A report from consultant Wood Mackenzie warned that on-site energy facilities–often microgrids–for AI data centers need utility-scale equipment to buffer the impact of AI ramping.

Lithium-ion batteries can be used to address this, but they could potentially run through their useful lifespan quickly, the report said.

Another option is using flywheel energy storage to meet bursts of demand–in AI data centers, cannabis growing facilities and other industrial operations, said Brian Battle, COO at Beacon Energy, which offers flywheel storage and is fielding numerous requests from AI data centers, he said.

Flywheel energy storage involves accelerating a cylindrical rotor assembly to store electrical energy as rotational kinetic energy. Flywheel systems absorb excess power and send it back to the grid or microgrid when needed, according to Beacon Power.

Flywheels can help stabilize the grid

Giant flywheels are being deployed in the U.K. to help stabilize the grid.

The technology is most helpful for projects that call for high power, numerous charge-and-discharge cycles, and a long cycle life, according to Sandia National Laboratories. In one example from Sandia, a facility that stores energy in flywheels can source or sink 20 MW for 15 minutes.

In a Beacon Power project, an existing 5.5-MW off-grid microgrid serving a 150,000-square-foot cannabis growing facility in central Illinois, the facility needed to meet a narrow band of acceptable temperatures, 69.8 degrees to 70.2 degrees. As a result, the off-grid, gas-fired microgrid kept cycling on and off, Battle said.

Once the temperature rises to, say, 72 degrees, the cooling might kick in until it drops it to 68 degrees, and then it'll turn off.  “In this case, they're trying to keep temperatures in a very narrow range, between about 69.8 and 70.2 degrees,” Battle said.

Large demand swings at a cannabis growing facility

The microgrid originally had four natural gas engines providing power to the facility. It was subject to demand swings like those seen in AI data centers–sometimes power surges lasting less than a second– and experienced frequency and voltage instability. That resulted in power brownouts and blackouts, Battle said.

In  2024, Beacon Power installed two flywheels that store 30 kWh and last for about 5 minutes at full load, he said. This was an add-on to the existing system and didn’t involve replacing any components.

The payback on the flywheels was less than 2 years, Battle claimed. That’s because the flywheels  cut the gas engines’ runtime by 15% to 20%. The cannabis facility also saves on engine maintenance and fuel.

Flywheels can be less expensive in high-demand swing scenarios

In general, flywheels are less expensive than lithium-ion batteries in high-load-swing situations because they don’t need to be replaced for about 30 years, while batteries might need to be replaced once or so a year because the constant cycling degrades the batteries, Battle said.

AI data centers experience similar but bigger spikes in demand as they receive AI requests.  A data center may respond to numerous requests at once, and the load cycles quickly. “They've got thousands of processors in the data center. So you can see that this is going on all the time,” Battle said.

Other systems can meet AI data-center demand swings.

On-site power providers like VoltaGrid say they can meet these demand swings utilizing battery storage. In 2025, VoltaGrid announced it would deploy 2,300 MW of infrastructure to support Oracle Cloud AI data centers. When the deal was announced, Oracle noted that AI data centers are highly variable and often create demand swings. Oracle’s collaboration with VoltaGrid will lead to innovations that dampen those swings, Oracle said.

It makes sense in some instances to combine lithium-ion storage with flywheels.

“If you wanted to store solar that you generate during the day, and you want to use it to power a plant at night, that's going to be more the workload of a battery,” Battle explained.

“If you really want to store energy to use all night, we would use a battery in our solution,” he said.  “But we might do the hybrid of a battery and a flywheel together.” Flywheels paired with batteries can serve as a “shock absorber,” making an energy system more stable.

Beacon has installed flywheels for utilities, absorbing excess power and injecting it when needed to maintain the correct voltage and frequency.

The AI data-center demand swings are relatively new, so data center operators are just beginning to look at flywheels, said Rex Roehl, project development director at Beacon Power.

Companies that offer flywheels–Torus, Vycon Energy and Active Power, for example–now see a market in AI data centers.

Modeling uncovers demand bursts

Data center operators and others modeling AI demand on a grid or microgrid are now seeing the size and frequency of AI data center demand swings. That’s how the flywheel became more attractive to them, said Roehl. “They're doing their own modeling and they're seeing these big swings that they have to cover for a matter of seconds. That's something that they've never had to think about before because they always just relied on the big grid to do that,” Roehl said.

It’s possible that regulators and utilities will begin creating regulations requiring data center operators to address AI demand swings, Battle said.

Ben Hertz-Shargel, global head of grid transformation and large loads at Wood Mackenzie, said fast ramp times in AI data centers can wreak havoc.

“There is the risk of AI ramping up where the workloads can go up and down in tens of megawatts within  milliseconds, which is an incredibly fast ramp rate that reciprocating engines and turbines aren't able to deal with directly,” he said. “So you need to have equipment that buffers the impact of those hammer blows of load steps.”