Artificial intelligence's energy problem does not end with the search for new generation capacity. Turbines, photovoltaics, transmission lines, and grid connections are essential, but the race for large compute campuses is bringing a second front to light: how those campuses absorb, modulate, and interrupt their electricity consumption.
Virginia, and specifically the Ashburn area, is the clearest example. On July 22, 2026, a transmission line fault knocked out more than 3 gigawatts of load in seconds across the world's largest concentration of data centers. This was not an issue of insufficient power generation: the critical point was the simultaneous response of numerous facilities to a grid anomaly.
A previous incident in 2024 had already revealed the same vulnerability. The failure of a single surge arrester led to the disconnection of roughly 60 sites in Virginia, totaling a load in the range of 1,500 megawatts. On that occasion, many of the outages were traced back to protection systems that, detecting repeated voltage sags, disconnected the facilities according to their design logic.
When Multiple Loads React the Same Way
A power grid can manage faults and consumption fluctuations if user reactions remain distributed across time and space. Large industrial plants, buildings, and residential users have different profiles and do not necessarily respond at the same speed to a disturbance. AI-dedicated campuses, however, introduce a new combination: massive power draw, highly concentrated IT components, and similar behavior across sites built with analogous electrical designs.
According to an analysis published by MIT Technology Review, an AI campus can alter up to 70% of its power draw in milliseconds during a training run. When thousands of GPUs spin up simultaneously, the shift looks nothing like a traditional, gradual increase in demand. Similarly, when faced with an upstream sign of instability, systems can choose to quickly disconnect the data center from the grid to protect extremely high-value computing infrastructure.
For an individual operator, this is a prudent decision. For the power system, however, it becomes a risk if dozens of sites adopt the same measure simultaneously. A sudden loss of gigawatts of demand can be just as destabilising as a sudden loss of generation, because it forces the grid to rebalance production and consumption within extremely tight timeframes.
This is where the expansion of AI collides with an infrastructure built on different assumptions. Electrical protection systems and connection standards were developed in an era when even a large load was measured on a much smaller scale compared to the gigawatt campuses currently being planned. This does not mean existing facilities were poorly designed: it means the size and ramp rate of the load are changing the environment in which they must operate.
The limits of the traditional electrical chain
The typical data center configuration delivers medium-voltage electricity to the facility, steps it down to low voltage, and routes it through uninterruptible power supplies, or UPS, before reaching the racks. These systems are designed to keep equipment powered during an outage, providing the time needed for emergency generators or other backup reserves to come online.
In the model described by the source, that system reveals three limitations when scaled to AI workloads. The first involves location and capacity: the UPS is situated close to the IT equipment, and its batteries are suited to covering short-duration blackouts, not necessarily to continuously compensating for rapid, deep load swings.
The second concerns the operating mode. To reduce conversion losses, many operators use configurations where utility power reaches the racks directly via a bypass. It is an efficiency-driven choice, but it means that some compute load fluctuations can reach the grid unattenuated, and that extremely brief transients from the grid can hit the facility before a transfer switch can intervene.
The third limitation lies in the protection logic. Rules designed to disconnect a site after a specific sequence of voltage drops may be reasonable to prevent local damage. However, when applied across a vast and homogeneous set of large data centers, they turn a disturbance into a de facto coordinated disconnection. The 2024 incident in Virginia is cited specifically as an example of this collective effect.
Bringing protection closer to the grid
The solution proposed in the MIT Technology Review analysis—content labeled as sponsored and associated with ON.energy—involves shifting protection and conditioning electronics from inside the data center to the medium-voltage level, that is, the point where large facilities receive electricity from the grid. This refers to voltages of at least 13.8 kilovolts and modules located near the substation, rather than in the rooms housing servers and uninterruptible power supplies.
The idea is not merely to have a backup that kicks in after detecting a fault. The system would remain permanently in the power path, absorbing load fluctuations and isolating downstream equipment from disturbances. Under this approach, the goal is to present a smoother consumption profile to the grid even when computing loads shift rapidly.
Such an architecture would also carry implications beyond reliability. If the interface between the data center and the utility is consolidated into a single medium-voltage block, the interconnection process could be simpler for the grid operator to assess compared to a facility comprising transformers, UPS, chillers, pumps, and switchgear that must be examined separately. The source claims this could shorten permitting timelines and allow computing hardware upgrades without repeating the entire interconnection study.
Nonetheless, several aspects remain to be proven in the field. Promises regarding permitting times, footprint density, or backup system costs depend on local regulations, the grid operator, individual campus design, and contractual supply terms. Furthermore, a new electrical architecture does not eliminate the need to generate and transmit more electricity: it can make the load more manageable, but it cannot create energy capacity out of thin air.
A responsibility that cannot remain on the sidelines
The lesson from the Virginia incidents is that data centers are no longer just large grid consumers. At the power levels projected for AI, they become elements with a significant impact on power grid stability. Utilities, campus developers, suppliers of UPS systems and medium-voltage equipment will therefore need to discuss not only how much power to deliver, but also how that power will be used under normal conditions and during a fault.
For operators, investing in protection capable of handling rapid fluctuations can mean reducing downtime risk and freeing up space previously occupied by internal electrical infrastructure. For grid operators, the priority is preventing safeguards designed to protect a single building from amplifying a disturbance on a regional scale. Finally, for the AI sector, this represents a concrete industrial constraint: GPU availability matters little if the electrical system powering them is not designed for their collective behavior.
