Utility supply remains the fastest path to power for most AI campuses, but the queue is where that path stalls. The nature of the load is now the limiting factor. 

Utility planners have stopped treating AI facilities as ordinary commercial loads. NERC's Large Loads Working Group classifies them as emerging large loads¹ because of how they behave on the system, not how many megawatts they draw. That behavior stopped being theoretical on July 10, 2024, when a lightning arrestor failure on a 230 kV line in the Eastern Interconnection triggered a sequence of faults lasting 42 to 66 milliseconds each. Voltage sagged to between 0.25 and 0.40 per unit during those faults. Data centers in the area, running conventional UPS systems, read the disturbances as a threat and disconnected. Roughly 1,500 MW of load dropped simultaneously, and 1,260 MW of it stayed off the grid for hours². The line was already recovering. The load didn't know that, because nothing in its architecture was designed to.

Voltage at the point of interconnection during the July 10th, 2024, Eastern Interconnection event. Six faults in under two minutes, each 42 to 66 milliseconds, triggered simultaneous data center load loss.²
 

That is the difference between a stable asset and a disturbance risk, and it is now the first question in any interconnection review. Available megawatts are the second question. A facility that fails the first inherits delays and reinforcement costs that were never in the pro forma.

Why the Decision Belongs in the Base Design
An MV-UPS cannot be a late-stage procurement choice. Stability has to be designed into the point of interconnection (POI) interface, because the interface is what the utility studies. This is underscored by Open Compute Project’s recent specification, authored by hyperscalers, which recommends a proactive design approach to power stability³. 

An interconnection review asks two separate questions. The capacity question is whether there is thermal and voltage headroom at the point of interconnection. If there isn't, the answer is a substation or line build on the utility's own timeline. The behavior question is how the load interacts with grid strength at that point, and it is assessed separately: NERC's guideline directs planners to study short-circuit ratio conditions specifically for harmonic and oscillation risk, and to hold large loads to the limits in IEEE 519-2022⁴. A facility that steps hard and trips on a sag like the one in the July 2024 event fails the behavior test even when it clears the capacity test, and the resulting mitigation, whether ramp-rate limits or a mandated curtailment agreement, lands on the developer.

Decoupling the load from the source changes what the planner sees. The MV-UPS sits at medium voltage between the utility and the compute halls, operating grid-forming on the load side while precisely regulating what the grid side draws. Integrated batteries absorb the sub-second load steps and hold through voltage and frequency events that would trip a conventional UPS, including sags in the range the July 2024 incident recorded⁵. The utility sees a stable, predictable interface instead of the raw behavior of the GPU hardware behind it⁶. This is the same decoupling principle validated on electric arc furnace loads, which present a workload profile similar to the rapid fluctuations of AI training⁶, applied here to a load with a far higher cost of interruption.

What the Architecture Delivers
The clearest gain is fewer study triggers: when ramp rate and harmonics are controlled at the load, the conditions that push a project into extended dynamic studies and mitigation mandates are addressed before the application is filed.

The design also holds across supply scenarios. With the stability block between source and load, the same medium-voltage architecture works whether the site is grid-connected on day one, bridged with on-site generation while the interconnection completes, or islanded later, because meter position and generation source don't change what the interface itself has to do.

The commercial case follows directly from the reliability case: an interface that rides through grid-side voltage events protects GPU cluster uptime in a way that more generators and longer battery runtime cannot, because neither changes how the load looks to the grid.

The Next Design Requirement
The industry has long bought reliability as redundancy. More generators and longer runtime still belong in the design. They don't change how a facility responds when a fault like the July 2024 event hits it.

The question for developers and EPCs is no longer whether the grid can power the load. It is whether the facility is engineered to ride through a disturbance instead of adding to it. 

Stability designed into the interface from day one is what makes that answer yes before the interconnection application is filed.

Ready to see how MV-UPS can move your next site forward? Answer this four-question check.

References

1. NERC Large Loads Working Group, White Paper: Characteristics and Risks of Emerging Large Loads, July 2025.
2. NERC, Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions, published January 8, 2025 (documenting the July 10, 2024 Eastern Interconnection event).
3. Open Compute Project Data Center Facility Project, Interface Specification Between Hyperscalers and Utility Grid for Disturbance Performance (v1.0.0), September 2026.
4. NERC Large Loads Working Group, Reliability Guideline: Risk Mitigation for Emerging Large Loads, May 2026 (Final); IEEE Standard 519-2022.
5. GE Vernova, Power Stability Fact Sheet: MV-UPS for Data Centers, Rev1, August 10, 2026.
6.  GE Vernova, When the Load Becomes the Disturbance: Lessons from Steel for AI Data Centers, September 29, 2026