How Much Electricity It Actually Takes to Use AI May Surprise You

A normal text-only AI prompt currently uses roughly 0.24 to 0.34 watt-hours (Wh) of data-center electricity. That is a fraction of a watt-hour—roughly the energy used by a 100-watt television in less than nine seconds. It is considerably lower than the older estimate of about 3 Wh per ChatGPT query that still appears widely online.

That number is not a universal price tag for AI use. Long documents, reasoning modes, agentic workflows, image generation, and video generation can require many times more electricity. The best way to understand the impact is to separate the energy used by one task from the much larger electricity demand created when billions of tasks, model training runs, and data centers are counted together.

#	Preview	Product	Price
1		Charger Compatible with HP Laptop Computer 65W 45W Smart Blue Tip Power Adapter	$9.90	Buy on Amazon
2		Ostrich Charger Compatible with Dell Laptop Computer 65W 45W Round Tip Power Adapter	$9.90	Buy on Amazon
3		VJYUIJAY Universal 65W USB C Laptop Charger Compatible with HP chromebook Lenovo Dell Acer Asus...	$9.99	Buy on Amazon
4		65W 45W USB C Laptop Power Replacement Adapter Charger for Lenovo Chromebook/Yoga/ThinkPad/HP...	$15.99	Buy on Amazon
5		Replacement for HP Laptop Charger 65W 45W Computer Charger Smart Blue Tip Power Adapter	$27.99	Buy on Amazon

The short answer: about 0.24–0.34 Wh for an ordinary text prompt

The best public evidence available as of August 9, 2026 puts a typical text interaction with a modern, large-scale AI service in the range of approximately 0.24–0.34 Wh of operational electricity.

• 0.24 Wh: Google’s measured median for a Gemini Apps text prompt in May 2025.
• About 0.30 Wh: Epoch AI’s independent estimate for a typical GPT-4o text query.
• 0.34 Wh: The average ChatGPT-query figure stated by OpenAI CEO Sam Altman in June 2025.

These figures are useful reference points, not a universal electricity tariff. They involve different products, models, dates, measurement boundaries, and definitions of a query. Google reported a production measurement with a detailed methodology; Epoch AI produced a model-based estimate; and Altman’s ChatGPT figure was not accompanied by equivalent public methodology.

#1 Best Overall

Charger Compatible with HP Laptop Computer 65W 45W Smart Blue Tip Power Adapter

• POWER SPECS - Output Max: 19.5V 3.33A 65W (also compatible with 19.5V 2.30A 45W) Input Volt Range: 110-240V /(Blue Tip Size: 4.5 x 3.0mm) Compatible with HP Laptop Power Supply AC Adapter Charger Cord
• Compatible with HP Smart Blue Tip Models 65W/45W: Elitebook 830 / 840 G3 G4 G8 / 850 G3 G4 / ProBook 430 G8 / 440 G8 / 450 G3 G8 / 640 G2 G3 G4 / 650 G2 G3 / Pavilion x360 / Stream 11 Pro G3 / Stream 11 Pro G4 EE / Chromebook 11 EE G2 G3 G4 / Spectre x360 x2 / Notebook 15 / Zbook / Part Numbers: SK90195333 710412-001 741727-001 213349-109 854054-002 L24008-001 L25296-001 L25296-002 SK90A195231 TPN-LA15
• SAFETY- The Leading American Consumer Product Testing Laboratory, Lists This HIBBO Product as Meeting Their Standards for Electrical Safety and Design in The United States and Canada. Don't Buy Potentially Inferior or Dangerous Chargers That Can Harm Your Laptop or Worse
• PORTABLE DESIGN - One Adapter for Many Models Compatible with HP Laptop. And the Weight is So Light Which is An Incredibly Small Footprint the Laptop AC Travel Adapter Slips into Bags, Purses and Even Pockets with Ease While You Go Anywhere
• OUR GUARANTEE - As a Professional Notebook Power Supplier, Our Products are in Compliance with Top Industry Standards, Include Numerous Safety Mechanisms, Including Protection Against Short Circuiting, Overvoltage, Overcurrent, and Internal Overheating. That's Why We Offer an Industry-Leading Return and Exchange Policy: Within 30 Days of Purchase. Additionally, LIFETIME from The Date of Purchase, We Will Exchange Your Product Should it Become Defective

$9.90

Buy on Amazon

For ordinary text use, the practical conclusion is straightforward:

A typical text AI prompt uses about a quarter to a third of a watt-hour, but the cost can rise to several watt-hours or more when the system performs extensive reasoning, processes a long context, generates images, or creates video.

