DriversRecommendedOutdated drivers can make a good PC feel brokenScan driver issues before chasing fixes manually.Scan NowHalloween DealsAmazon USHalloween week deals: useful picks before month-endAmazon US: current deals, useful picks and tech finds.Check DealsSlow PC?RecommendedPC slow today? Run a repair scan before it gets worseResolve common Windows issues and optimize system performance.Scan Now×
FREEDOM251

Freedom251 › Blog › AI energy use

How to Estimate the Energy and Water Costs of Running an AI Model

By Freedom251 Team · · 10 min read

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

There is no universal energy or water cost per AI prompt. A defensible estimate starts by defining the workload and measurement boundary: which model and task, how much input and output work it performs, what hardware serves it, how fully that hardware is used, whether idle capacity and host systems are counted, and whether “water” means on-site cooling water or a broader lifecycle footprint. State the date, geography when known, assumptions and an uncertainty range. If the provider does not publish request-level and facility data, present a scenario—not a falsely precise number.

What an estimate is actually measuring

“Running an AI model” can mean training a model once, serving millions of inference requests, or running a single request in an application. Those activities have different workloads and accounting boundaries. A text completion, image generation job and video render also exercise different amounts and types of hardware.

As an Amazon Associate I earn from qualifying purchases.

Before calculating, write down five things:

  • Activity: training, fine-tuning, batch inference or interactive inference.
  • Task: text, image, video, speech or another workload.
  • Work volume: number of requests, input and output tokens, images, seconds of video, or another reproducible unit.
  • Boundary: accelerator only, IT equipment, or production serving including facility overhead.
  • Scope and date: location or fleet, operating period, and whether water is operational cooling water or a lifecycle estimate.

Token counts are useful descriptors for text inference, but they do not determine energy by themselves. The model architecture, context length, batching, hardware, utilization and software all matter.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Choose the electricity boundary before doing arithmetic

Accelerator-only energy

The narrowest boundary counts the active GPU, TPU or other accelerator while it computes. It is useful for comparing a kernel or a hardware configuration, but it omits equipment needed to serve a production request. It can therefore understate the energy allocated to a user request, especially when utilization is low.

#1 Best Overall
Sale
TOZO A1 Wireless Earbuds Bluetooth 5.3 Light Weight in Ear Headphones | IPX5 Waterproof,2 Mics for AI Calls,Immersive Premium Sound Bass Headset with Charging Case,32 EQ via App Black
  • [Ultra-Lightweight Ear Buds Designed for Small Ears] Each earbud weighs only 3.7g and features a compact, ergonomic in-ear design made especially for small ears. Secure, low-profile, and comfortable for workouts, all-day wear or overnight sleeping.The ultra-slim profile sits flush in your ear with no hard edges poking or rubbing, so you can even sleep on your side without discomfort.
  • [Immersive Stereo Sound with TOZO OrigX Technology] TOZO OrigX tuning delivers clear vocals, balanced mids, and natural stereo sound for music, podcasts, and videos.
  • [Long Battery Life for Daily Use] Get up to 7 hours of playtime on a single charge, with up to 32 hours total using the charging case—ideal for workdays, commuting, and extended listening sessions.
  • [Bluetooth 5.3 & Stable Connection] Bluetooth 5.3 provides fast pairing, stable wireless performance, and reduced dropouts as you move around home or office.
  • [Deep Bass with Clear Vocals] High-performance drivers produce punchy bass while keeping vocals clean and detailed for everyday listening.

IT energy for serving

A fuller IT estimate includes the accelerator plus host CPU and DRAM, networking and storage that support the service. It may also allocate provisioned machines that are powered on but waiting for work. How idle capacity is allocated is a consequential assumption: a busy shared cluster and a lightly loaded reserved cluster can do the same user work with different energy per request.

Facility energy

Data-center overhead includes cooling, power conversion, lighting and other non-IT loads. If your starting figure is IT energy, a simplified facility estimate is:

facility energy (kWh) = IT energy (kWh) × PUE

Do not multiply by PUE a second time when the published figure already includes facility overhead. PUE must match the same site, period and accounting boundary as the IT number.

