Ask ChatGPT a question, and somewhere in a warehouse full of humming servers, a graphics chip that draws as much power as a hair dryer wakes up for a fraction of a second. That tiny burst of electricity happens billions of times a day. So how much energy does ChatGPT use, really? The honest answer surprises most people: a single text prompt costs about as much electricity as running a standard LED light bulb for a couple of minutes. Yet multiply that by 2.5 billion prompts a day, and you get a power bill that rivals a small city.
The confusion around this topic runs deep. Headlines toss around scary numbers, viral posts claim every question drains a bottle of water, and the actual research often gets buried. That gap matters, because energy decisions shape electricity prices, water supplies, and climate goals in real communities. In this guide, you will learn exactly what one prompt costs in watt-hours, how training compares to daily use, why an AI video costs hundreds of times more than a paragraph of text, how ChatGPT stacks up against streaming and driving, which viral claims fall apart under scrutiny, and what you can do to shrink your own AI footprint.
The Short Answer: What One ChatGPT Prompt Costs in Electricity
Let’s start with the number everyone wants. A typical ChatGPT text prompt uses roughly 0.3 watt-hours of electricity, which equals about one-tenth of the energy needed to run a 10-watt LED bulb for 10 minutes, or roughly 1/40th of a smartphone charge. OpenAI’s CEO Sam Altman put the figure at about 0.34 watt-hours per average query. The independent research group Epoch AI landed in nearly the same place, estimating around 0.3 watt-hours for a standard GPT-4o response of about 500 words.
That figure represents a big drop from the number that dominated news coverage for years. Earlier estimates pegged each ChatGPT query at 2.9 watt-hours, roughly ten times higher. Those older calculations relied on 2023-era hardware, guessed at model sizes, and assumed servers ran far below full efficiency. Chips got faster, models got leaner, and researchers got better data. The old number stuck around anyway, which explains why you still see it quoted today.
Here is how the leading published estimates compare:
| Source | Energy per text query | Notes |
|---|---|---|
| Sam Altman (OpenAI) | 0.34 Wh | Stated average across ChatGPT queries |
| Epoch AI | ~0.3 Wh | Modeled for GPT-4o, ~500 output tokens |
| Google (Gemini) | 0.24 Wh | Median text prompt, full-stack measurement |
| Older 2023 estimates | 2.9 Wh | Based on outdated hardware assumptions |
| Long reasoning query | 3-40 Wh | Models that “think” before answering |
Keep one thing in mind as you read the rest of this article. OpenAI does not publish audited, per-model energy data. Every number here comes from company statements, academic modeling, or measurements of similar systems. Treat them as solid ballpark figures rather than precise meter readings.
How Your Prompt Turns Into Electricity
To understand the numbers, it helps to follow a prompt on its journey. When you hit enter, your text travels to a data center where specialized chips do the heavy lifting. Those chips, usually NVIDIA GPUs like the H100 or newer Blackwell models, each pull between 700 and 1,200 watts at full tilt. A single server rack packed with them can draw 120 kilowatts, about as much as 40 American homes running at once.
The step-by-step process
- Tokenization: The system chops your prompt into tokens, small chunks of text roughly three-quarters of a word each. This step costs almost nothing.
- Prefill: The model reads your entire prompt at once and builds a mathematical memory of it. Long prompts and big attached documents make this step more expensive.
- Decoding: The model generates the answer one token at a time, running billions of calculations for each token. This stage dominates the energy bill.
- Cooling and overhead: Every watt the chips burn turns into heat. Cooling systems, power converters, and networking gear add another 10 to 20 percent on top in a modern facility.
- Delivery: The answer travels back through the network to your screen, adding a tiny fraction of a watt-hour.
That overhead in step four has a name: power usage effectiveness, or PUE. A PUE of 1.1 means the facility spends 10 cents of supporting power for every dollar spent on actual computing. Hyperscale operators like Microsoft, which hosts most of OpenAI’s compute on Azure, report PUE figures between 1.1 and 1.2. Older enterprise data centers often sit closer to 1.5 or higher, which is one more reason the old ChatGPT estimates ran high.