The electricity figures above generally describe provider-side inference—the computing needed to serve the request in a data center. They usually do not include the electricity used by your phone or laptop, the energy used to manufacture hardware, model training, or every part of the network and storage system.

First, what does electricity per AI request mean?

Several terms are easily mixed together:

• Power is the rate of electricity use, measured in watts (W).
• Energy is the amount used over time, measured in watt-hours (Wh) or kilowatt-hours (kWh). One kWh equals 1,000 Wh.
• Inference is the process of running a trained model to produce an answer, image, audio clip, video, or other output.
• Serving is the production infrastructure that receives requests and performs inference for users.
• Training is the earlier process of adjusting a model’s parameters using large datasets. It is a separate electricity cost from serving users.
• Operational electricity covers the electricity used by computing equipment and the facilities supporting it.
• Embodied energy and emissions refer to impacts from manufacturing chips, servers, buildings, and other equipment.

That is why the precise phrase is energy per prompt or electricity per prompt, not watts per query. A watt is a rate; a watt-hour is an amount.

There is also no perfectly stable unit called an AI query. Depending on the source, it might mean one user message, one completed answer, one image, or an entire backend workflow. A single visible question can trigger retrieval, tool calls, retries, safety checks, multiple model calls, or hidden reasoning before the answer appears.

Why Google’s 0.24 Wh figure is especially useful

Google’s production-scale measurement is one of the clearest public disclosures because it explains what is inside its estimate. Google reported that the median Gemini Apps text prompt used 0.24 Wh in May 2025.

In its comprehensive serving boundary, Google included:

• The active AI accelerator.
• Host CPU and DRAM.
• Provisioned machines that were idle but needed to be available for service.
• Data-center overhead, including cooling and power conversion.

The accompanying technical methodology says the estimate excludes the user’s device, external networking, model training, and data storage.

That boundary matters. Google also reported a narrower estimate of about 0.10 Wh when looking only at the active accelerator. Its comprehensive estimate of 0.24 Wh was therefore about 2.4 times higher. Counting only the GPU or other accelerator can understate the electricity required to operate a real service, while counting every possible lifecycle impact would produce a different figure again.

How the other current estimates compare

Epoch AI independently estimated that a typical GPT-4o text query used approximately 0.3 Wh under its assumptions. Its analysis accounts for factors such as model size, token counts, accelerator hardware, utilization, and production serving conditions. It also explains why long inputs can change the result substantially. See Epoch AI’s GPT-4o energy analysis.

Sam Altman stated that an average ChatGPT query used approximately 0.34 Wh in a June 2025 post. That is valuable as a company executive’s indication of scale, but OpenAI has not published a comparable full measurement methodology for the number. It should not be treated as an independently audited figure for every ChatGPT interaction. The statement appears in Altman’s post about AI.

The differences between 0.24, 0.30, and 0.34 Wh are not evidence that one source must be wrong. They may reflect different model fleets, workloads, dates, averages versus medians, and accounting boundaries.

Why the old 3 Wh estimate is misleading for ordinary text use

The often-repeated claim that a ChatGPT query uses approximately 3 Wh came from an earlier model-based calculation. It was not fabricated, but it was a scenario estimate based on assumptions that are less representative of a normal current text prompt.

Epoch AI’s 2025 reassessment placed a typical GPT-4o query at about 0.3 Wh—roughly one-tenth of the older estimate. The main reasons for the difference included:

• More efficient H100 hardware compared with older A100-based assumptions.
• More realistic assumptions about typical input and output lengths.
• Modern production batching, in which requests share hardware efficiently.
• Higher utilization of commercial serving infrastructure.
• Different assumptions about how many model parameters are active for a request.

The 3 Wh figure may still be relevant to an unusually long, inefficient, or computation-heavy workload. It is simply outdated as a headline estimate for an ordinary current text prompt. The more accurate question is not whether 3 Wh is universally wrong, but which workload and measurement boundary produced the number.

