For one person, the carbon difference between asking a cloud chatbot and running a model on your own phone is small either way. Google measured the median Gemini text prompt at 0.24 watt-hours and 0.03 grams of CO2e, and a phone generating a reply locally uses energy in the same ballpark. Pick on-device AI for privacy and offline use; don’t pick it, or avoid it, to save the planet.
The details are more interesting than the headline, though, because they show where AI’s energy actually goes.
How much energy does one cloud AI prompt use? #
In August 2025 Google published the first production measurement of AI serving from a major provider. Its technical paper and summary put the median Gemini Apps text prompt at:
- 0.24 Wh of energy, about what a TV uses in nine seconds
- 0.03 g CO2e of emissions
- 0.26 mL of water, about five drops
Google counted more than the chip doing the math. Of that 0.24 Wh, 58% went to the AI accelerators, 25% to the host CPU and memory, 10% to idle machines kept ready for traffic spikes, and 8% to cooling and power conversion. Google also reported that energy per median prompt fell 33-fold over the previous 12 months, thanks to better models and hardware.
OpenAI hasn’t published a comparable study. Sam Altman wrote in June 2025 that an average ChatGPT query uses about 0.34 Wh, without a methodology.
Two caveats. A median hides the heavy tail: long reasoning chains, image generation and “deep research” jobs use many times more. And the 0.03 g figure reflects Google’s clean-energy purchasing, which your home grid may not match.
How much energy does a phone use to run a local model? #
Nobody publishes a single number for this, because it depends on the phone, the model and the length of the reply. You can estimate it from your battery.
A phone battery rated 5,000 mAh at a nominal 3.87 volts stores about 19 Wh (5 Ah × 3.87 V). Running a language model keeps the GPU busy, so the drain is visible. Suppose a 20-minute chat with 15 replies takes 10% of that battery:
- 10% of 19 Wh is 1.9 Wh.
- Charging loses some energy as heat, so call it a little over 2 Wh from the wall.
- Spread over 15 replies, that’s roughly 0.13 to 0.15 Wh per reply.
That’s an illustration, not a benchmark, but it lands in the same range as the cloud figures. Try it on your own phone: note the battery percentage, chat for a while, and check again.
| Scenario | Energy per reply | Where the number comes from |
|---|---|---|
| Median Gemini text prompt | 0.24 Wh | Google’s measurement, May 2025 data |
| Average ChatGPT query | 0.34 Wh | OpenAI CEO’s blog, no method published |
| Local 4B model on a phone | ~0.1 to 0.2 Wh | Battery arithmetic above, varies widely |
| Boiling 1 liter of water | ~93 Wh minimum | Physics: 4.18 kJ/kg·K × 80 K |
The kettle row is there for scale. One kettle of tea costs about as much energy as a few hundred chatbot prompts from either source.
Why data centers are efficient, and why phones still keep up #
Per unit of computation, a data center wins. Its accelerators run many users’ requests at once in batches, stay busy most of the time, and are built only for this work. A phone’s GPU serves one person, sits idle between messages, and wasn’t designed around language models.
The phone claws back ground in other places. There’s no idle server fleet waiting for traffic, no cooling plant, and no network hop for every message. The text of a chat is only a few kilobytes, so the network share of a cloud prompt is small anyway, but the idle and overhead share in Google’s numbers was 18%.
A local model does have a one-time cost: the download. A 2.7 GB model file crosses the internet once. Estimates of network energy per gigabyte vary by an order of magnitude depending on who’s counting, so treat it as a one-off in the same league as streaming an hour or two of HD video, not something you pay per message.
Energy isn’t carbon: your grid matters #
Carbon is energy multiplied by how dirty the electricity is. The U.S. Energy Information Administration puts 2023 U.S. generation at about 0.81 pounds of CO2 per kWh, which is roughly 0.37 grams per watt-hour.
So a 0.15 Wh local reply charged from an average U.S. outlet works out to about 0.06 g of CO2. That’s in the same range as a cloud prompt, and both are tiny. On a coal-heavy grid a local reply costs more; on a hydro or nuclear grid, less.
The bigger number is the phone itself #
Manufacturing a phone produces far more emissions than charging it for years. Apple’s own product environmental reports attribute most of an iPhone’s lifetime footprint to production, not use. That leads to the one piece of advice that really moves the needle: run local AI on the phone you already have. Buying a new phone just to run bigger models costs more carbon than years of prompts in either direction.
When does the choice actually matter? #
For typical personal use, it doesn’t much. A few habits keep both your battery and your footprint down, and they apply to cloud and local AI alike:
- Leave reasoning mode off unless you need it. Thinking models write out their working, which multiplies the tokens generated. Artificial Analysis found that the small Qwen 3.5 models produced hundreds of millions of output tokens to finish its test suite, several times more than a larger model in the same family.
- Use the smallest model that answers well. Each generated token reads most of the model from memory, so a 5.7 GB model does roughly twice the work per token of a 2.7 GB one. We explain that in which AI models can run on a phone.
- Cap long replies. Ask for a short answer, or lower the max-tokens setting.
Personal LLM makes the second habit easy to follow. It shows a live tokens-per-second readout on every reply and includes a built-in benchmark for any downloaded model. A model that answers faster keeps the GPU busy for less time, which is the part that shows up on your battery. Thinking mode is off by default and you can switch it on per chat. For how to read those speed numbers, see how fast a local LLM runs on your phone.
So which should you pick? #
Pick on other grounds, because carbon won’t decide it.
| If you care most about… | Better choice |
|---|---|
| Keeping chats and photos off anyone’s server | On-device model |
| Working on a plane or with no signal | On-device model |
| The hardest questions and the latest facts | Cloud chatbot |
| Phone battery on a long travel day | Cloud if you have signal, or a small local model |
| Lowest total footprint | Whatever runs on hardware you already own |
For more on what local processing does and doesn’t protect, read what on-device AI is and why it matters for privacy. If you’re curious why phones can do this at all, our explainer on AI chips and GPUs covers the hardware.
Frequently asked questions #
Is ChatGPT bad for the environment? #
Per prompt, the published figures are small: about 0.34 Wh for an average ChatGPT query according to OpenAI’s CEO, and 0.24 Wh for a median Gemini prompt by Google’s measurement. The concern is scale, since billions of prompts a day plus model training add up to real data-center demand. Your personal share from everyday chatting is a tiny fraction of what your fridge or car uses.
Does running AI on my phone drain the battery? #
Yes, noticeably while it’s generating, because the GPU works hard. A long session can take several percent of your battery and warm the phone. Smaller models, shorter replies, and leaving thinking mode off all reduce the drain.
Is on-device AI more energy efficient than cloud AI? #
Per unit of computation, data centers are more efficient because they batch requests on specialized chips. Per reply, a small local model ends up in a similar range because it skips server overhead and runs a much smaller model. Neither is clearly greener for a single user.
How much water does an AI prompt use? #
Google reported 0.26 mL, about five drops, for a median Gemini text prompt, mostly from data-center cooling. A phone running a model locally uses no cooling water, though the power plant that supplies your grid may use some.