In-Depth Guide
The Environmental Cost of AI: What the Data Actually Shows
Concerns about AI's environmental impact range from warranted to substantially overstated, depending on what is being measured and how it is framed. Training a large language model like GPT-4 consumes enormous energy — estimates range from 50 to 500 megawatt-hours, comparable to the lifetime driving of dozens of cars. But that is a one-time cost. The ongoing inference — running the model to answer queries — is far more modest, and inference is what scales with usage.
Per-Query Energy Consumption
OpenAI published figures in 2025 showing ChatGPT consuming approximately 0.34 Wh per query — about 10 times the energy of a Google search. Google published comparable figures for Gemini at 0.24 Wh per query. These are the most credible publicly available per-query figures, and they are the basis for this calculator. For context: a 0.34 Wh query equals 0.00034 kWh. Your refrigerator uses roughly 1–2 kWh per day. A single hour of air conditioning uses 1,000–3,500 Wh. In absolute terms, the energy per AI query is very small.
Why Grid Carbon Intensity Matters More Than Model Efficiency
The CO₂ produced by an AI query depends less on the model's efficiency and more on the carbon intensity of the electrical grid powering the data center serving your request. A query routed through a data center powered by Pacific Northwest hydroelectric power (roughly 50–100 g CO₂/kWh) produces 3–6 grams of CO₂. The same query routed through a Midwest coal-heavy grid (800+ g/kWh) produces 30–40 times more. Choosing between ChatGPT and Gemini has a small effect on your carbon footprint. The location of the data center serving your request has a much larger one — and users generally cannot control this.
AI Carbon in Context
A New York to Los Angeles round-trip flight generates approximately 750 kg of CO₂ per passenger. At 100 AI queries per day at the US average grid intensity (420 g/kWh), a full year of heavy AI usage generates approximately 5.2 kg of CO₂ — about 0.7% of one domestic flight. The environmental impact of individual AI query volume is real but small. The larger and more complex environmental story involves data center construction, hardware manufacturing, water cooling, and the embodied carbon of the chips and servers that run these systems.
What Providers Are Doing
Major AI providers have made significant renewable energy commitments. Google's data centers have been carbon-matched since 2007. Microsoft Azure, which runs OpenAI's infrastructure, has committed to 100% renewable energy by 2025 and carbon negativity by 2030. However, renewable energy "matching" means purchasing renewable energy credits equivalent to consumption — it does not guarantee that the specific electrons powering your query come from a renewable source. The grid that serves a data center at any given moment reflects regional generation mix, which varies hour by hour.