Rank #2

Ostrich Charger Compatible with Dell Laptop Computer 65W 45W Round Tip Power Adapter

• [POWER SPECS] Output Max - 65W, 19.5V 3.34A (also for 45W laptop as well), Input Range - 100 - 240V. /(DC Connecter Size: 4.5 x 3.0mm) Compatible with Dell Charger Laptop AC Adapter Power Cord
• Compatible with ALL 65W/45W Dell laptops with round power connector, including Inspiron, XPS, Vostro series & more
• SAFETY- The Leading American Consumer Product Testing Laboratory, Lists This Ostrich Product as Meeting Their Standards for Electrical Safety and Design in The United States and Canada. Don't Buy Potentially Inferior or Dangerous Chargers That Can Harm Your Laptop or Worse
• [OUR GUARANTEE] As a Professional Notebook Power Supplier, Our Products are in Compliance with Top Industry Standards, Include Numerous Safety Mechanisms, Including Protection Against Short Circuiting, Overvoltage, Overcurrent, and Internal Overheating. That's Why We Offer an Industry-Leading Return and Exchange Policy: Within 30 Days of Purchase. Additionally, LIFETIME from The Date of Purchase, We Will Exchange Your Product Should it Become Defective
• [What You Get] 1 * Power Adapter, 1 * Power Cord, 1* 4.5mm x 3.0mm to 7.4mm x 5.0mm Converter Cable, which can be used for 7.4mm * 5.0mm connector laptops

$9.90

Buy on Amazon

The workload ladder: not all AI requests are alike

Here are indicative figures from the available research. These are not universal prices charged by AI providers, and several are model-based estimates or controlled benchmark results rather than direct measurements of commercial products.

The figures in this table should not be added together as though they were standardized measurements. Their value is showing the size of the workload differences.

Long prompts and uploaded documents can change the math

Text generation has two important computational phases. The system first processes the input context, then generates the output. A long input can therefore add a large upfront cost even if the eventual answer is short.

Under its assumptions for GPT-4o, Epoch AI estimated:

• A typical text query: approximately 0.3 Wh.
• A 10,000-token input: approximately 2.5 Wh.
• A 100,000-token input: nearly 40 Wh.

These are model-based estimates, not direct production measurements of every commercial service. They do not mean that every long answer costs 40 Wh. The actual result depends on input-token count, output-token count, model architecture, serving utilization, and whether the system caches or reuses context.

For example, uploading a long PDF may create a substantial processing cost on the first question. A short follow-up in the same conversation may not repeat the entire document-processing cost in the same way if the service reuses cached context. That behavior varies by product, model, and implementation, so it should not be assumed without provider data.

Why reasoning and agentic AI can cost much more

A standard chatbot may generate a relatively short response. A reasoning or agentic system may do considerably more work before returning one visible answer. It may:

• Generate hidden or visible intermediate reasoning tokens.
• Retry, verify, or compare possible answers.
• Search the web or retrieve documents.
• Execute code.
• Call external tools.
• Generate several candidate answers.
• Maintain a much longer context.
• Make multiple model calls inside one user-visible task.

A Microsoft Research analysis estimated a median of about 0.34 Wh for a frontier-scale query under realistic serving assumptions. When test-time computation used roughly 15 times more tokens, the estimate rose to 4.32 Wh—about 13 times the baseline.

That is a scenario-specific result, not a claim that every reasoning model uses 13 times as much electricity. The important distinction is between:

• A user-visible message: the question and answer displayed in the interface.
• A backend task: potentially several or dozens of model calls, searches, or tool operations.
• An agentic session: a workflow whose electricity use cannot be inferred from the first prompt alone.

The International Energy Agency’s 2026 analysis gives the broader warning that video-generation, reasoning, and agentic tasks can consume hundreds or thousands of times more energy per query than simple text generation. That is a category-level warning, not a universal multiplier for every product.

Image generation is a different category

Images generally require substantially more computation than a short text response because the model must construct many pixels through repeated generation steps. But there is no single reliable number for an AI-generated image.

A 2025 experiment covering 17 image-generation models found median model energy values ranging from 0.086 Wh to 4.08 Wh per image—a difference of up to roughly 46 times. The study found that:

• Doubling image resolution increased energy by about 1.3 to 4.7 times, depending on the model and configuration.
• Prompt length did not have a statistically significant effect under the study’s conditions.
• Quantization did not consistently reduce energy and sometimes increased it in the tested setups.

Those results come from the 2025 image-generation experiment, not from a universal measurement of ChatGPT, Gemini, Midjourney, or another commercial service.

A separate 2024 controlled benchmark published at FAccT reported an average of 2.907 Wh per image across its tested conditions. Its least efficient image model used 11.49 Wh per image. In that benchmark, text generation averaged 0.047 Wh per inference, making image generation more than 60 times as energy-intensive as text generation in that particular test.