Here is the practical takeaway: output length drives cost more than anything else. Asking ChatGPT for a one-sentence answer might use 0.05 watt-hours. Asking it to write a 2,000-word essay could use 1.5 watt-hours or more. The prompt you type is cheap. The words the model writes back are what cost money and power.
Training vs. Inference: Where the Power Actually Goes
People often assume the giant energy cost of AI comes from building the model. That used to be true. Today, the balance has flipped hard toward everyday use.
Training GPT-3 consumed an estimated 1,287 megawatt-hours of electricity and produced roughly 550 tons of carbon dioxide, according to research published by Google and Stanford. That equals the annual electricity use of about 120 American homes, or the emissions from driving a gas car about 1.4 million miles. GPT-4 required far more. Public estimates put its training run somewhere between 50 and 62 gigawatt-hours, roughly 40 to 50 times GPT-3’s total. Newer frontier models likely cost more still.
Those numbers sound enormous until you compare them to inference, the technical term for running the model after training. If ChatGPT handles 2.5 billion prompts per day at 0.3 watt-hours each, that comes to 750 megawatt-hours daily. At that rate, everyday use burns through the entire energy cost of training GPT-3 in about 41 hours. Even a 50-gigawatt-hour training run gets matched by roughly 67 days of ordinary chatting.
| Activity | Estimated energy | Equivalent |
|---|---|---|
| Training GPT-3 (one run) | ~1,287 MWh | 120 US homes for a year |
| Training GPT-4 (estimated) | ~50,000 MWh | 4,600 US homes for a year |
| One day of ChatGPT prompts | ~750 MWh | 70 US homes for a year |
| One year of ChatGPT prompts | ~274,000 MWh | 26,000 US homes for a year |
Industry analysts now estimate that inference accounts for 60 to 90 percent of the total lifetime energy an AI model consumes. That shift changes the whole conversation. Making models more efficient to run matters far more than making them cheaper to train, and it explains why chip makers pour so much effort into inference-optimized hardware.
Adding It All Up: ChatGPT’s Total Footprint at Scale
Individual prompts look harmless. Scale changes the picture. OpenAI reported handling roughly 2.5 billion prompts per day in 2025, serving around 800 million weekly users. Run the arithmetic at 0.3 watt-hours per prompt and you land near 274 gigawatt-hours per year just for ChatGPT text responses.
To picture that, imagine a mid-sized American city of about 70,000 people. That entire city’s residential electricity use for a year roughly matches what ChatGPT burns answering questions. Add API traffic from businesses, image generation, voice mode, and continuous model training, and OpenAI’s real total climbs well past that figure, likely into the range of one to a few terawatt-hours annually.
Now zoom out further. The International Energy Agency estimated that all data centers worldwide consumed about 415 terawatt-hours in 2024, roughly 1.5 percent of global electricity. The agency projects that number could reach around 945 terawatt-hours by 2030, more than Japan’s total electricity use today. AI drives a large share of that growth, though streaming, cloud storage, cryptocurrency, and ordinary web services still account for the majority.
In the United States, the trend hits harder because data centers cluster in a few places. Northern Virginia, central Ohio, and parts of Texas and Arizona now host massive campuses. A Lawrence Berkeley National Laboratory report found data centers used about 4.4 percent of US electricity in 2023 and could reach 6.7 to 12 percent by 2028. Local grids feel that strain first, which is why utility bills and interconnection queues have become part of the AI conversation.
Here is what those big numbers mean in practice:
- ChatGPT alone remains a small slice of global electricity, well under 0.02 percent.
- The bigger issue is speed of growth, not current size. Grids need years to add capacity.
- Concentration matters more than totals. One county can absorb a huge share of new demand.
- Efficiency gains have so far kept pace with usage growth, but that race stays close.
Not All Prompts Cost the Same: Text, Images, Reasoning, and Video
The 0.3 watt-hour figure applies to a normal text exchange. Ask for something fancier, and the meter spins much faster. Understanding these differences gives you real control over your footprint.
