location-specific values.
Claimed by
Independent: Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren (UC Riverside, UT Arlington)
Paid for by
not stated
Dispute
none
History
Added.
Llama 3.1 405B text response, MIT Technology Review analysis
a mean of 1.86 Wh per text response (published as: an estimated 6,706 joules total (6,706 J / 3,600 = 1.863 Wh))
On average, this model needed 3,353 joules, or an estimated 6,706 joules total, for each response.
One response from Llama 3.1 405B to a variety of text-generating prompts
Scope
data centre
Counts
measured GPU energy (3,353 joules per response on average); estimate for cooling, other computations and other demands (GPU energy doubled)
Leaves out
not stated
Model or version
Llama 3.1 405B
Place
not stated
Period measured
not stated
Method
GPU energy measured on Nvidia H100 GPUs by Jae-Won Chung (ML.Energy, University of Michigan), then doubled to estimate the whole operation, following a 2024 Microsoft paper.
Claimed by
Independent: MIT Technology Review (James O'Donnell and Casey Crownhart)
Paid for by
Tarbell Center for AI Journalism (grant)
Dispute
none
History
Added.
Llama 3.1 8B text response, MIT Technology Review analysis
0.032 Wh per text response (published as: an estimated 114 joules (114 J / 3,600 = 0.0317 Wh))
When tested on a variety of different text-generating prompts, like making a travel itinerary for Istanbul or explaining quantum computing, the model required about 57 joules per response, or an estimated 114 joules when accounting for cooling, other computations, and other demands.
One response from Llama 3.1 8B to a variety of text-generating prompts
Scope
data centre
Counts
measured GPU energy (about 57 joules per response); estimate for cooling, other computations and other demands (GPU energy doubled)
Leaves out
not stated
Model or version
Llama 3.1 8B
Place
not stated
Period measured
not stated
Method
GPU energy measured on Nvidia H100 GPUs by Jae-Won Chung (ML.Energy, University of Michigan), then doubled to estimate the whole operation, following a 2024 Microsoft paper.
Claimed by
Independent: MIT Technology Review (James O'Donnell and Casey Crownhart)
Paid for by
Tarbell Center for AI Journalism (grant)
Dispute
none
History
Added.
Average ChatGPT query, per OpenAI's chief executive
a mean of 0.34 Wh per ChatGPT query
a mean of 0.32 mL of water (kind not known) per ChatGPT query (published as: about 0.000085 gallons of water (converted at 1 US gallon = 3,785.41 mL))
People are often curious about how much energy a ChatGPT query uses; the average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.
Google's August 2025 paper notes that the disclosure gives no explanation of the measurement boundary or method, so it cannot be compared with other estimates. https://arxiv.org/abs/2508.15734
History
Added.
Le Chat 400-token response, Mistral life-cycle analysis
1.14 g CO2e per 400-token response
45 mL of water (consumption) per 400-token response
the marginal impacts of inference, more precisely the use of our AI assistant Le Chat for a 400-token response - excluding users’ terminals: 1.14 gCO₂e, 45 mL of water, and 0.16 mg of Sb eq.
The marginal impact of inference for one 400-token response from Le Chat
Scope
not stated
Counts
marginal impacts of inference; upstream emissions, such as from manufacturing servers
Leaves out
users' terminals
Model or version
Le Chat (the study is a life-cycle analysis of Mistral Large 2)
Place
not stated
Period measured
not stated
Method
Life-cycle analysis with Carbone 4 and the French ecological transition agency (ADEME), peer-reviewed by Resilio and Hubblo, following the AFNOR Frugal AI methodology, the GHG Protocol Product Standard and ISO 14040/44; water is Water Consumption Potential and electricity emissions are location-based.
Claimed by
Company: Mistral AI
Paid for by
none
Dispute
none
History
Added.
Median Gemini Apps text prompt, comprehensive boundary
a median of 0.24 Wh per text prompt
a median of 0.03 g CO2e per text prompt
a median of 0.26 mL of water (consumption) per text prompt
Applying our comprehensive measurement methodology, we find that the median energy consumption for a Gemini Apps text prompt in May 2025 is 0.24 Wh. ... a median Gemini Apps text prompt generates 0.03 gCO2e and consumes 0.26 mL of water when measured comprehensively.
active AI accelerators (prefill and decode, including networking between accelerators in the same AI computer); host CPU and DRAM; idle machines provisioned for reliability and low latency; data centre overhead (PUE); embodied (Scope 1 and Scope 3) emissions of the AI accelerators and host CPU and DRAM (CO2 figure); water consumed for cooling in data centres (water figure)
Leaves out
external networking; end-user devices, including edge compute; LLM training and data storage
Model or version
Gemini Apps (all models serving the app, including supporting models for scoring, ranking, classification and routing)
Place
Google data centre fleet (fleet-wide average emission and water factors)
Period measured
May 2025
Method
Measured from internal telemetry across Google's AI serving fleet; emissions apply the previous calendar year's fleet market-based grid emission factor plus embodied Scope 1 and 3 emissions, and water applies the fleet consumptive water usage effectiveness (ISO WUE Category 2) to energy per prompt.