The FAccT result is best used to demonstrate the difference between modalities, not to assign a fixed electricity cost to every image. Commercial services may use different hardware, batching, model versions, resolutions, and quality settings. The benchmark is available as a FAccT 2024 paper.

Rank #3

VJYUIJAY Universal 65W USB C Laptop Charger Compatible with HP chromebook Lenovo Dell Acer Asus Samsung Google Computer Type C Power AC Adapter

• USB C Laptop charger:Watt: 65W 45W Input :100-240V 1.5A 50-60Hz Output:5V-3A or 9V-3A or 15V-3A or 20V -3.25A,Connector:USB Type-C; VJYUIJAY Laptop Charger Fast charging Compatible with More USB C laptops,For use with compatible devices only,Do not exceed the rated power,Use certified cables,and Avoid placing in enclosed, high-temperature areas
• VJYUIJAY Computer Charger Type C Compatible with ThinkPad L390 L480 L490 L580 L590 E480 E580 E585 E490 E590 P51s P52s P43s P53s ,T470 T470s T480 T480s T490 T495 T590, X270 X380 X390 X395 X1 Carbon 5th 6th 7th Generation;Yoga S730 720 730 910 920 720-12IKB 720S-13IKB 730-13IKB 910-13IKB 920-13IKB, X270 X280 X380 X390 X395 Yoga, X1 Tablet 2nd 3rd;Flex 11 chromebook ,13 Chromebook 2nd Generation;100e 300e 500e C330 C340 S330 S340 C930 C940 C740 Yoga
• VJYUIJAY Laptop Charger Type C Compatible with Chromebook X360 11 12 14 15;14a-na0020nr 14a-na0010nr 14b-ca0013dx 14b-ca0010nr 14b-ca0015cl 14b-ca0023dx 14b-ca0025cl 14b-ca0036nr 14c-ca0000 14-ca0053dx 14-ca0043cl 14-ca003cl 14-ca0065nr 14-ca061dx 11-ae051wm 11-ae001tu 11-ae027nr 11-ae001nr 11-ae002nr 11-ae010nr 11-ae020nr 11-ae027nr;Spectre X2 X360/Elite X2/Pavilion X2/Elite x2 1012 G1 1012 G2 210 /Envy X2
• VJYUIJAY Chromebook Charger Type C Compatible With Chromebook 3100 3300 3380 3400 3500 5190 5300 5400 7200 7300,Latitude 5420 5520 5320 7410 7310 2-in-1 P28T P29T P30T P86F;XPS 12 9250 XPS 13 9300 9310 9350 9360 9370 9380 XPS 15 9550;3310 2-in-1 3390 2-in-1 5175 2-in-1 7200 2-in-1 7210 2-in-1
• VJYUIJAY 65W USB C Laptop Charger Compatible with Spin 11 13 R13 15 311 315 CP311 CP713 C933 CB5-312T R751T SF713 SP714 CB311 CB314 CB314 CB514 CB515 CB714 CB715 CP5-471 CP311 CP315 CP511 CP713 R721 R751 R752T R851

$9.99

Buy on Amazon

Video generation can be dramatically more energy-intensive

Video generation should be treated separately from both text and images. The system may need to generate or refine many frames at high resolution, preserve motion and consistency between frames, add audio, upscale the result, and repeat the process for multiple variations.

The IEA identifies video generation as an emerging use that can consume hundreds or thousands of times more energy than simple text generation. A 2026 inference-energy analysis also found that some video workloads can use more than 100 times the energy of image generation, while showing large differences based on GPU utilization and task type.

There is no honest universal number for the electricity in one AI video. The result depends heavily on:

• Video duration and frame rate.
• Resolution.
• The number of diffusion or sampling steps.
• The video model’s architecture.
• Audio generation.
• Upscaling, interpolation, and other post-processing.
• The number of variations or failed attempts.
• Whether the service offers a fast or high-quality mode.

A one-second low-resolution preview and a high-resolution cinematic clip are both called video generation, but they are not comparable electricity workloads.

How much energy does AI training use?

Training is the separate, upfront electricity bill behind a model. It involves processing large datasets repeatedly while adjusting the model’s parameters. Inference is the continuing cost of serving users after the model exists.

Public training estimates vary considerably:

These estimates may not cover the same things. Depending on the source, a training figure may exclude or include some combination of failed experiments, preliminary runs, fine-tuning, reinforcement learning, data processing, evaluation, cooling, facility overhead, hardware manufacturing, or later retraining.