Reasoning models
Newer models that “think” before answering generate a long internal chain of reasoning tokens that you never see. A model might produce 10,000 hidden tokens to deliver a 200-word answer. Since energy scales with tokens generated, a single deep reasoning query can cost 10 to 70 times more than a standard chat response. That puts long reasoning tasks in the 3 to 40 watt-hour range, closer to microwaving a snack than flipping on a light.
Image generation
Research from Hugging Face measured image models at roughly 2.9 watt-hours per generated image for the most demanding systems, though efficient models come in far lower. That means one AI image costs about as much as ten text prompts. Generating four variations to pick your favorite multiplies that by four.
Video generation
Video sits in a completely different league. Analysis published by MIT Technology Review estimated that producing a five-second AI video consumed about 3.4 million joules, which works out to roughly 944 watt-hours. That single clip uses more electricity than 3,000 text prompts, or about as much as running a microwave for an hour straight.
| Task | Rough energy cost | Text prompts equivalent |
|---|---|---|
| Short text answer | 0.05 Wh | 0.2x |
| Standard chat response | 0.3 Wh | 1x |
| Long document draft | 1-2 Wh | 3-7x |
| Deep reasoning query | 3-40 Wh | 10-130x |
| One generated image | ~3 Wh | 10x |
| Five-second video clip | ~940 Wh | ~3,100x |
Consider a real scenario. A marketing team spends an afternoon making a campaign. They run 50 text prompts (15 Wh), generate 30 images and keep three (90 Wh), and produce two 10-second video clips (about 3,800 Wh). The videos account for 97 percent of the session’s energy. If that team wants to cut its footprint, video is the lever, not the chatting.
Water and Carbon: The Two Costs Behind the Kilowatt-Hours
Electricity is only part of the story. Data centers also drink water and, depending on the grid, emit carbon. Both get misreported constantly.
Water use
Most large data centers cool their servers with evaporative systems that consume fresh water. Sam Altman stated that an average ChatGPT query uses about 0.000085 gallons, which comes out to roughly 0.32 milliliters, or about one-fifteenth of a teaspoon. Google published a comparable figure for Gemini at 0.26 milliliters per median text prompt. To put that in perspective, you would need to send about 3,000 prompts to consume one liter of water directly.
Those figures cover on-site cooling. Researchers who include the water used to generate the electricity itself, at power plants, arrive at higher totals. Even so, the widely shared claim that one conversation “drinks a 500ml bottle” came from a 2023 study that assumed 10 to 50 exchanges, older hardware, and a specific water-stressed region. Newer, cooler-running facilities that use closed-loop or air cooling changed the math considerably.
Carbon emissions
Carbon depends entirely on where and when the electricity comes from. Run 0.3 watt-hours on the average US grid at roughly 0.4 kilograms of CO2 per kilowatt-hour, and you get about 0.12 grams of CO2 per prompt. That is less than the emissions from breathing for a minute. Google reported 0.03 grams per Gemini prompt thanks to a cleaner power mix and carbon-free energy purchases.
- A prompt on a coal-heavy grid can emit five to ten times more carbon than the same prompt on a hydro or nuclear grid.
- Time of day matters. Midday solar output can cut a data center’s carbon intensity by half.
- Manufacturing the chips carries embodied carbon that most estimates leave out entirely.
- Sending 1,000 prompts still emits less carbon than driving one mile in a typical gas car.
None of this makes AI free. It does mean the individual guilt many users feel is misplaced. The meaningful decisions happen at the level of data center siting, grid mix, and cooling design, not at the level of whether you ask one more question.
