Watts and Watt-Hours — Reading Electricity's Two Numbers Like a Speedometer and an Odometer

Two Numbers on the Dashboard

There are two numbers on a car dashboard. The 80 kilometers per hour that the needle points to, and the cumulative 35,000 kilometers printed in the small window below. The first is how fast you are going right now, and the second is how far you have come in total so far. No driver confuses the two.

Electricity has the exact same two numbers. Watts and watt-hours. Watts tell you how hard you are using it right now like a speedometer, and watt-hours tell you how much you used in total like an odometer. However, the news places these two side by side in the same article. "A data center of X gigawatts" and "X terawatt-hours a year" appear in the same article, but it is hard to know how the two numbers connect just by looking at the article.

This article carries one dashboard all the way through. The only thing borrowed from the analogy is the relationship between the numbers, that multiplying speed by time gives distance.

Watts is how hard you are using it right now, and watt-hours is how much you used in total by using it at that intensity for a certain time.

Watts are the rate, watt-hours are the amount used

Let us suppose an electric kettle has 2,000 watts written on the side, and an electric fan has 50 watts on the back. These numbers are examples used in this article, and they differ for actual products. This number is electrical power. Power is the amount of energy electricity moves in 1 second, measured in a unit called watts. The kettle uses power 40 times more strongly than the fan. If you compare it to a speedometer, the needle of the kettle is pointing 40 times higher than that of the fan.

The energy moved by electricity becomes heat that warms water in the kettle, and movement that turns the blades in the fan. 「Using electricity」 in this article means this.

Now, we boiled water with the kettle for 3 minutes, and ran the fan for 2 hours. Watt-hours are the amount of energy calculated by multiplying power by the time it was used, and using 1 watt for 1 hour is 1 watt-hour. Since 3 minutes is one-twentieth of an hour, the kettle used 100 watt-hours by multiplying 2,000 watts by one-twentieth of an hour. The fan multiplied 50 watts by 2 hours, which is also 100 watt-hours. The rate differs by 40 times, but the amount used is the same.

Transferring this to an odometer, it is the same. A car driven for 3 minutes at 400 kilometers per hour and a car driven for 2 hours at 10 kilometers per hour both travel 20 kilometers. You cannot know how much was used by looking at watts alone, and you cannot know how strongly it was used by looking at watt-hours alone.

There is one spot in the name that is easy to confuse. While speed is distance divided by time, watt-hours are watts multiplied by time. The direction is the opposite of 「watt-hours per hour」. Therefore, a 2,000-watt kettle remains 2,000 watts whether it runs for 1 minute or 5 minutes while turned on, and only the watt-hours increase as much as the time it was turned on.

Stairs rising by 1,000 times

Watt-hour is a small unit, so in practice words are attached in front of it. Kilo, mega, giga, and tera are prefixes attached in front of units to scale them up by 1,000 times; kilo is 1,000, mega is 1 million, giga is 1 billion, and tera is 1 trillion. That is the kilo of kilometer. A 2,000-watt electric kettle is 2 kilowatts, and if you leave this kettle turned on for 30 minutes, it uses 1 kilowatt-hour.

The unit of usage printed on a household electricity bill is the kilowatt-hour. A month for one house fits within this scale. Mexican households that replaced their refrigerators with efficient new models saw their monthly usage drop by an average of 11 kilowatt-hours, about 7 percent (DOI: 10.3386/w18044, working paper). Since 11 kilowatt-hours is around 7 percent, the amount these households used in a month before replacement was roughly 150 to 160 kilowatt-hours (rough calculation in this article). The average effect of a randomized program repeatedly encouraging households to save was also 9 kilowatt-hours per month, and the savings predicted by the model varied widely from 1 to 33 kilowatt-hours per house (DOI: 10.1093/ej/ueaf028). The change that such replacements or recommendations brought to a single home was in the single or double digits of kilowatt-hours per month.