Training energy is also not automatically charged to one user. It can be amortized across every inference request made during the model’s useful life. A model that serves billions of requests can spread a large training cost across many interactions. A model that is quickly replaced or serves relatively few users may have a much higher training cost per request when calculated on a lifecycle basis.

Conversely, the ordinary 0.24–0.34 Wh prompt figures generally describe serving electricity and should not be read as including a share of every training run, chip factory, server, or building.

What billions of prompts look like

A small individual electricity cost can become a substantial continuous load when multiplied by very large usage. In July 2025, OpenAI said its tools were receiving more than 2.5 billion messages per day, including more than 330 million per day in the United States, according to its global economic analysis.

Using 2.5 billion messages and the 0.34 Wh average as a purely illustrative calculation:

• 2.5 billion × 0.34 Wh = approximately 0.85 GWh per day.
• That would equal approximately 0.31 TWh per year.

This is not OpenAI’s measured electricity consumption. The message count is not necessarily a count of completed model generations. Messages may use different models and tools; some may be short, cached, rejected, or routed to smaller models; and some may trigger much more backend computation. The 2.5 billion figure is from July 2025 and should not be treated as an August 2026 usage count.

The calculation is useful only for understanding scale: a fraction of a watt-hour multiplied by billions of interactions becomes a large, round-the-clock infrastructure demand. OpenAI also reported, with Harvard researchers, approximately 700 million weekly ChatGPT users in a separate usage study. User counts and message counts are different measures, but both illustrate why per-request efficiency does not make system-wide demand irrelevant.

The broader data-center electricity picture

AI is part of the data-center electricity story, but it is not the whole story. Data centers also run cloud software, websites, databases, video services, enterprise applications, storage, networking, and non-AI workloads.

The IEA’s 2026 update projects global data-center electricity consumption to rise from approximately 485 TWh in 2025 to 950 TWh in 2030, equivalent to about 3% of global electricity demand by 2030. Electricity use by AI-focused data centers is projected to triple over that period.

The same analysis reports that:

• Global data-center electricity demand grew approximately 17% in 2025.
• Electricity use by AI-focused data centers grew approximately 50% in 2025.
• An advanced AI server rack could have peak demand equivalent to roughly 65 households by 2027.
• AI-server power density increased approximately 11-fold from 2020 to 2025 and is expected to rise further.

The 950 TWh projection is for all data centers, not just generative AI. It should not be attributed entirely to AI. A useful analysis keeps separate the electricity used for AI training, AI inference, other cloud services, and the facilities that support all of them.

Rank #4

65W 45W USB C Laptop Power Replacement Adapter Charger for Lenovo Chromebook/Yoga/ThinkPad/HP Chromebook/Dell Laptop USB Type C Power Adapter

• [65W USB C Charger] Input:100-240V~1.2A 50/60Hz, Output:5V-3A 9V-3A 15V-3A 20V-3.25A, Power: 15W/27W/45W/65W
• [Compatible with HP Chromebook Charger] Chromebook X360 11 12 14 15;11-ae027nr 11-ae001nr 11-ae002nr 11-ae010nr 11-ae020nr 11-ae030nr11-ae044cl 11-ae001tu 11-ae051wm 14a-na0020nr 14a-na0010nr 14b-ca0013dx 14b-ca0010nr 14b-ca0015cl 14b-ca0023dx 14b-ca0025cl 14b-ca0036nr 14c-ca0000 14-ca0053dx 14-ca0043cl 14-ca003cl 14-ca0065nr 14-ca061dx 15-db 15-de0000 15-de0523dx 15-bl152nr/ HP Spectre X2 X360;Elite X2, Pavilion X2, HP Elite x2 1012 G1 1012 G2 210 G2; HP Envy X2 and more.
• [Compatible with Lenovo Laptop Charger] Thinkpad T14 T14S T14P T15 T15P T16 E14 E15 E16 S2 2nd 3rd Gen S2Yoga 2nd 3rd and more;Lenovo Chromebook 100e 300e 500e C330 C340 S330 $345 C630 C630-13 C630-13Q50 C340-11 C340-15 S340-14 c740 s940 14w 14e S345-14AST;ThinkPad X1 Carbon T480 T480s T580 T580s X390 P52S,A275 A285 A475A485 T470s T480s T580 T490 E480 E580 E485 E585 E490; LenovoThinkBook; Lenovo Ideapad; Lenovo Chromebook.
• [Compatible with Dell Laptop Charger] Latitude 5440 5430 5420 5410 5400 5550 5540 5530 5520 5510 5500 5350 5340 5330 5320 2-in-1 5310 5300 5590 5490 5290 5190 7650 7640 7530 7520 7450 7440 7430 7420 7340 7330 7320 7310 7350 7370/XPS 12 9250 9250 2-in-1 XPS 13 9340 9333 9310 9305 9300 9350 9360 9365 2-in-1 9370 9380 7390 7380 7370/Dell Chromebook 7486 2-in-1 5190 2 in 1 3400 3380 3110 2-in-1 3100 3100 2-in-1.
• [Your Satisfaction Is Our Priority] If you encounter any issues while using,please feel free to contact our customer service team.We offer 7 x 24h customer service.