How ChatGPT Compares to Everyday Activities and Rival AI Tools
Numbers without context mean nothing. Here is how a ChatGPT prompt lines up against things you already do without a second thought.
| Activity | Energy used | ChatGPT prompts equivalent |
|---|---|---|
| One ChatGPT text prompt | 0.3 Wh | 1 |
| One Google search (2009 estimate) | 0.3 Wh | 1 |
| Charging a smartphone fully | ~12 Wh | 40 |
| Running a laptop for one hour | ~50 Wh | 167 |
| Streaming video for one hour | ~77 Wh | 257 |
| Boiling a kettle of water | ~110 Wh | 367 |
| Microwaving for five minutes | ~100 Wh | 333 |
| Running a clothes dryer once | ~3,000 Wh | 10,000 |
| Driving one mile in a gas car | ~1,100 Wh | 3,667 |
| One round-trip flight, New York to London | ~1,000,000 Wh per passenger | 3.3 million |
Look at that last row. One transatlantic flight equals more than three million ChatGPT prompts. If you used ChatGPT 100 times every single day for 90 years, you still would not match a single vacation flight. That comparison does not excuse AI’s growth, but it does put personal usage in proportion.
Comparisons between AI tools follow similar patterns. Google reports 0.24 watt-hours for a median Gemini text prompt. Anthropic, Meta, and Mistral run models of broadly similar scale, and independent testing suggests all major chatbots land within the same order of magnitude for standard text. Mistral published a full lifecycle assessment showing a 400-token response from its large model produced about 1.14 grams of CO2 equivalent, higher than the per-prompt figures above because it counted hardware manufacturing and training amortization too.
The bigger differences show up between model sizes, not between companies. A small model running on your own laptop might use 0.005 watt-hours per response. A frontier model in reasoning mode might use 40. That is an 8,000-fold spread within the same category of tool.
Myths and Misconceptions That Keep Spreading
Bad numbers travel fast. Here are the claims that come up most often, and what the evidence actually shows.
- “Every ChatGPT query wastes a bottle of water.” The real figure sits near a third of a milliliter per prompt for on-site cooling. The bottle claim came from an older study measuring dozens of exchanges under worst-case conditions.
- “Saying please and thank you costs OpenAI millions in electricity.” Extra polite words add a handful of tokens. Altman did mention that courtesy costs “tens of millions of dollars,” but that spans billions of users over years and works out to a rounding error per person.
- “AI uses more power than entire countries.” All data centers combined, including every non-AI service, used about 1.5 percent of global electricity. AI represents a fraction of that, though a fast-growing one.
- “Each query costs 2.9 watt-hours.” That number came from 2023 hardware assumptions. Current estimates run roughly ten times lower.
- “Not using AI meaningfully lowers grid demand.” Data centers run at high utilization regardless of any individual user. Personal restraint has almost no measurable effect compared to policy and infrastructure choices.
- “Training is the main energy cost.” Inference now dominates, accounting for most of a model’s lifetime energy use.
One more subtle mistake deserves attention. People frequently compare AI energy use to the alternative of doing nothing, when the fair comparison is doing the same task another way. If ChatGPT drafts a report in 30 seconds that would have taken you two hours on a laptop, the laptop time alone costs about 100 watt-hours. The AI version, even generously counted, might cost 2. Efficiency comparisons need both sides of the ledger.
That said, be careful with rebound effects. When something becomes cheap and easy, people do far more of it. Cheaper text generation does not automatically reduce total energy use if it triggers ten times more content creation. Economists call this the Jevons paradox, and it may be the single biggest factor in AI’s long-term energy story.
Practical Ways to Cut Your AI Energy Footprint
You cannot rewire a data center, but you can make choices that add up, especially if you use AI heavily for work or run a team.
Habits that genuinely help
- Pick the right model for the job. Use a fast, lightweight model for simple questions and save reasoning models for problems that truly need step-by-step thinking.
- Ask for the output length you need. Requesting “three bullet points” instead of letting the model write six paragraphs cuts energy proportionally.
- Write clearer prompts the first time. Every regeneration doubles the cost of that answer.
- Batch related questions into one conversation instead of starting fresh threads that resend context.
- Skip image and video generation when a stock photo or plain text will do. Those tasks dwarf everything else.
- Trim long attachments. Feeding a 200-page PDF when 3 pages matter wastes prefill compute.
- Turn off automatic AI features you never read, like auto-summaries on every email.
