How Data Centers Draw Power

Five public documents describe how data centers draw power. Three points come first. U.S. data centers may use about half of their grid connection, by one estimate. An AI training cluster's draw can swing in under a second. And one kind of grid sensor may misread a fast swing. Each point comes from a document named below, with its limits.

On the software side, a training job is work you can schedule. On the power side, it is demand you have to meet, second by second. This question sits where the two meet. The size of a connection does not tell you either.

Connection Size and Peak Draw

Start with the connection. People often describe a data center by its power capacity. Berkeley Lab's 2025 update on U.S. data center energy use says the phrase can mean two things. One is the nameplate capacity: the IT hardware load the facility is designed to serve. The other is the connection asked of the grid. That is the most power the data center requests from the utility or grid operator, usually through a formal request. It covers the whole site: computers, cooling and other systems, plus design margins.

How much of that connection do U.S. data centers use today? About 50%, the lab estimates. It is an estimate, and the lab says current use "is not well documented." It gives two reasons for the share: high redundancy requirements and maintenance needs. It adds that the industry is making a concerted effort to raise the share. The effort covers new ideas in power system design, energy balancing and demand management. And it says the new generation needed for new loads is "not inherently equal" to the connection requested. When the lab turns its energy forecast into connection size, it assumes an average utilization rate of 50%.

Think of a house with two water heaters, one kept as a spare. The pipes are sized for both. Most days only one is working.

A national laboratory in Golden, Colorado, looked at peak power. It compared the peak power of AI computing with the power the computers were rated for. That rating leaves out cooling. The laboratory's researchers measured the power of AI jobs on the laboratory's own computing center. They took a reading every tenth of a second. Then they scaled those readings up to two simulated data centers.

The first was a 10 MW colocation data center running a mix of AI training and fine-tuning jobs. That size did not count cooling and other auxiliary loads. In most runs, every server node was busy at once at some point. Even so, power peaked at about 7.30 MW. That is about 73% of the center's rated power design.

The second was a 1 MW center answering users' questions with a large language model. Its peak reached 80% of its rated power.

The paper names some limits itself. The measurements came from one specific hardware setup. Its patterns of use were adapted from outside datasets, and may not fully reflect large commercial or enterprise sites. The results are "strictly tied" to the workload power profiles it assumed. The simulations left out cooling, power distribution losses and other loads that are not IT. The paper says those might be less than 10% of the total load at large AI data centers of the latest design. It also says the results raise questions about spare room across the whole facility. One is whether that room should count when the infrastructure is sized.

A grid connection, by Berkeley Lab's definition, covers cooling and design margins. The rating in Golden leaves cooling out. So I would not read one as the other: the size of the connection and the peak draw of the computers can be two different numbers.

When the load asks for one number and draws another, you can size for the request and carry capacity that goes unused. Or you can size for a lower draw and add machines as the campus grows. Neither document says which is right for a given campus.

The Swing in a Training Cluster's Draw

Ordinary data centers usually draw power fairly smoothly. So says a report from the Pacific Northwest National Laboratory (PNNL), dated April 2026. Their work is many separate tasks. The ups and downs of those tasks tend to average out over time.

An AI training facility works differently. Its graphics processors work in parallel, in tightly synchronized training cycles. Each cycle has two phases. In the first, the processors compute hard and draw a lot of power. In the second, they mostly move data and coordinate with each other, and they draw less. Repeat that over and over, and the report says you get a "sustained oscillatory load profile." In plain words, a draw that keeps swinging. Because the work is coordinated, power can rise and fall at the same moment across many servers.

How fast? NERC, the North American Electric Reliability Corporation, published a white paper in July 2025. It says AI training in large clusters generally has rapid swings. They come during training and while the job saves its progress, a step called a checkpoint. The switch between the two "may happen in under one second." NERC also says AI data centers can have load profiles that are "periodic, repetitive, and sustained."

How a Grid Sensor Can Misread a Fast Swing

One tool for watching the grid is the phasor measurement unit, or PMU. PNNL's report shows how one can misread a fast swing. Its example is a large crypto-mining data center, not an AI cluster.

In October 2024, ERCOT, the Electric Reliability Council of Texas, saw that load's power swing as it ramped. A PMU at the point of connection reported 30 times a second. It showed a swing of about 7.5 Hz. Hz, or hertz, means cycles per second. The PMU showed it as 25 MW from peak to peak.

A fault recorder at the same location took more samples each second. It showed the actual swing. It was about 23 Hz. It measured 50 MW from peak to peak.

The report explains the gap. The PMU's limited reporting rate made it show the wrong frequency, which the report calls aliasing. It also shrank the measured size a great deal. Reporting faster may not be enough. Even at a rate high enough to avoid aliasing, there is a second limit. The PMU's built-in filtering "may still attenuate" the faster parts of a swing.

You have seen this in old films. A wagon wheel can seem to turn slowly, or even backward, because the camera takes too few pictures a second. The PMU did the same to the swing.

The report concludes that PMUs are generally not well suited to watching swings much faster than about 5 Hz. They remain highly effective for slow swings, it says. They may misrepresent fast ones.