$15.99

Buy on Amazon

How much would your own AI use add to your electricity bill?

For ordinary text use, you can make a transparent estimate without pretending it is exact:

Annual kWh = (prompts per day × estimated Wh per prompt × 365) ÷ 1,000

Using 0.30 Wh per ordinary text prompt as a round midpoint:

The $0.15 rate is only an example. Replace it with your electricity rate, and remember that these figures cover estimated provider-side inference rather than the electricity used by your device, home network, or the broader lifecycle of the service.

For mixed workloads, use separate assumptions rather than treating every activity as a text prompt:

Annual kWh = [(text prompts × text Wh) + (reasoning tasks × reasoning Wh) + (images × image Wh) + (videos × video Wh)] ÷ 1,000

Useful published starting ranges are:

• Ordinary text: 0.24–0.34 Wh.
• Reasoning or high test-time computation: approximately 0.3–4.3 Wh in published scenarios.
• Image generation: approximately 0.086–11.49 Wh across the cited studies and test conditions.
• Video: highly variable and potentially orders of magnitude above text and image generation.

These are examples, not standardized rates. If you generate 100 images or repeatedly create video variations, those activities can dominate an otherwise modest text-chat total.

What makes one AI request use more electricity?

1. Output length: Generating more tokens generally requires more decoding work.
2. Input length: Long documents and large context windows increase the computation needed before generation.
3. Model size and architecture: Larger or less efficient models usually require more computation, although mixture-of-experts routing makes simple parameter-count comparisons incomplete.
4. Reasoning depth: Hidden reasoning and test-time computation can multiply the number of generated tokens.
5. Tool use: Search, retrieval, ranking, code execution, and external calls add backend work.
6. Modality: Images and videos generally require more computation than short text. Audio has its own workload characteristics.
7. Resolution and duration: These are particularly important for images and video.
8. Serving utilization: Batching and shared production infrastructure can make each request more efficient than an isolated benchmark.
9. Hardware generation: Newer accelerators can deliver more computation per watt.
10. Data-center overhead: Cooling, power conversion, idle capacity, and facility design change the total.

Important edge cases

Regenerating an answer

Five regenerations are not one prompt operationally. Each regeneration may invoke another model completion, even if the interface displays only one final answer. The same applies to repeatedly asking for minor stylistic changes.

Uploading a long PDF

The initial upload or first question may require significant context processing. A short follow-up may be cheaper if context is cached, but caching behavior varies. Do not multiply a long-document estimate by every follow-up without evidence about how the service handles context.

Using a thinking or reasoning mode

The same user wording can consume different amounts of electricity if one request is routed to a reasoning model and another is answered by a faster model. A visible question does not reveal how many hidden tokens or backend calls were used.

Generating image batches

Four images in one request may use less electricity than four completely separate requests if the service shares setup work. That efficiency cannot be assumed without provider data, however. The safest personal estimate is to count the number of images produced and apply a range.

Free versus paid accounts

Subscription price does not reliably indicate energy per request. Paid users may receive larger models, faster hardware, priority capacity, or higher limits, while free users may be routed differently. The model, workload, and serving conditions matter more than the price label.

AI search versus conventional search

Traditional search may return links after relatively limited ranking and retrieval. AI search may retrieve documents, rank them, summarize them, cite them, and generate a response—but it may also use small specialized models rather than a frontier conversational model.

The frequently repeated comparison that AI uses 10 times as much energy as a Google search is an outdated simplification. The often-cited 0.3 Wh per Google search figure came from Google’s 2009 estimate. It should not be treated as a current universal search baseline. Full Fact’s review explains the historical comparison.