Tools and resources worth knowing
Several free resources help you dig deeper. The ML CO2 Impact calculator estimates emissions for training runs. CodeCarbon tracks the energy of code you run yourself. Electricity Maps shows real-time grid carbon intensity by region, which matters if you schedule large batch jobs. Hugging Face maintains the AI Energy Score leaderboard, which benchmarks open models on energy per task. Epoch AI and the International Energy Agency publish regularly updated analysis on AI compute and data center demand.
Questions beginners ask most
- Does a longer conversation cost more? Yes. The model reprocesses earlier messages as context, so thread 40 grows more expensive than thread 1.
- Does voice mode use more energy? Somewhat. Speech recognition and synthesis add processing on top of the text generation, though the increase is modest compared to images.
- Is running an AI model locally greener? Often yes for small models, since you skip data center overhead. But consumer GPUs run less efficiently per calculation, so large models usually stay greener in the cloud.
- Does deleting my chat history save power? No. Storage costs are negligible compared to computation.
- Do paid plans use more energy than free ones? Paid tiers unlock larger models and reasoning modes, so heavy paid usage does typically cost more energy per prompt.
Where AI Energy Use Is Heading Next
Two forces pull in opposite directions right now. Efficiency improves fast, and demand grows faster. Which one wins over the next decade will determine whether AI becomes a footnote or a headline in energy policy.
On the efficiency side, progress has been remarkable. Cost per unit of AI performance has dropped by roughly an order of magnitude per year in some benchmarks. Chip makers keep delivering more calculations per watt with each generation. Techniques like quantization, which shrinks the precision of model weights, and mixture-of-experts architectures, which activate only a slice of the model for each token, cut energy dramatically without much quality loss. Distillation lets companies compress a huge model into a small one that runs at a fraction of the cost.
On the demand side, the picture looks very different. Companies plan multi-gigawatt data center campuses. AI agents that run for hours without human input will generate far more tokens than chat ever did. Video generation is going mainstream. Utilities in several US states have already delayed retiring fossil plants to meet new data center load.
Energy supply is shifting to match. Watch these developments over the next few years:
- Nuclear deals. Tech companies have signed agreements to restart shuttered reactors and fund small modular reactor development specifically for data centers.
- Liquid cooling. Direct-to-chip and immersion cooling handle high-density racks while cutting water and electricity used for air conditioning.
- Custom silicon. Purpose-built inference chips from Google, Amazon, and startups often beat general-purpose GPUs on energy per token.
- Carbon-aware scheduling. Systems that shift flexible workloads to times and places with clean power can cut emissions without cutting service.
- Mandatory reporting. Regulations in the European Union and several US states now require data centers to disclose energy and water use, which will replace guesswork with real data.
- Waste heat reuse. Some facilities already pipe server heat into district heating systems for nearby homes.
Transparency remains the biggest gap. Right now, the public relies on estimates and occasional blog posts because no major AI lab publishes audited, model-by-model energy figures. That will likely change under regulatory pressure. When it does, expect the numbers in articles like this one to get sharper, and possibly to surprise us in both directions.
So where does that leave you? A single ChatGPT prompt uses roughly 0.3 watt-hours of electricity, a third of a milliliter of water, and about a tenth of a gram of CO2. Those amounts sit well below a Google search from a decade ago, far below streaming a show, and in a completely different universe from driving or flying. Training the models cost a lot, but everyday use now dominates the total. And within everyday use, the spread is enormous: a quick answer costs almost nothing, while a five-second AI video costs more than 3,000 text prompts combined.
The real story is not individual guilt. It is scale, speed, and where the power comes from. Billions of small prompts add up to the electricity demand of a mid-sized city, and that demand is growing faster than most grids can expand. The good news is that efficiency has improved roughly tenfold in just a couple of years, cleaner power keeps coming online, and better disclosure rules are arriving. Understanding these numbers puts you ahead of most of the conversation, so use AI where it genuinely helps, skip the wasteful video experiments you will never watch, and push for the transparency that makes honest accounting possible.