Conversely, in the same Mexican replacement project, households that replaced their air conditioners used an average of 6 more kilowatt-hours per month, with the increase being greater in summer. The authors' explanation is that efficient machines cost less to run so they get used more, and the authors consider that this effect was large for air conditioners but not large for refrigerators (DOI: 10.3386/w18044). Since a watt-hour captures both power and time, even if you replace a machine, if you run it more, the amount used can increase.

If you climb the stairs all the way to the top, a country's year appears. Terawatt-hour is a unit for measuring the amount of electricity, and 1 terawatt-hour is 1 billion kilowatt-hours. Taking houses like those Mexican households that use about 150 kilowatt-hours a month, it is the amount that a little over 6 million households use for a month (calculation in this article). From kilo to tera is three steps, that is, 1 billion times.

Recounting the Used Amount — Average Power

Multiplication works in reverse, too. Dividing the used amount by time gives the intensity. Average power is the energy used over a period divided by that time, which is the intensity you would get if you used that much steadily the whole time. On a dashboard, this is the average speed. If you traveled 240 kilometers during a day, your average speed was 10 kilometers per hour. That is true even if you actually drove fast for a few hours and stood still the rest of the time.

A single person's lifestyle can also be counted this way. There is a calculation that if you add up all the energy for transportation, aviation, housing, food, goods, and water used by an American adult and convert it into an average flow per second, it comes to about 6,700 watts maintained throughout the day (DOI: 10.21203/rs.3.rs-9962169/v1, preprint not peer-reviewed). This is a baseline established to compare the resources required for human work and AI work, and it includes not only electricity but also energy spent on car fuel and making food. The watt is a unit measuring how fast energy flows, so it is used for energy outside of electricity, which is why fuel and food can be counted this way. If you use 6,700 watts continuously for 24 hours, it is about 160 kilowatt-hours per day, an amount combining energy that is not electricity (calculation from this article).

This division is the bridge connecting the two numbers in the news. One year is 8,760 hours. Dividing "how many terawatt-hours per year" by 8,760 hours gives how many gigawatts on average, and multiplying "how many gigawatts" by 8,760 hours gives the amount if it were used at that intensity all year round.

Two Numbers of a Power Plant

Solar power plants cannot generate electricity after the sun sets. Because the power generated in this way varies from time to time, power plants also need two numbers. The number attached to a power plant nameplate is on the speedometer side. Capacity is the maximum electric power the power plant is designed to continuously produce under specified conditions, written in watts. In terms of a car, it is the top speed. Generation is the amount of electricity actually produced by the power plant over a period, written in watt-hours. In terms of a car, it is the distance actually driven over a year.

No car runs at top speed all year round. Capacity factor is the ratio of the generation actually produced over a year divided by the amount that would have been produced if the capacity had run continuously all year round. Let us calculate using numbers made up for this article. If a power plant with a 100-megawatt capacity runs at maximum without stopping for 8,760 hours, it makes 876,000 megawatt-hours, or 876 gigawatt-hours. If this power plant actually makes 219 gigawatt-hours in a year, the capacity factor is 25 percent.

There are several reasons why the capacity factor falls short of 100 percent. Solar power stops when the sun sets, wind turbines turn less when the wind is weak, and every power plant has days off for maintenance. Even if the capacity is the same, the electricity made in a year can be different, so 「percent of capacity」 and 「percent of electricity」 are different things.

Two numbers on the consumer side — peak demand

At seven in the evening, a rice cooker, an air conditioner, and a washing machine run all at once in a home. Even if the kilowatt-hours used in a month are the same, a home that bunches consumption like this and a home that spreads it evenly through the day are separated by another single number. Peak demand is the electricity power at the time when the consumer used electricity the most intensely during a given period. In terms of cars, it is the speed at the moment it ran the fastest during a month. In actual metering, this "time" is captured as a very short interval and measured by the average intensity of that interval. The amount used and peak demand, these two numbers move separately.