Claimed by
Company: Google
Paid for by
none
Dispute
Researchers quoted in press coverage, including Shaolei Ren (UC Riverside) and Alex de Vries-Gao, said the water figure leaves out water consumed by the power plants that supply the data centres, and that market-based emissions do not show the carbon actually emitted. https://www.pcworld.com/article/2886730/experts-are-skeptical-about-googles-ai-water-consumption-claims.html
The calculator uses this figure: AI chat prompt (Google's median Gemini text prompt, May 2025).
History
Added.
Median Gemini Apps text prompt, active accelerators only
a median of 0.1 Wh per text prompt
a median of 0.02 g CO2e per text prompt
a median of 0.12 mL of water (consumption) per text prompt
When we apply this non-comprehensive methodology that only considers active TPU and GPU consumption, we estimate the median Gemini text prompt uses 0.10 Wh of energy, emits 0.02 gCO2e, and consumes 0.12 mL of water. This is an optimistic scenario at best and substantially underestimates the real operational footprint of AI.
The median Gemini Apps text prompt in May 2025, measured with a narrower method Google calls the existing approach
Scope
servers
Counts
active AI accelerators (TPU and GPU) only
Leaves out
host CPU and DRAM; idle machines; data centre overhead (PUE); external networking; end-user devices; LLM training and data storage
Model or version
Gemini Apps (all models serving text prompts)
Place
Google data centre fleet, subsampled to the 10% most energy-efficient data centres
Period measured
May 2025
Method
Google's own narrow comparison case, counting only active accelerator energy and sampling the most energy-efficient data centres; the companion paper (arXiv 2508.15734) states that all its emissions figures are market-based and its water figures are consumption.
Claimed by
Company: Google
Paid for by
none
Dispute
none
History
Added.
Long DeepSeek-R1 query on DeepSeek's own servers, How Hungry is AI? benchmark
a mean of 29.078 Wh per long query (10,000 input, 1,500 output tokens)
17 g CO2e per long query (10,000 input, 1,500 output tokens)
DeepSeek-R1 and DeepSeek-V3 hosted on DeepSeek’s own servers consume 29.078 Wh and 13.162 Wh, while the same models on Azure use just 7.410 Wh and 3.696 Wh, over 70% less energy. ... DeepSeek-R1 (DS) consumes over 200 mL of water and emits approximately 17 gCO2e per long query
One long-form DeepSeek-R1 query (10,000 input tokens, 1,500 output tokens) served from DeepSeek's own data centres
Scope
data centre
Counts
GPU power; non-GPU server power (CPUs, SSDs, network, cooling control electronics); data centre overhead (PUE)
Leaves out
embodied emissions and water from hardware manufacturing and supply chains (Scope 3); training
Model or version
DeepSeek-R1, hosted by DeepSeek
Place
DeepSeek data centres in China (PUE and WUE taken as the average of the thirty most efficient data centres in China; carbon factors from the regions of its known or reported deployments)
Period measured
not stated
Method
Estimated from public API throughput and latency data combined with inferred hardware configurations (Nvidia H800), rated GPU and non-GPU power, PUE and regional carbon intensity, reported as the mean of Monte Carlo draws.
Claimed by
Independent: Nidhal Jegham et al. (University of Rhode Island and others)
Paid for by
not stated
Dispute
none
History
Added.
Short GPT-4o query, How Hungry is AI? benchmark
a mean of 0.42 Wh per short query (100 input, 300 output tokens)
A single short GPT-4o query consumes 0.42 Wh (±0.13 Wh), exceeding the footprint of a Google search (0.30 Wh) by approximately 40%.
One short GPT-4o query (100 input tokens, 300 output tokens)
Scope
data centre
Counts
GPU power; non-GPU server power (CPUs, SSDs, network, cooling control electronics); data centre overhead (PUE)
Leaves out
embodied emissions and water from hardware manufacturing and supply chains (Scope 3); training
Model or version
GPT-4o (March 2025 version)
Place
Microsoft Azure data centres (Azure-reported PUE; carbon and water factors from the locations of Microsoft data centres)
Period measured
not stated
Method
Estimated from public API throughput and latency data combined with inferred hardware configurations, rated GPU and non-GPU power, and provider-specific PUE, reported as the mean of Monte Carlo draws (the paper's energy table gives 0.423 ± 0.085 Wh for the same case).
Claimed by
Independent: Nidhal Jegham et al. (University of Rhode Island and others)
Paid for by
not stated
Dispute
none
History
Added.
Frontier-scale model query on H100 nodes, Microsoft researchers' estimate
a median of 0.31 Wh per query
For frontier-scale models (>200B parameters) on H100 nodes, we estimate a median energy of 0.31 Wh/query (IQR 0.16–0.60), indicating widely cited estimates are overstated by 4–20×.
Bottom-up Monte Carlo model from token throughput, node power and PUE under production serving assumptions (batching, FP8); interquartile range 0.16 to 0.60 Wh.