Published November 6, 2026
By NZero
A large data center can draw 200 to 500 megawatts from the grid, roughly the demand of 200,000 to 500,000 households. It can also be built in 18 to 24 months, while new transmission and grid upgrades typically take four to ten years. That timing gap is now one of the defining challenges for US power systems. In Texas, grid operator ERCOT faces close to 400 gigawatts of data center connection requests, almost five times the state’s record peak demand of 85.5 gigawatts.
Flexibility, meaning a data center’s ability to reduce what it draws from the grid when the system is under stress, is increasingly proposed as part of the answer. A new analysis by energy think tank Ember examines two ways to deliver it. The first, load shifting, moves computing work to other times or places. The second switches a facility to its own onsite generation or storage during stress events. For US operators, the findings point to a practical strategy: start with the workloads that can move, then add onsite power to go further.
Data center capacity is expected to at least double in most regions between 2025 and 2030. But grids are designed around their single highest-demand hour, not the average one. What counts for reliability is how much data center load sits on the system at that peak.
In Texas, Ember estimates data centers account for about 6% of peak demand, a share expected to stay roughly steady through 2030 even as the market grows. Actual draw also differs from nameplate capacity. US data centers run at an average utilization of around 48%, so their contribution at peak depends on how hard their equipment is working at that moment.
Peak contribution also carries a cost. Grid operators secure enough firm capacity to cover the highest-demand hours, and those costs are shared across customers. In the capacity auction run by PJM, the largest regional grid operator in the US, data center load was responsible for $6.3 billion of the $16.4 billion in capacity charges for 2028–2029. Reducing data center demand during the hours that set these requirements can therefore benefit the wider system, not just individual facilities.
Load shifting works best where computing tasks can be paused or rescheduled without affecting users. AI training is the clearest example, with training runs typically able to flex 25% to 55% of their power demand. In demonstrations under the Electric Power Research Institute’s DCFlex initiative, Emerald AI cut power consumption by up to 40% within a minute and held the reduction for several hours without compromising its highest-priority jobs.
Flexibility varies by workload type:
Based on a global estimate of the 2030 workload mix, traditional and inference workloads make up about 87% of capacity. Adjusted for utilization, around 3% of data center capacity is available for load shifting. In Texas, California and New York, that translates into peak demand reductions of roughly 0.1% to 0.6%. Load shifting is a useful first layer, and it leaves room for a second tool to deliver more.
Onsite generation, sometimes called bring-your-own-capacity, keeps every workload running while the facility switches its power source from the grid to its own generation or storage. Because operations do not change, it works for traditional, inference and training workloads alike.
Ember’s Texas case study shows the scale of the difference. Two planned facilities, Stargate Abilene and Comanche Circle, together represent about 5.6 gigawatts of new load due online by 2030, drawing around 2.7 gigawatts on average. During a grid stress event:
The approach is already gaining ground. Developers have announced 40 gigawatts of onsite generation for Texas data centers. Texas Senate Bill 6 requires new large loads of 75 megawatts or more that connect after 31 December 2025 and have onsite backup generation to be curtailable during grid emergencies.
Onsite assets may be needed only briefly. A modeling study of six large US sites with flexible grid connections found grid power remained available in more than 99% of hours, with onsite resources needed for only 40 to 70 hours a year.
The outcome depends on what is built. Ember notes that gas-based onsite generation risks fossil fuel lock-in, while clean generation and storage avoid it. Grid operators also need to coordinate switching, since many facilities dropping off the grid at once could affect system stability.
Load shifting and onsite generation address different parts of the same challenge. Load shifting makes better use of existing grid and generation assets, while onsite generation delivers larger relief during the hours that matter most. The right mix depends on each facility’s workloads, location and connection terms.
For US data center operators, practical steps include:
Each of these steps depends on accurate, timely energy data. Operators that understand their load profile can show grid planners when and how much demand can be reduced, turning flexibility from a promise into a resource the grid can count on. That data foundation is where energy management platforms such as NZero can support data center teams as demand keeps growing.