What’s actually slowing this PC down?

Pick the symptom - the matching free tool is one click away.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Electricity calculation: a repeatable method

  1. Define one accounting unit. For example, one completed text request with its measured input and output tokens, or one generated image at a stated resolution.
  2. Measure or obtain energy per unit. Identify whether the value is accelerator-only, total IT or facility-inclusive, and record hardware, utilization and time period.
  3. Scale to the workload. Multiply energy per request by the number of requests (or by the number of other work units).
  4. Convert units. If the result is in watt-hours, divide by 1,000 to obtain kilowatt-hours.
  5. Convert energy to money separately. Multiply kWh by the applicable electricity tariff, including demand or time-of-use charges only if your calculation is designed to include them.
  6. Publish the assumptions and a range. Use measured percentiles or low/base/high scenarios when utilization, token mix or facility factors are uncertain.

In compact form:

total kWh = (Wh per unit × number of units) ÷ 1,000

electricity cost = total kWh × your tariff ($/kWh)

For a service that reports only IT energy, apply PUE once and label the result. For a provider that reports facility-inclusive energy, use it directly and document that no further PUE adjustment was made.

What published figures show—and why they cannot be generalized

Example Reported value Boundary and conditions What it demonstrates
Gemini Apps median text prompt 0.24 Wh Google’s comprehensive method; median prompts measured in May 2025; includes active accelerator power, host CPU and DRAM, idle machines and data-center overhead. A full-stack provider estimate can be higher than an accelerator-focused figure.
Gemini Apps median text prompt 0.10 Wh Google’s narrower method for the same May 2025 prompt population, excluding several components. The boundary alone can more than double the reported result.
Frontier-scale model query 0.34 Wh median; IQR 0.18–0.67 Wh Microsoft Research bottom-up model for an H100 node under stated workload, utilization and PUE assumptions. A modeled result is conditional on its hardware and operating assumptions.
Frontier-scale query with test-time scaling 4.32 Wh median Microsoft Research scenario using 15 times more tokens than a typical query. Additional reasoning or token generation can materially raise energy per query.

These values describe different systems and boundaries. They are examples, not a universal benchmark. A Gemini Apps median prompt is not interchangeable with an H100-node model, and neither value predicts every commercial model or prompt.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Worked electricity examples

Scaling a provider value to a monthly workload

Suppose you use the reported 0.24 Wh median value only as a scenario for 1,000 comparable text prompts:

Rank #2
Sale
Smart Watch for Men Women Compatible with iPhone Samsung Android Phone 1.69 inch Full Touch Screen IP68 Waterproof Bluetooh Fitness Tracker Heart Rate/Sleep Monitor (Black)
  • 【IP68 Waterproof Multifunctinal Smart Watch】Torjalph fitness activity tracker smart watch is IP68 waterproof technology ,you can wear it when swimming,handwashing,it's rain-proof which means that you perform multiple exercise modes outside without worrying about sweat or rain corroding your fitness smart watch,and this watch equipped with a 230 mAh large capacity battery magnetic charging ,and this sports activity tracker watch can be used 5-7 days If fully charged.
  • 【Larger 1.83'' Full Screen Watch & Multi-dial Selection】1.83 inch full HD color touch screen smart watch for men / women provides excellent 240*284 resolution image quality ,This Bluetooth smart watch offers a better touch experience and remarkable immersion and ,more than 200 watch faces for you to choose from App"Da Fit ", you can change the dial style according to your mood to wear keeping the world on your wrist
  • 【Accurate Activity Tracker Smart Watch 】sports watch activity tracker uses sweat and dust resistant breathable silicone strap, our Fitness Tracker Smart Watch for Android iOS Phones supports 12 sports modes,walking,running,bicycle,swimming,badminton,basketball,football,rope skipping. This bluetooth calls smart watch for men women is lightweight and stylish, bringing more convenient and relaxing sports experience while also providing more protection.
  • 【Your 24 Hour Personal Daily Health Monitor】Fitness watch for women men full touch screen IP67 waterproof bluetooth smartwatch for Android iOS Phones with heart rate blood pressure monitor multifunction Andriod smart watch with steps counter sleep monitor sports fitness watch, monitoring blood oxygen saturation allows you to understand your physical condition in real time to track fitness progress ,to know your health status in time.
  • 【More Functions Bluetooth Call Smart Watch】This fitness activity smart watch full screen for iPhone and Android phones includes a bunch of new features including:The bluetooth smart watch can hang up the phone, receive and display WhatsAPP /Text /Facebook /Twitter /Messenger reminder.Support sedentary reminder, alarm clock, stopwatch, timer,weather information,music control, remote camera control compatible with Android 5.0 and above or iOS 9.0 and above smartphones.