Peak demand can be reduced by changing how it is used. An optimization model tailored to an actual home in Ontario, Canada automatically determines when and how much to run the refrigerator, dishwasher, washing machine, lighting, and heating and cooling. Depending on what goal was set, electricity and gas bills decreased by up to 20 percent, and peak demand decreased by up to 50 percent. Whether these two values came out together in the same operation is not in the abstract, and this is a simulation result that maintained the comfort desired by the homeowner (DOI: 10.1109/tsg.2012.2212032). Apart from this study, there is something you can know just by following the definition. If you spread out evenly what you used to bunch up, the peak demand goes down even if the amount used is the same.

The power grid is a system made by connecting power plants and places that use electricity with wires to run as one, and in the industry, it is also called the grid. The power grid must receive not only the amount used but also the most intense moment. Therefore, there are power generation facilities built separately to prepare for the time when it is used the most, and a policy question becomes whether the consumer reducing use or shifting time can replace the investment required for those facilities and power grid reinforcement (DOI: 10.2139/ssrn.2831868, a policy brief written regarding Western Europe).

The same direction comes out in power grid models as well. There is a study that assumed emergency situations where the capabilities of generators, lines, or both were cut in a test power grid model consisting of 30 nodes. In that auxiliary experiment, when the total amount used by data centers was kept the same while demand overlapped at one time, unsupplied energy increased by 34.4 percent more. Under which emergency condition and compared to what the increase is are not written in detail in the abstract. The authors conclude that data center demand clustered in time can increase the impact on the endurance of the power grid without increasing total usage (arXiv:2607.06958v1, preprint, model study). Even if the amount used is the same, if it is bunched up all at once, the burden placed on the power grid can grow.

The Two Numbers of a Battery

When you install rooftop solar and a battery together at home, you buy solar in kilowatts and batteries in kilowatt-hours. A study that calculated combinations using actual usage records from 4,190 households in Zurich, Switzerland, set the costs at 2,000 euros per kilowatt of solar and 1,000 euros per kilowatt-hour of battery (DOI: 10.1016/j.apenergy.2018.03.185). The kilowatt of solar is the intensity it can produce under standard, fixed sunlight conditions, which is its capacity.

Batteries also have two numbers. Battery capacity is the amount of energy that can be stored in a battery and drawn out later, written in watt-hours. Car-wise, it is the distance you can go on a full fuel tank. Battery output is the limit of power the battery can send out, written in watts. Car-wise, it is the speed you can reach. Depending on the product, the value you can sustain continuously and the value you can sustain briefly are sometimes listed separately.

Let us do the math using numbers made up for this article. If you draw power at 5 kilowatts from a battery with a 10-kilowatt-hour capacity, it lasts for 2 hours; if you draw it at 1 kilowatt, it lasts for 10 hours. This is a simple division ignoring losses. A battery with a continuous output of 5 kilowatts cannot sustain 7 kilowatts continuously, even if its capacity is large. Battery output refers to how strong, and battery capacity refers to how much in total, and dividing capacity by output roughly gives the time it can last at that intensity.

The place where checking these two numbers separately matters most is in batteries where output and energy ratings are independent of each other. Vanadium redox flow batteries are like that, and their size must be determined by considering the required output and the energy to be stored respectively (DOI: 10.1109/tste.2015.2404780). A 2008 study reported that in wind farms, it is difficult for a single storage technology to have both large output and large capacity, and proposed a combination where another storage device handles short, strong surges created by gusts while the battery handles long fluctuations (DOI: 10.1109/pesc.2008.4592198).

The size of a battery attached to the power grid is also written with two numbers: the energy rating and the output rating. If a battery supports peak demand instead, power outages decrease and new facility investments can be delayed. The same study adds that the larger the battery, the greater the blow to supply stability when it becomes unusable (DOI: 10.1109/access.2020.3024846). The peak demand from the previous section meets the battery's output here.

