GPT-4o Environmental Impact
OpenAI's standard multimodal model
- Architecture
- Multimodal Transformer (decoder-only)
- Context
- 128,000 tokens
- Provider
- OpenAI
Energy per query
0.43 Wh
about the same as a Google search (0.3 Wh)
CO2 per query
0.19 g
US East (Virginia) grid (450 gCO₂/kWh)
Water per query
2 mL
~667 queries to fill 1 litre
Processing location
Azure US East / Sweden
Provider
OpenAI
Category
Text / Chat
Grid carbon intensity
450 g CO2/kWh (25% renewable)
How does GPT-4o compare?
Detailed Breakdown
Energy Consumption
GPT-4o consumes approximately 0.43 Wh per short query (100 input / 300 output tokens). Energy scales dramatically with context length: medium queries (~1,000 tokens each way) use approximately 1.2 Wh, while a 100K-token input can reach ~40 Wh. Sam Altman disclosed 0.34 Wh as an average; Epoch AI independently estimated 0.3 Wh; Jegham et al. measured 0.42 Wh (±0.13). We use the Jegham figure as it includes full infrastructure overhead.
Power Source & Carbon
OpenAI runs primarily on Microsoft Azure data centers, with key inference regions in US East (Virginia) and Sweden Central. Microsoft has contracted 34 GW of carbon-free electricity across 24 countries, but its location-based Scope 2 emissions doubled from 4.3M to ~10M metric tons CO2 between 2020 and 2024. The Virginia data center hub sits on a grid with approximately 450 g CO2/kWh carbon intensity — one of the dirtier grid regions in the US. OpenAI has also expanded to AWS and Google Cloud as of 2025.
Water Usage
Each GPT-4o short query consumes approximately 1.5 mL of water, based on Jegham et al. infrastructure-aware estimates. Sam Altman disclosed approximately 0.32 mL per query as an average. Water use scales with energy consumption — longer queries with more tokens use proportionally more. Microsoft's Azure data centers use evaporative cooling systems, though Microsoft has launched zero-water cooling designs in late 2024 to reduce this. Microsoft's average Water Usage Effectiveness (WUE) has improved 39% from 0.49 to 0.30 L/kWh between 2021 and 2024.
About GPT-4o
GPT-4o is the workhorse behind ChatGPT — the model that most people interact with when they use OpenAI's flagship product. With over 100 million weekly active users, it is almost certainly the single largest source of AI inference energy consumption in the world. At 0.43 Wh per query, it sits in a middle ground: efficient enough that casual use has a negligible footprint, but at the scale of hundreds of millions of daily queries, its aggregate impact is substantial. Understanding GPT-4o's environmental cost matters because it is, for many people, their entire experience of AI.
GPT-4o in Context
At global scale
With an estimated 100M+ daily users averaging 10 queries each, GPT-4o consumes roughly 430 MWh of electricity per day — enough to power 14333 homes.
Your yearly GPT-4o footprint
At 25 queries per day, your annual GPT-4o usage consumes 3.9 kWh — comparable to running a LED light bulb for a month. That produces 1.7 kg of CO₂.
Key Insights
OpenAI GPT Family
How energy efficiency has evolved across versions.
What does your GPT-4o usage cost the planet?
Use our calculator to estimate your personal environmental footprint based on how often you use GPT-4o.
Calculate My ComputeFrequently Asked Questions
How much energy does GPT-4o use per query?
Each GPT-4o query consumes approximately 0.43 Wh of energy. This is about the same as a traditional Google search (~0.3 Wh).
What is GPT-4o's carbon footprint?
Based on the carbon intensity of Azure US East / Sweden, each query produces approximately 0.19 g of CO2. The grid in this region has a carbon intensity of 450 g CO2/kWh with 25% renewable energy.
How much water does GPT-4o use?
Each query consumes approximately 2 mL of water, primarily used for cooling the data centers that process the request.
How does GPT-4o compare to a Google search?
A GPT-4o query uses about the same as a Google search in terms of energy. A Google search uses approximately 0.3 Wh, while GPT-4o uses 0.43 Wh.
Technical Details
Architecture
Multimodal Transformer (decoder-only)
Context window
128,000 tokens
Release date
2024-05-13
Open source
No
Training data cutoff
2024-06