0.24 Wh × 1,000 = 240 Wh = 0.24 kWh

The monetary charge is then 0.24 × your electricity tariff. The calculation is valid only for prompts comparable to that measured population and only under the provider’s comprehensive boundary. It does not include a separate, unverified lifecycle allocation for manufacturing or electricity generation.

Showing a range instead of a single point

For the Microsoft Research modeled case, 1,000 queries would correspond to 340 Wh at the median, with an interquartile range of 180–670 Wh under the same H100-node assumptions. With the test-time-scaling scenario, the median would instead be 4,320 Wh for 1,000 queries. This contrast is a workload scenario, not evidence that every long answer uses exactly 4.32 Wh.

Estimating water: define WUE and the water scope

Water figures are especially easy to misread because “water used” can mean different quantities. Water Usage Effectiveness (WUE) is commonly reported as liters per kilowatt-hour:

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

WUE (L/kWh) = annual water used for the stated purpose ÷ annual IT energy (kWh)

Microsoft defines its WUE as annual liters used for humidification and cooling divided by annual IT kWh. Its FY25 global WUE was 0.27 L/kWh for July 1, 2024 through June 30, 2025, covering data centers Microsoft fully owns and controls that operated for 12 months. That is a fleet-specific operational factor, not a universal value for another provider or location.

On-site cooling-water estimate

When the numerator and denominator have matching boundaries, a simple estimate is:

cooling water (L) = relevant IT energy (kWh) × WUE (L/kWh)

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Use an IT-energy denominator with an IT WUE factor. Do not combine a facility-inclusive energy figure with a WUE defined against IT energy unless you adjust the accounting explicitly.

Rank #3
New Bee Guitar Cable 10ft Electric Instrument Bass AMP Cord | 6.35mm Jack, Durable Design, Stage and Home Studio Use, Low Noise Shielded Electric Cord for Bass, Keyboard, Amp, Mixer
  • [Low Noise High Fidelity] New bee guitar cord with 22AWG oxygen-free-copper center conductors provides enhanced signal transfer, the PVC insulation material and the high-density stranded copper double shield provide noise rejection for quiet operation
  • [Ideal for Choice] New bee 1/4 straight and right angle metal connectors, ensures a secure connection in any instrument, pedal, or amp jack. L shape great fro on stage guitar stands
  • [Wide Compatible] Suitable for mixer, power amplifier, microphone, effector, decoder, equalizer, electoronic organ, electric guitar, electric bass and other electronic equipments. Widely used in the installation of smart home, karaoke speaker, professional stage audio, etc
  • [Easy Disassemble] Making incidental repairs quick and easy! Very nice, flexible cable
  • [Dont Hesite to Order] 24 hours customer services and professional technology team are standing by

Consumptive versus withdrawn water

Google’s prompt-water method uses a Category 2 freshwater WUE defined as consumptive use—water input minus returned water—to represent water evaporated and unavailable for reuse. Google reports 1.15 L/kWh for both 2023 and 2024 in that factor. A withdrawal figure, a discharge figure and a consumptive figure answer different questions and should not be substituted for one another.

Provider examples for water and carbon

Reported result Conditions How to interpret it
0.26 mL of water per median Gemini Apps text prompt Google’s derivation using its 2024 average fleet WUE and the provider’s prompt-energy method. An operational provider estimate for that prompt population; not a complete lifecycle water footprint.
0.03 gCO₂e per median Gemini Apps text prompt Derived with Google’s 2024 average fleet grid carbon intensity. A related provider estimate; it does not turn the water number into a universal benchmark.
1.15 L/kWh freshwater WUE Category 2 Google factor reported for both calendar years 2023 and 2024. A specific consumptive-use factor for the stated fleet and method.
0.27 L/kWh global WUE; 1.17 global PUE Microsoft FY25 reporting population: fully owned and controlled sites operating for 12 months. Fleet-level factors with Microsoft’s stated definitions and fiscal-year boundary.