Returning to the Zurich study, under the most optimistic cost assumption (1,000 euros per kilowatt of solar and 250 euros per kilowatt-hour of battery), the best combination was an average of 4.4 kilowatts of solar and 9.6 kilowatt-hours of battery, and under that assumption almost all homes (99.9 percent) benefited from the battery. Calculated at the costs of that time without subsidies, only 0.1 percent of homes benefited by adding a battery (DOI: 10.1016/j.apenergy.2018.03.185). Kilowatts and kilowatt-hours appear side by side in a single sentence, with the former being intensity and the latter being amount.

Reading News Numbers — Data Centers

Now we read a data center article holding the dashboard. In the United States, the scale of data centers applying to connect to the power grid reaches hundreds of megawatts (DOI: 10.2139/ssrn.6614823, policy brief). A data center that research places as a representative case is sometimes a 20-megawatt one (DOI: 10.2139/ssrn.7336158, preprint), and when determining locations in a synthetic model mimicking the Texas power grid, 1-gigawatt and 2-gigawatt ones are placed and considered (arXiv:2605.14714v1, preprint). The number stating the application to connect and the number stating the scale are on the speedometer side, meaning intensity.

The first thing to do is multiplication. When hearing the phrase "1-gigawatt data center," multiply by 8,760 hours. That is 8,760 gigawatt-hours, or 8.76 terawatt-hours. 8.76 terawatt-hours is the amount when the intensity on the nameplate is fully packed throughout the entire year, and the amount actually used comes out by multiplying the average power by 8,760 hours. We also ask what the intensity on the nameplate includes. This is because the electricity used by the entire facility is the value for the computing equipment multiplied by a multiplier determined by the efficiency of cooling, power distribution, and power conversion (DOI: 10.63550/iceip.2026.20.71.004, conference paper). 1 gigawatt for servers alone and 1 gigawatt for the entire building are different numbers.

The second is division. The electricity used by Chinese data centers in 2018 was 161 terawatt-hours, which was 2.35 percent of the country's total consumption (DOI: 10.3389/fenrg.2026.1774740). Dividing by 8,760 hours, that is an amount with an average of about 18 gigawatts turned on all year round (calculation by this article). The 2030 projection from the same study ranges from 277 to 500 terawatt-hours depending on how much AI spreads and how much efficiency improves, spreading the upper and lower bounds by nearly twofold.

The third is looking at what time the number measured. An article that re-examined the global data center energy use estimates in 2020 saw that the growth in consumption slowed down thanks to efficiency improvements, and in the near term, policies could help continue those efficiency improvements (DOI: 10.1126/science.aba3758). Afterward, as AI spread rapidly, a diagnosis came out that the electricity used by AI data centers is increasing unprecedentedly (arXiv:2509.07218v6, preprint review). A 2025 preprint cited as background a projection that the electricity used by data centers globally will double by 2026 (arXiv:2511.08759v2). What it was doubled in comparison to is not in the abstract, and this article did not verify whether that projection was correct. The two diagnoses looked at different times. There was a time when efficiency improvements slowed the increase in consumption, and the diagnosis changed after AI spread. Therefore, when looking at data center numbers, we first check what time was measured, and if it is a future projection, what assumptions were placed.

Guesses about the share of data centers were far off. When asking 303 people gathered to mimic the U.S. demographic composition what percentage of current U.S. electricity data centers use, the median was 20 percent. What this survey paper (DOI: 10.2139/ssrn.7227338, preregistered survey and preprint) cited as a benchmark was the Lawrence Berkeley National Laboratory's 2024 estimate of 4.7 percent, and 92 percent of respondents guessed higher than that. This article did not directly view the original report by that laboratory. Why the guesses were off cannot be known from this result alone.

