There is a broad consensus among analysts that the rapid rise of data centers and the increased use of artificial intelligence (AI) will increase electricity demand. But by how much and how soon? Moreover, there are some experts who believe that data centers and AI electricity demand, like all other demand, has some inherent flexibility – which at a price and with added intelligence – can be managed to avoid overloading the limited capacity of the existing grid.Â
The latest report on the uptake of AI from theInternational Energy Agency (IEA) reckons that data centers accounted for roughly 1.5% of the world’s electricity consumption in 2024 or around 415 TWhs. By comparison UK consumed 263 TWhs in 2023 – i.e., data centers are electricity intensive, and all indications are that their consumption will continue to rise. The IEA believes that the number will more than double by 2030 to around 830 TWhs.
The IEA is mindful of the crucial role that renewables and natural gas fired plants will play in meeting this demand while pointing to the many challenges posed by grid connection queues. The agency is especially concerned about the risk that meeting data center load growth could adversely impact other goals, such as electrification, manufacturing growth, or affordability. The IEA is not alone.
Meeting the increased AI demand, however, is not different than meeting the increased demand from electrification of, say, road transport, which will most likely have an even bigger impact on demand. But just as the time, location and speed of charging electric vehicles (EVs) can be managed to avoid overloading the grid infrastructure at peak times, the same might apply to data centers, or so it is hoped.Â
In a post on 14 April 2025, Severin Borenstein, a professor at the University of California Berkeley asked “Can Data Centers Flex Their Power Demand? It is a timely and critically important question.Â
Borenstein’s starting point is to ask, “do data centers really need to strain the grid?”Â
For the “vast majority of hours” any network has “plenty of spare capacity to serve these new loads. But in a small number of hours each year, the extra demand could lead to price spikes and even shortages.Â
Borenstein correctly points out that the focus must be on the strain of new loads, data centers or otherwise, on the grid during “a very small number of hours” where the network may be overloaded. The rest of the time, there is plenty of spare capacity to serve the load.
This leads to the next question:Â
“Couldn’t data centers ramp down during those hours (when) demand pushes the limits of supply?”
Borenstein refers to research that “suggests that a bit of flexibility at peak times – ratcheting down demand during less than 1% of annual hours – could drastically reduce the grid constraints that are causing planners to slow interconnections for new data centers.”
The key issue is to test “the common narrative that data centers are so expensive that they must run at 100% of capacity once they are built” to justify the costs.
To answer this question Borenstein says we must ask data center operators “how much is it worth to run them up to 100% in each hour? Does that value exceed the full cost of incremental supply for the hour?”
“My intuition suggests it probably doesn’t in hours when the cost includes a spiking wholesale electricity price in a strained grid.”
He points out that data centers have two broad categories of activities:Â
§ Real-time response to data queries and information retrieval requests; and
§ Background tasks such as training AI models and processing tasks that are not time sensitive. Â
“Demand for the real-time activities fluctuates throughout the hours of the year, so those tasks can’t keep data centers running at or near 100% of capacity in every hour. The operators fill in the troughs in real-time usage with the background tasks. In hours with lots of real-time tasks, not many of the background tasks get done, and the opposite happens when real-time demand is low.”
Borenstein is not the first to point out that “electricity is a significant share of the cost of operating a data center, 40-70%” depending on the specifics including how much they pay for electricity, the type of supply contract they have and so on.Â
“Even if it is only 50% of operating costs on average, during the small number of constrained hours that portion of costs can easily rise by 10x to more than 20x, causing overall operating costs to increase from 5 to more than 10 times. It seems quite likely that a data center facing that additional cost would look for ways to ramp down or shut off completely for a few hours.”
“Ramping down doesn’t mean that real-time computing tasks get dropped or even postponed by more than a fraction of a second,” he notes. In some cases, the heavy computing load may be shifted to another facility in a different location where electricity is plentiful and cheap.
While moving tasks around creates lags or latency, they are typically measured in milliseconds, which may be easily tolerated by most customers.
“So, if straining the grid is so costly and the solutions are so easy, why aren’t we seeing data centers everywhere modifying their loads during the highest demand hours of the year?”
Borenstein’s own view is that it is less of a technical problem and more of an incentive problem. The data centers may not be adjusting loads not because they cannot but because they don’t have sufficient incentives to do so.
The answer, therefore, seems to be time-varying electricity prices that reflect real-time variations in the wholesale price and grid congestion. When and if customers have strong financial incentive to reduce consumption during the small number of hours when wholesale costs spike, they may respond. Moreover, the operators of large data centers have “people whose full-time job it is to manage electricity costs.”
“… it is perfectly possible to design a menu of prices that reflects the value of demand flexibility and appropriately include a hedge premium for customers who choose to have the utility buffer price fluctuations for them. Failing to do so raises costs, slows interconnection of variable loads, and undermines reliability.”
There are other promising developments including the fact that “many data centers are already managing the timing and location of computing tasks to minimize their carbon footprint. Some are also co-locating renewable generation with the facilities. So, they are already in the business of managing real-time energy supply and demand fluctuations. Why would we have utilities distort those decisions by offering prices that don’t reflect real-time costs?”
There may be other reasons such as the utility contracts with data centers which don’t “specify service level obligations” or other “critical features” such as “telling customers that there might be some service degradation – e.g., latency increase – when the grid in their area is strained.”
Borenstein ends his post with the following observation,Â
“These are early days for the data center business. It’s clear that pricing and service requirements have not caught up with the imperatives of large loads that are rapidly growing. Developing service contracts between utilities and the data centers, and between the data centers and their business customers, will be central to serving this exploding new demand in a cost-effective way without undermining system reliability.”
Given the sheer size and expected growth of data center demand, exploring their demand flexibility is worth considering. As with electric vehicles, if properly managed, their impact on the grid infrastructure can be minimized.
This article appeared in the June 2025 issue of EEnergy Informer, a monthly newsletter edited by Fereidoon Sioshansi who may be reached at [email protected]
Article at:
https://energyathaas.wordpress.com/2025/04/14/can-data-centers-flex-their-power-demand/