Cloud AI, local AI, and the device in your hand

Most published per-prompt figures focus on the provider’s data center. Google’s Gemini estimate, for example, excludes end-user devices and external networking while including the production AI computer and relevant facility overhead.

Best Value

Replacement for HP Laptop Charger 65W 45W Computer Charger Smart Blue Tip Power Adapter

• 【Compatible with HP Laptop Charger】Input Voltage: 100-240V 50-60Hz Output Max: 19.5V 3.33A 65W (also compatible with 19.5V 2.31A 45W), Connector Size: (Blue Tip Size: 4.5 x 3.0mm)PLEASE CONFIRM THE SPECIFICATIONS OF YOUR POWER ADAPTER BEFORE ORDERING
• 【Compatible with HP computer charger】This product has been tested to ensure compatibility with the following and many more Smart Blue Tip Models 65W/45W:Elitebook 830 / 840 G3 G4 G8 / 850 G3 G4 / Envy 13 15 M6 X360/Pavilion x360 11 13 15/ProBook 430 G8 / 440 G8 / 450 G3 G8 / 640 G2 G3 G4 / 650 G2 G3/ Stream 11 Pro G3 / Stream 11 Pro G4 EE / Chromebook 11 EE G2 G3 G4 /Spectre x360 x2 / Notebook 15 / Zbook
• 【Compatible with HP P/N】710412-001 741727-001 709985-002 709985-004 213349-109 854054-002 AD9043-022G2 714657-001 709985-001 740015-002 854055-004 740015-003 L24008-001 L25296-001 L25296-002 SK90195333 SK90A195231 hp power adapter
• 【safe and reliable】The Leading American Consumer Product Testing Laboratory, Lists This PIOEVTKA Product as Meeting Their Standards for Electrical Safety and Design in The United States and Canada. Don't Buy Potentially Inferior or Dangerous Chargers That Can Harm Your Laptop or Worse
• [ OUR GUARANTEE ] As a professional Notebook power supplier, Our products are in compliance with Top Industry Standards,.Including Numerous Safety Mechanisms and Protection Against Short Circuiting, Overvoltage, Overcurrent, and Internal Overheating. If there are any quality issues with the product within 12 months, please don't hesitate to contact us for after-sales service

$27.99

Buy on Amazon

With cloud AI, the dominant incremental computation generally occurs in the provider’s data center, although your phone or laptop still uses electricity to display the interface, transmit data, and sometimes perform local processing.

With local AI, the balance changes. Your computer’s CPU, GPU, memory, cooling, and display consume electricity directly, and there may be no remote inference electricity for that task. But manufacturing impacts and model-download costs still exist. A small local model may be more efficient for a simple task, while a high-power GPU running for a long time may not be.

Local AI is not automatically greener, and cloud AI is not automatically more efficient. A reasonable local estimate is based on the device’s actual wall power multiplied by the time spent generating the result, while also recognizing that lifecycle comparisons require more than operational electricity.

Electricity is not the same as carbon emissions

The same amount of electricity can produce very different emissions depending on whether it comes from coal, gas, nuclear, hydro, wind, or solar generation. Results also depend on the region, time of day, power contracts, and whether the analysis uses location-based or market-based accounting.

Google estimated that its median Gemini prompt produced approximately 0.03 grams of CO₂-equivalent under its comprehensive methodology. That is a Google-specific result incorporating its data-center locations, energy procurement, accounting choices, and model fleet. It is not a universal conversion from 0.24 Wh to carbon.

Hardware manufacturing and other Scope 3 impacts can add further emissions beyond the operational electricity used during inference. A full lifecycle figure therefore needs a stated boundary and methodology.

Water estimates vary even more by boundary

Water is related to, but different from, electricity use. It can be associated with data-center cooling, electricity generation, semiconductor manufacturing, and server manufacturing. Analysts must also distinguish water withdrawal from water consumption and account for local water stress.

Google estimated 0.26 milliliters of water for a median Gemini Apps text prompt under its comprehensive serving methodology. Mistral reported 45 milliliters of water for a 400-token Le Chat response, excluding users’ terminals, in its lifecycle analysis. The figures are not directly comparable because they use different boundaries, geographies, hardware assumptions, and accounting methods. Mistral’s analysis is available from its environmental-standard disclosure.

The lesson is broader than these two numbers: claims such as AI uses a specific number of drops of water need a date, geography, boundary, and methodology before they can be meaningfully compared.