The fourth is checking what the share is a share of. The previous Chinese study notes that renewable energy capacity at the end of 2024 was 1.889 billion kilowatts, accounting for 56 percent of total capacity (DOI: 10.3389/fenrg.2026.1774740). This is the share of capacity. Since capacity utilization rates may differ for each facility, there is no guarantee that the share of renewables in the electricity generated that year is also 56 percent.

Lastly, even with the same terawatt-hours, when it is used remains. In the model study (preprint) seen earlier, assuming emergency situations, even with the same total amount, overlapping demand at a certain time increased the energy that could not be supplied. That burden is not visible with annual terawatt-hours alone. "How many gigawatts" is the speedometer, "how many terawatt-hours per year" is the odometer, and between them is how long and how fully packed it ran.

Where to look next

The number for the electricity a data center uses in a year varies depending on when it was counted and what assumptions were made, and the detailed circumstances of that go beyond the scope of this article.

• "How Much Electricity Does AI Use — The Real Reason the Numbers Vary Across Sources" — Follows how the electricity used by data centers is counted and where those numbers diverge.
• "Is a Data Center an Inflexible Load — What Can Be Moved and What Cannot" — The task of lowering peak demand by shifting when electricity is used while keeping the total amount used the same.
• "Plugging In Is Harder Than Building Power Plants — What the Interconnection Queue Actually Determines" — The circumstances under which connection requests that came in by the megawatt stand in line.
• "Air Is No Longer Enough — Where Does the Heat Generated by Computation Go" — The cooling electricity that fills the gap between the servers' share and the building's entire share.

Source Papers

arXiv and Research Square preprints are pre-peer-review or independent of peer review status. Working papers, policy briefs, and conference papers are indicated in the text where applicable.