Training a Model and Using It Draw Power Differently

NERC says the power demand is different for training an AI model than for using one. Using a trained model on new input is called inference. It is common, NERC says, for a data center to be built for one or the other. And the rapid ramps of training and checkpoints are not seen in current inference methods.

Berkeley Lab describes inference as largely user-facing. Systems must be sized for peak use, when many users query the models at once. Outside those peaks, servers may run at relatively low utilization.

The lab calls how hard a server works its utilization. That is its average effort over a year, as a share of its maximum. Its 2024 report assumed an average utilization of 40% for AI servers doing inference. The lab says recent feedback from industry experts indicates it is "likely much lower." The update's main scenario, its Reference Case, assumes inference utilization reaching 20% by 2030. For training servers it assumes 80%. The lab adds that there is still "a lack of verifiable data" to confirm typical inference utilization.

In the simulations from Golden, the inference center's peak-to-average ratio was lower on average than the colocation center's. That ratio compares the highest draw with the average draw. The paper says this could suggest that inference centers disturb the grid less. It says the harm from full servers is "arguably more detrimental" to an inference center than to a colocation center, because inference centers provide real-time and often critical services. So, it says, inference centers might pose less risk to grid stability than colocation centers, at similar utilization levels. But they are more likely to run at lower average utilization, to keep their service from degrading.

The software decides when a cluster saves its progress: how often it saves a checkpoint, and how it restarts from one. Those are settings in the code. NERC says the switch between training and saving a checkpoint may happen in under one second.

What an Idle Server Draws

A server that is waiting still uses power. Berkeley Lab assumes an idle AI server draws 20% of its rated power. That figure rests on measurements reported by another national laboratory.

The lab also reports feedback from industry experts. In commercial AI work, it says, the feedback indicates that idle power may be a much higher share of rated power than the available measurements show. Idle power is shaped by a server's power settings. These may be changed from the factory settings to keep servers more active, so they respond faster.

A hardware architect at a major chip maker confirmed the lab's 20% in a follow-up conversation. But it was not confirmed whether that share may change once end users install and manage the hardware. Because of that uncertainty, the lab tests higher idle power in one of its scenarios. That scenario assumes 30%.

Why Fast Swings Matter, and a Draft Limit

Why does a fast swing matter to the grid? ERCOT gave its board an update in September 2026. It says large loads, particularly those tied to AI training at data centers, can produce rapid swings in power use. Those swings can disturb the shafts of synchronous generators, it says, and may lead to equipment fatigue or damage.

ERCOT had a draft proposal on this. It "will recommend" a limit of 10 MW per 5 seconds on how much a large load's power may vary. As of that update, it was a draft.

PNNL's report looks at what it calls the sub-synchronous range. For a 60 Hz system, it says that range, in everyday use, is generally between about 5 Hz and 59 Hz. At the higher frequencies in that range, a load's swings can interact with the torsional modes of nearby synchronous generators. Torsional means twisting: these modes are part of how the shaft of a turbine and its generator moves. If the swings keep exciting those modes, twisting force can build up in the shafts. The report says that could add mechanical stress and long-term fatigue.

In the past, PNNL says, these torsional interactions were short-lived. They faded once fixes were put in place. But swings from tightly synchronized computing in large AI clusters "may persist continuously" while the facility runs normally.

A generator feels a load swing as a change in torque on its shaft. Within a second, a swing can be met by the machine's controls, by stored energy or by the grid. PNNL says swings at the higher frequencies of that range can interact with the torsional modes of nearby generators. If they keep exciting those modes, twisting force can build up in the shaft. That is one reason I would want to know which of the three is doing the work.

NERC says the bulk power system, the main high-voltage grid, "was not designed to operate with large and persistent subsynchronous oscillations." It also says that, broadly speaking, the reliability risks of forced oscillations, swings that a load itself drives, "are not always well defined."

Interruptions, Checkpoints and Ride-Through

What happens when the power to an AI cluster is cut? NERC's paper cites data center industry experts on this. Power systems for AI data centers, they note, commonly leave out an uninterruptible power supply, or UPS, for the computers. Instead, checkpoints are saved at regular intervals. If the power to the computers is interrupted, the checkpoint may be restored after the interruption ends.

ERCOT's board update lists a risk it calls ride-through. NERC says ride-through is defined mainly by how long a load stays connected during a voltage or frequency disturbance. The risk ERCOT describes is sensitive computing equipment tripping offline during normal grid events. That, it says, leads to imbalance and instability on the grid. Its fix is ride-through requirements for large computing loads. The Public Utility Commission of Texas (PUCT) approved them in July, effective on August 1.

The cluster may restart from its last checkpoint. The grid has none. That is where the two trades have to meet.

Gaps in the Data

Berkeley Lab says future research should study how a data center's demand changes over time. The aim is to understand better how the grid, or power made on site, can serve these loads.

The Golden laboratory's paper notes that existing power data is "largely proprietary." It is also reported at different levels of detail. That makes it hard to estimate a whole facility's use and to plan the infrastructure. The paper's own measurements are public.

So before a single machine is ordered, I would ask for three answers in writing. How much of its connection will the campus really use? How fast can its draw swing? And how was that measured?

Who has measured the data center in front of you, and how often did their instruments look?