For the Google prompt result, the paper’s method subtracts data-center overhead energy from total energy before multiplying by the selected freshwater WUE. That produces an operational prompt-water estimate, not a full account of water consumed to generate electricity or manufacture hardware.

Why water footprint estimates vary by place and time

On-site cooling is only one part of an AI system’s water footprint. Electricity generation can consume water, and semiconductor, server and facility manufacturing add embodied impacts. Local climate, cooling design, grid mix, season and the hour of operation can change results. The 2024 study Making AI Less “Thirsty” discusses these spatial and temporal differences and notes that public runtime and embodied-water data remain limited.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

If you include indirect water, create a separate line item with its own geography, time period and factor source:

  • Operational site water: cooling and humidification at the data center.
  • Electricity-related water: water associated with the generation mix supplying the servers.
  • Embodied water: manufacturing and construction impacts allocated to the hardware and facility.

Do not add a generic multiplier to a provider’s on-site figure. If an indirect factor is unavailable, say that the estimate excludes it.

Converting an estimate into a transparent worksheet

A spreadsheet or script should preserve the inputs instead of hiding them in one headline number. Use one row per workload scenario with columns such as:

Input Example entry Why it matters
Model and task Text inference Different models and modalities have different compute patterns.
Requests Measured count Sets the scale of the estimate.
Input/output work Token counts or image/video units Captures workload mix; tokens alone are not an energy meter.
Energy per unit Wh, with percentile or scenario Core measured or modeled quantity.
Energy boundary Accelerator, IT or facility Prevents accidental comparisons of unlike values.
PUE Only when starting from IT energy Adds facility overhead once.
Electricity tariff Your applicable $/kWh Converts physical energy to money; tariffs vary by location and contract.
WUE Matching L/kWh factor Converts energy to the stated operational water quantity.
Water scope Consumptive on-site, withdrawal, or lifecycle Defines what the liters mean.
Date and geography Measurement period and region Factors and grid conditions change over time and place.

Then calculate separate outputs:

  1. total energy = energy per unit × work units
  2. facility energy = IT energy × PUE (only when required)
  3. electricity cost = facility or IT kWh × tariff, depending on the selected boundary
  4. operational water = matching IT kWh × matching WUE

Keep low, central and high rows when the uncertain inputs are utilization, token volume, PUE or WUE. The resulting interval is more honest than adding decimal places to an uncertain average.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Support on Ko-Fi

How to compare two published estimates fairly

Put the estimates side by side only after aligning the following axes:

Rank #4
BSIMB 15.6" Digital Family Calendar & Chore Chart, No Subscription Fees
  • Digital Family Calendar: BSIMB digital calendar wall touch screen puts an end to your chaotic life. Designed for busy families, the digital calendar keeps you organised and in step with your family. Assign different colours and tasks to each family member to make family scheduling easier and more intuitive.
  • Easy Setup and Auto-Sync: Plug in, connect to WLAN and auto-syn your electric calendar in minutes. Digital wall calendar compatible with Google calendar, iCloud calendar, Outlook calendar, Cozi calendar and Yahoo calendar. Add events directly on the device or through the ecalendar app. Electronic calendar stays current whether you're at work or traveling—no manual updates needed.
  • All-In-One Home Assistant: Smart calendar builds habits and encourage independence with personalised chores. Display daily meal plan to reduce the daily ‘What's for dinner?’ stress. To-do lists keep track of everything you want. Turning chores into valuable achievements with stars and rewards, making life easier and more enjoyable.
  • Manage Anytime Anywhere: BSIMB electronic calendar for family stay connected with the powerful ecalendar app. Access and manage your digital wall calendar anytime, anywhere with free features like stars & rewards, photo screensaver, chore charts, meals plans, to-do lists etc. Electronic planner paid features like Magic Import tailor advanced tools for users with special needs.
  • Stay in Touch with Loved Ones: Share photos or videos from your phone to electronic wall calendar with app, instantly update your life moment, and keep in touch with your family, especially those who are far away; you can also surprise your smart calendar digital wall by sending him a wish card; it's also great for keeping tech-illiterate elders in touch with the youngsters.
  • same model class and task;
  • similar input and output token volume or equivalent work;
  • same serving hardware or a clearly described hardware model;
  • similar utilization, batching and idle-capacity treatment;
  • same inclusion of host CPU, memory and networking;
  • same treatment of facility overhead and PUE; and
  • same water definition, denominator, geography and time period.

