To calculate the energy consumed by a ChatGPT or Gemini (Google AI Mode and Overview) request, we follow the methodology proposed by Rincè et al. 2025 (and utilized in the Ecologits Calculator, a tool that estimates the environmental footprint of generative AI models at inference). We define Erequest as the energy consumed to generate a response in kWh: \[E_{request}=E_{server}\cdot {PUE} / 1000\] where Eserver is the energy consumed by the server and PUE is the power usage effectiveness of a data center (or in other words, PUE measures the efficiency of a data center by comparing the total amount of power entering the facility to the portion used directly by information technology (IT) equipment). To determine PUE, we use self-reported values by Microsoft and Google for their data centers, as OpenAI reportedly leases Microsoft data centers for ChatGPT models (Microsoft 2025; Google 2026). Unfortunately, the locations Microsoft has values for are limited so we are forced to use high-level values for the Americas, Asia Pacific, Europe & Middle East & Africa, or Global ranging from 1.16 to 1.28. This is an area of uncertainty due to high-level values and use of assumptions as the specific, local-level data we desired was not available. On the other hand, Google reports PUE values for each of its data centers for the year of 2025, giving specific values for certain countries and states (where their data centers are located) ranging from 1.05 to 1.13.
Eserver is then defined as: \[E_{server} = E_{server\backslash GPU}+{GPU}\cdot E_{GPU}\] where Eserver\GPU considers server energy consumption without GPUs, EGPU is the amount of energy consumed by a single GPU throughout a request period based on the number of output tokens and the model architecture, and GPU represents the total number of GPUs needed for a given search.
Eserver\GPU in Wh is then defined as \[E_{server\backslash GPU}(\delta T) = \delta T \cdot W_{server\backslash GPU}\cdot \frac{GPU}{\#GPU_{installed}}\cdot \frac{1}{B}\cdot \frac{1000}{3600}\] where δT is the generation latency (the time it takes the model to respond to an inference), Wserver\GPU is a fixed power consumption of 1.2kW, #GPUinstalled is assumed to be 8, and B represents a batch size of 64.
The latency is then defined as \[\delta T(\#T_{out},P_{active},\delta T_{request}) = min\{\#T_{out}\cdot f_{L}(P_{active},B), \delta T_{request}\}\] \[f_{L}(P_{active},B) = \alpha P_{active} + \beta B + \lambda\] \[\delta T_{request}=\#T_{out}\cdot T_{per-sec}\] where #Tout is the number of output tokens generated by a request, Pactive is the number of active parameters in billions (we assume a value of 300B for GPT, 240B for Google AI Mode, and 100B for Google AI Overview following Rincè et al. 2025), α = 6.78 × 10-4, β = 3.12 × 10-4, λ = 1.94 × 10-2 (Rincè et al. 2025), and Tper-sec is the assumed number of tokens generated per second (65 for GPT 5.5, 159 for Google AI Mode, and 272 for Google AI Overview), following Artificial Analysis 2026.
GPU, the number of GPUs needed for a single search, is then determined by: \[GPU(P_{total},Q,M_{GPU}) = \left\lceil log_{2}(\left\lceil \frac{M_{model}}{M_{GPU}} \right\rceil ) \right\rceil\] \[M_{model}=1.2 \cdot \frac{P_{total}\cdot Q}{8}\] where Ptotal is the total number of parameters in the model which we assume to be 900B for ChatGPT and Google AI Mode, and 300B for Google AI Overview (Rincè et al. 2025), Q quantized to 16 bits, MGPU is the memory available on one GPU (assumed to be 80GB), and Mmodel represents the required memory to host the LLM for inference.
Lastly, EGPU, the amount of energy consumed by a single GPU for a request, is defined as \[E_{GPU}(\# T_{out},P_{active}) =\# T_{out} \cdot \alpha e^{\beta B}\cdot P_{active} + \lambda\] where α = 1.17 × 10-6, β = -1.12 × 10-2, and λ = 4.05 × 10-5.
To calculate water consumption for each generated response we follow methods from “Making AI Less Thirsty” (Li et al. 2025). This requires determining water usage effectiveness (WUE) for both onsite and offsite of data centers and energy efficiency onsite (PUE) to combine to calculate the total water footprint: \[{WaterOperational} = E \cdot [{WUE}_{onsite} + {PUE} \cdot {WUE}_{offsite}]\] where E is the energy consumed by the model for a request (calculated following the methods above) and \[{WUE}_{offsite} = \frac{\sum_k b_{k,t} \cdot EWIF_{k}}{\sum_kb_{k,t}}\] where bk,t denotes the amount of electricity generated from fuel type k at time t for the grid serving the data center and EWIFk denotes the electricity water intensity factor for fuel type k.
To determine PUE (Power Usage Effectiveness, or energy efficiency of the data center) and WUEonsite (water efficiency onsite a data center) values, we use self-reported values by Microsoft and Google for their data centers, as OpenAI reportedly leases Microsoft data centers for ChatGPT models (Microsoft 2025; Elsworth 2025). Unfortunately, the locations Microsoft reports values for are limited so we are forced to use high-level values for the Americas, Asia Pacific, Europe & Middle East & Africa, or Global ranging from 0.03 to 0.34 L/kWh. Google, on the other hand, reports only 1 average WUEonsite value across all their data centers: 1.15 L/kWh. This is an area of uncertainty due to high-level values and use of assumptions as the specific, local-level data we desired was not available.
To determine WUEoffsite (water efficiency of energy production offsite) values, we found datasets containing the electricity mix of the energy grid by location and water efficiency for different energy production fuel types. For the energy grid, we found a dataset containing percentages of each fuel type in the energy grid by state in the U.S. in the year 2023 from the U.S. EPA's Emissions & Generation Resource Integrated Database (eGRID), a comprehensive source of data on almost all electric power generated in the U.S. This dataset is based on available plant-specific data for all U.S. electricity generating plants that provide power to the electric grid (U.S. Environmental Protection Agency 2025). We also use a dataset of the Canadian energy grid by province from the year 2023 created by Canada Energy Regulator (Canada Energy Regulator 2026). Outside of the U.S. and Canada, we found yearly, country-level electricity generation data for over 200 geographies, collected from multi-country datasets and national sources for the year 2024 (Ember 2026). For water efficiency we found a dataset containing L/kWh values for the following fuel types: wind, solar, gas, geothermal, nuclear, coal, oil, hydropower, and biomass (Jin et al. 2019; Visualizing Energy 2023). These datasets were utilized in combination to calculate the electricity water intensity factor for WUEoffsite.
Carbon emission calculations follow a similar structure to water consumption methods, but are inspired by the methods in Rincè et al. 2025. The carbon emissions, C for a single request are estimated by \[C=E_{request} \cdot F_{em}\] where Erequest is the energy consumed (calculated using methodology above) and Fem is the carbon intensity of the electricity mix being used at the given location. To determine Fem we use the same electricity mix datasets of the U.S. by state, Canada by province, and country level estimates for 200 geographies as we do for water consumption calculations. We use the carbon intensity in g/kWh of the fuel types coal, gas, wind, solar, biomass, hydro, nuclear, and geothermal (reported by Ember 2022). These datasets were utilized in combination to calculate the electricity carbon intensity factor Fem for the location of a single request.