How to reduce AI electricity use without pretending one user controls the grid

• Match the model to the task. Use a smaller or faster model for simple classification, summaries, formatting, or brainstorming when it performs adequately.
• Reserve reasoning modes for tasks that need them. Deep reasoning is valuable for some problems but unnecessary for every short question.
• Specify the required format and length. A focused request can reduce unnecessary output and repeated revisions.
• Limit repeated regenerations. Give clearer constraints or edit a good response locally rather than requesting many near-identical completions.
• Reduce image and video quality when acceptable. Lower resolution, shorter duration, fewer variations, and fast modes can reduce computation.
• Combine related requests. One well-structured answer may replace several separate calls, although an extremely long prompt can itself increase input-processing energy.
• Reuse useful results. Saving and editing a suitable response avoids regenerating it for minor changes.
• Measure organizational use directly. Companies should seek model-level token counts, accelerator power, utilization, facility overhead, and location-specific energy data rather than relying on generic calculators.

Do not shorten prompts solely because they sound wasteful. Output length, model choice, reasoning depth, resolution, and the number of attempts may matter more. In the 2025 image study, prompt length had no statistically significant relationship with energy under the tested conditions.

What a credible AI-energy claim should tell you

Before comparing two numbers, check whether each source identifies:

• The product or model.
• Whether the task was text, image, audio, video, reasoning, or agentic.
• Input and output length.
• The measurement date.
• Whether the number is a median, mean, estimate, benchmark, or direct measurement.
• Whether active accelerators, host systems, idle capacity, cooling, and facility overhead are included.
• Whether the user’s device, network, storage, training, hardware manufacturing, and failed runs are included.
• The geography and carbon-accounting method if emissions are reported.

A narrow benchmark may be useful for comparing two model configurations. A production measurement may be more useful for estimating real service operation. Neither automatically answers every version of the question.

Frequently Asked Questions

How much electricity does one normal ChatGPT or Gemini text prompt use?

The best current public reference range is approximately 0.24–0.34 Wh of provider-side data-center electricity for an ordinary text prompt. Google reported a measured 0.24 Wh median for a Gemini Apps text prompt, while Epoch AI estimated about 0.3 Wh for GPT-4o and Sam Altman stated an average ChatGPT-query figure of 0.34 Wh. The figures are not perfectly comparable or universal.

Is the old estimate of 3 Wh per ChatGPT query still accurate?

It is outdated as a headline estimate for an ordinary current text prompt. Newer hardware, realistic token assumptions, production batching, and higher utilization help explain why current ordinary-text estimates are closer to 0.24–0.34 Wh. A long-context or computation-heavy task can still reach several watt-hours or more.

Does using AI on my phone use electricity from my home?

Cloud AI mainly performs inference in the provider’s data center, while your phone uses additional electricity for its display, networking, and local processing. Local AI shifts the computation to your own CPU or GPU. Neither cloud nor local AI is automatically more energy-efficient in every situation.

Is AI more environmentally harmful than a normal Google search?

There is no reliable current universal multiplier. The often-repeated comparison uses Google’s old 2009 estimate of 0.3 Wh per search, not a current measurement applicable to every search. AI search can involve retrieval, ranking, summarization, and generation, but it may also use smaller specialized models. The workload and measurement boundary matter.

The Bottom Line

Bottom line: One ordinary text AI prompt is usually a small electricity event—about 0.24–0.34 Wh, or less than nine seconds of a 100-watt television. But that modest individual number changes quickly with long documents, hidden reasoning, tool calls, image generation, and video. The larger environmental and infrastructure story comes from billions of requests, energy-intensive training, hardware, facility overhead, and the data centers being built to serve them.

Quick Recap

Bestseller No. 1

Charger Compatible with HP Laptop Computer 65W 45W Smart Blue Tip Power Adapter

$9.90

Bestseller No. 2

Ostrich Charger Compatible with Dell Laptop Computer 65W 45W Round Tip Power Adapter

$9.90

Bestseller No. 3

VJYUIJAY Universal 65W USB C Laptop Charger Compatible with HP chromebook Lenovo Dell Acer Asus Samsung Google Computer Type C Power AC Adapter

$9.99

Bestseller No. 4

65W 45W USB C Laptop Power Replacement Adapter Charger for Lenovo Chromebook/Yoga/ThinkPad/HP Chromebook/Dell Laptop USB Type C Power Adapter

$15.99

Bestseller No. 5

Replacement for HP Laptop Charger 65W 45W Computer Charger Smart Blue Tip Power Adapter

$27.99

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.