If an axis cannot be aligned, retain both figures with their qualifications rather than ranking one as “more efficient.” A narrower accelerator-only result can look better simply because it omits production-serving components.

Common mistakes and how to correct them

Turning a token count into an energy claim

Mistake: assuming a fixed Wh-per-token value applies to every model and context.

Correction: use token counts to describe the workload, then obtain a model- and hardware-specific measurement or scenario that includes utilization and serving overhead.

Free tools Windows power users keep installed

One-click scans. No signup required.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Applying PUE twice

Mistake: multiplying a facility-inclusive provider number by PUE again.

Correction: identify whether the source reports IT or facility energy and apply PUE once only to IT energy.

Mixing incompatible WUE numbers

Mistake: using a site’s cooling WUE with a fleet-wide energy figure, or treating consumptive liters as withdrawals.

Correction: match numerator, denominator, site population and period; label the water quantity explicitly.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Calling an operational figure a lifecycle footprint

Mistake: presenting cooling water as the total water cost of AI.

Correction: list operational, electricity-related and embodied components separately, and mark unavailable components as excluded.

Reporting a mean without checking the distribution

Mistake: using an arithmetic mean when a small number of long or unusually expensive prompts skews the distribution.

Correction: report the statistic that matches the source’s method. Google’s paper says it uses the daily median as its aggregate value because the distribution is skewed and the arithmetic mean is sensitive to extreme prompts. That is a statement about Google’s analysis, not a universal statistical rule.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Hiding uncertainty behind a precise decimal

Mistake: publishing a single number when provider-side runtime, facility or lifecycle inputs are missing.

Correction: show a range or named low/base/high scenarios and identify which unresolved input drives the spread.

What you can and cannot claim

You can claim that a defined workload consumed an estimated amount of energy under a stated boundary, or that multiplying matching IT kWh by a stated WUE gives an operational cooling-water estimate. You cannot claim that one published prompt value applies to all models, providers, regions or prompt lengths. Nor can an on-site WUE result be presented as a full lifecycle water footprint without additional electricity-generation and embodied-impact accounting.

No public source reviewed here supplies enough request-specific runtime, facility and lifecycle data to calculate an exact footprint for every commercial model request. When those details are unavailable, preserve the unknowns in the result instead of silently filling them with a generic factor.

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

A concise reporting template

Use wording like this in a technical report:

“For [model/task], we estimate [low–high] Wh per [unit] for [date and location]. The boundary includes [accelerator/host/idle/facility items]; PUE was [value or not applied]. At [number] units, that is [kWh], and at [tariff] the electricity charge is [currency amount or formula]. Operational water is estimated at [liters] using [WUE] L/kWh, defined as [consumptive/withdrawal and scope]. Electricity-generation and hardware-embodied water are [included/excluded/unknown].”

This format keeps the number attached to the workload, boundary, date and uncertainty that give it meaning.

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.

AI energy useAI water usedata center efficiencyPUESustainabilityWUE

Leave a Reply

Your email address will not be published. Required fields are marked *

Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.

Recommended PC Tool
Recommended PC Tool
Windows Errors? Fix Them Before They SpreadFree repair scan
Crashes, No Sound, or Screen Glitches?Free driver scan

Two free Windows tools

One Free Minute Could Fix That PC

Before you go - each of these free tools takes about a minute and tackles what quietly slows a Windows PC down.

Special offer. View Outbyte info, uninstall instructions, EULA, and Privacy Policy.