Why AI Carbon Footprint Estimates Range from 0.03g to 68g Per Query
Search "carbon footprint of a ChatGPT query" and you'll find numbers that disagree by three orders of magnitude — from roughly 0.03 grams of CO2 to over 60 grams for the same basic action. That's not because anyone is lying; it's because almost no AI company publishes complete, verified per-query energy data, so every estimate is built on different assumptions about hardware generation, data center efficiency (PUE), how many GPUs a company runs, and how many queries they handle per day. Older estimates from 2022–2023 often assumed inefficient early-generation hardware; newer estimates reflect faster, more efficient chips and better-optimized data centers, so the trend over time is downward even as usage grows.
What We Actually Know: The Only Two Disclosed Figures
As of mid-2026, only two AI providers have published concrete per-query energy numbers. OpenAI's CEO stated in June 2025 that an average ChatGPT query uses about 0.34 watt-hours of electricity. In August 2025, Google published a methodology showing its median Gemini text prompt uses about 0.24 watt-hours, producing roughly 0.03 grams of CO2e — a low figure driven by Google's heavily optimized, partly carbon-matched data centers rather than a different definition of "query." Anthropic, Microsoft, Perplexity, and xAI have not published comparable per-query figures, which is exactly why so many competing estimates exist for those services — they're built by outside researchers reverse-engineering hardware deployments rather than drawn from company disclosures.
This calculator applies the U.S. average grid emission factor (rather than each company's actual, often cleaner, data center power mix) so that all three query types are compared on the same basis you'd use for any other home energy calculation on this site. That makes the numbers a useful order-of-magnitude estimate for your own usage, not an official corporate carbon accounting figure.
What About AI Image Generation?
Image generation is a separate case from text chat, and it's consistently more energy-intensive per request. A peer-reviewed 2024 study in Scientific Reports estimated DALL-E 2's footprint at roughly 2.2 grams of CO2e per image — in the same range as a heavier text query — while Midjourney's own CEO has described image generation as requiring "thousands of trillions of operations," pointing to a similar or higher figure depending on settings. Independent trackers like CalcWolf estimate a typical image generation request lands at roughly 2–30 times the energy of a simple text prompt, with resolution, the number of diffusion steps, and any upscaling all pushing the number higher. Video generation (Sora-style clips) is more intensive still, with some estimates putting a short clip at the energy equivalent of 50+ text queries. As with text, no major image or video generation provider currently publishes verified per-request energy figures, so treat the image-generation option in this calculator as an order-of-magnitude estimate built from the available independent research, not an official figure from any AI company.
Putting the Numbers in Perspective
Even a heavy personal AI habit — a few hundred queries a day, every day — typically works out to a few kilograms of CO2 per year once you run the math. The average U.S. resident's total carbon footprint is about 16 metric tons per year (see our Carbon Footprint Calculator). That means personal AI use usually sits at the gram-to-kilogram level, while your total footprint is measured in tons — roughly a thousand times larger. Cutting back on AI queries is a far smaller lever for your personal footprint than changes to diet, transportation, or home heating and cooling. The more consequential environmental question is about AI's aggregate, industry-wide electricity demand as usage scales into the billions of daily queries — not any single person's habits.
How to Reduce Your AI Carbon Footprint (Without Giving Up AI)
- Match the model to the task. Use a fast, lightweight model for simple questions and save larger reasoning models for tasks that actually need multi-step logic — reasoning-heavy queries can use several times more energy than a simple prompt.
- Avoid unnecessary re-generations. Regenerating a response repeatedly for minor wording tweaks multiplies the energy cost of a single task.
- Batch related questions into one longer prompt instead of many short back-and-forth queries where practical — this reduces the fixed overhead per exchange.
- Keep perspective. If you're looking for the highest-impact places to cut your footprint, home energy, transportation, and diet changes (see our carbon footprint reduction guide) will move the needle far more than AI usage habits.