• "Cash for Coolers" (2012, working paper) — DOI: 10.3386/w18044 — In a Mexican home appliance replacement program, replacing a refrigerator reduced consumption by an average of 11 kWh per month (roughly 7%), while replacing an air conditioner increased it by 6 kWh per month (larger in summer). This is explained by the framework that more efficient machines are cheaper to run and thus used more, and this effect is considered important for air conditioners and unimportant for refrigerators.
• "Using Machine Learning to Target Treatment: The Case of Household Energy Use" (2025) — DOI: 10.1093/ej/ueaf028 — The average effect of a household energy-saving encouragement program is a reduction of 9 kWh per month, and the predicted reduction range is roughly 1 to 33 kWh per household.
• "To AI or not to AI? Per-task energy and water demands of AI inference are orders of magnitude lower than human labor across all scenarios" (2026, preprint) — DOI: 10.21203/rs.3.rs-9962169/v1 — Summing the lifecycle energy by sector for one US adult yields roughly 6,700 W of continuous power (the paper's main conclusion is comparing resource use per task between AI inference and human labor).
• "Optimal Operation of Residential Energy Hubs in Smart Grids" (2012) — DOI: 10.1109/tsg.2012.2212032 — Applying a residential appliance operation optimization model to actual homes in Ontario, electricity and gas bills were reduced by up to 20% and peak demand by up to 50% while maintaining comfort depending on the objectives (simulation, simultaneity of the two values is not in the abstract).
• "Demand-Side Flexibility for Energy Transitions: Policy Recommendations for Developing Demand Response" (2016, policy brief) — DOI: 10.2139/ssrn.2831868 — It recommends evaluating whether demand response can be an alternative to conventional peak demand equipment and grid reinforcement investments from a risk governance perspective.
• "Evaluating Grid Resilience in the Era of Ever-Increasing Data Centers" (2026, preprint) — arXiv:2607.06958v1 — In an emergency analysis (generator, line, and tie deratings) of an IEEE 30-bus model, a supplementary experiment showed that with energy matched and overlapping data center demand, unserved energy increased by 34.4%, concluding that resilience impacts can grow even without an increase in total usage.
• "Optimal Sizing of a Vanadium Redox Battery System for Microgrid Systems" (2015) — DOI: 10.1109/tste.2015.2404780 — Because the power and energy ratings of a vanadium redox battery are independent, sizing must consider the required power and energy capacity together.
• "A power electronic interface for a battery supercapacitor hybrid energy storage system for wind applications" (2008) — DOI: 10.1109/pesc.2008.4592198 — Because single-technology storage devices struggle to provide both large power and large energy capacity together, a scheme combining flow batteries and supercapacitors for wind power applications is proposed.
• "Probabilistic Peak Demand Matching by Battery Energy Storage Alongside Dynamic Thermal Ratings and Demand Response for Enhanced Network Reliability" (2020) — DOI: 10.1109/access.2020.3024846 — The battery size has two ratings: energy and power; if the battery supports peak demand, supply interruptions decrease and new facility investments can be postponed, but larger batteries cause greater harm to reliability when they are unavailable.
• "Economic assessment of photovoltaic battery systems based on household load profiles" (2018) — DOI: 10.1016/j.apenergy.2018.03.185 — Using data from 4,190 households in Zurich, at the costs at the time (solar 2,000 €/kWp, battery 1,000 €/kWh), 0.1% of homes profit from a battery, while under the most optimistic costs 99.9% profit, with an average optimum of 4.4 kWp solar and 9.6 kWh battery.
• "The Missing Half of the Queue: Why FERC Should Require Load-Side Interconnection Queue Disclosure" (2026, policy brief) — DOI: 10.2139/ssrn.6614823 — Unlike the generation side, queues for large-scale demand applying for connection at the scale of hundreds of megawatts are not publicly disclosed.
• "Governing the Energy Footprint of Artificial Intelligence for Sustainable Development" (2026, preprint) — DOI: 10.2139/ssrn.7336158 — Analyzing a representative 20-megawatt data center, it finds that efficiency alone cannot guarantee sustainability amid growing AI demand.
• "Flexibility-Aware Framework for Efficient Planner-Initiated Siting of Data Center" (2026, preprint) — arXiv:2605.14714v1 — Evaluating the siting of 1 GW and 2 GW data centers under flexible operation conditions in a 2,000-bus synthetic model patterned after Texas.
• "AI Energy Elasticity and Data Center Energy Consumption Growth Scenarios to 2030" (2026, conference paper) — DOI: 10.63550/iceip.2026.20.71.004 — Total data center electricity consumption is factored into the IT share multiplied by an infrastructure multiplier determined by cooling, power distribution, and power conversion efficiencies.
• "Analysis of China's power development and the impact of data center energy consumption" (2026) — DOI: 10.3389/fenrg.2026.1774740 — China's data center electricity consumption was 161 TWh in 2018 (2.35% of the total) with a 2030 projection of 277 to 500 TWh, and renewable energy capacity at the end of 2024 was 1.889 billion kW (56%).
• "Recalibrating global data center energy-use estimates" (2020) — DOI: 10.1126/science.aba3758 — Efficiency improvements have slowed the growth of data center energy use, and policies can help sustain those efficiency gains in the near term.
• "Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects" (2025, preprint review) — arXiv:2509.07218v6 — The spread of AI is driving an unprecedented increase in electricity consumed by AI data centers, and it outlines the challenges that this load poses to the power grid over a three-hour scale.
• "Grid Operational Benefit Analysis of Data Center Spatial Flexibility: Congestion Relief, Renewable Energy Curtailment Reduction, and Cost Saving" (2025, preprint) — arXiv:2511.08759v2 — Against the backdrop of projections that global data center electricity consumption will double by 2026 (base year not specified), it models and evaluates the grid benefits of spatial flexibility.
• "Grid Expectations: Americans' Overestimation of US Data Center Electricity Use" (2026, preprint) — DOI: 10.2139/ssrn.7227338 — In a pre-registered survey (N=303), the median estimate of data center share of US electricity use was 20%, higher than LBNL's 2024 estimate of 4.7% used as a benchmark by the paper, and 92% of respondents overestimated the benchmark.