The period after a grid disturbance can become an important part of the electrical response of a large AI site, particularly as the site begins restoring demand. A fault can push voltage outside normal operating conditions, trigger protective responses, interrupt portions of a compute workload and leave the electrical system moving toward a new operating point. The grid operator then has to understand not only what disappeared during the disturbance, but also what the connected load intends to do once voltage and frequency recover. That second movement matters because modern compute sites can contain large populations of electronically controlled loads whose electrical response can depend on power-conversion controls, protection settings and other coordinated site systems. A site that returns almost exactly as it left can therefore produce a different system response from one that restores demand through a more controlled sequence.

That question has become more visible as grid operators encounter large computational loads whose electrical behavior does not always resemble the traditional industrial demand assumptions embedded in older planning studies. ERCOT has moved directly into detailed requirements for large computational loads, while PJM has been examining large-load ride-through and disturbance behavior following a major data-center load event in Virginia. In Great Britain, the system operator is now funding dedicated work to understand how transmission-connected data centers respond to faults, voltage disturbances and subsequent active-power recovery. These developments do not amount to one common rule, and they should not be treated as evidence that every operator has adopted the same technical position. They do, however, point to a common engineering concern: the behavior of a large computational site after a disturbance can affect the wider power system and therefore warrants consideration alongside its behavior during the disturbance itself.

ERCOT’s First Realization — The Return Can Be Bigger Than the Fault

ERCOT’s recent work provides the clearest public evidence that large computational loads require a more detailed treatment of disturbance behavior than a simple connected-or-disconnected model. In 2025, ERCOT reported that it had observed several events in which multiple large loads disconnected following transmission faults, while the dynamic models used in studies did not indicate that those facilities should trip for the same faults. That mismatch prompted the system operator to examine the actual electrical response of electronically controlled loads alongside the behavior represented in planning models. The concern was not limited to the immediate loss of demand, because a simultaneous response from many facilities can alter frequency and voltage behavior across the system and can complicate the sequence that follows fault clearing.

The disturbance does not end when the voltage comes back

The next layer is the return profile itself, where the electrical system encounters the effects of control decisions made inside the site as voltage recovers. ERCOT’s published testing material explicitly addresses the recovery of active power after a voltage disturbance, including how a large electronic load should return toward its pre-disturbance operating condition after the grid voltage recovers. That approach treats recovery as a dynamic response rather than an automatic consequence of fault clearing, while the timing of individual site systems can influence when electrical demand resumes. A site can therefore experience an internal sequence in which protection clears, control systems validate electrical conditions and different site systems progressively return to service. If several large sites follow comparable control logic, their recovery responses could become correlated rather than being distributed independently over time.

Recovery becomes a system variable

ERCOT’s progression from observed load-loss events to formal ride-through requirements shows why recovery behavior has become part of the technical treatment of large computational loads. The operator’s final 2026 interim assessment identified insufficient voltage ride-through capability among groups of large computational loads as a potential source of substantial load loss, reinforcing the connection between individual control behavior and system reliability. The resulting requirements address how large computational loads should behave across different voltage conditions and how their active-power consumption should recover after voltage returns toward normal operating conditions. That architecture effectively gives the recovery curve a place inside the technical conversation, even though the rules do not amount to a universal “AI snapback standard” across power markets. The significance lies in the direction of the engineering work: operators are increasingly interested.

That shift also explains why the phrase “fault ride-through” can become too narrow when engineers discuss the next generation of computational interconnections. Ride-through answers an essential question about whether equipment remains connected or responds appropriately while the grid experiences abnormal electrical conditions, but the system still has to deal with whatever happens after the voltage disturbance clears. A controlled return can spread the restoration of active power across different internal blocks, while an uncontrolled return can align multiple electrical and computational processes around the same recovery trigger. ERCOT’s current requirements therefore provide an early indication that the interconnection model for AI sites is becoming more behavioral, with voltage response, active-power response and recovery characteristics considered together rather than as isolated technical attributes.

PJM’s Shift in Questioning — Not If You Stayed On, But How You Came Back

PJM’s recent experience has pushed the discussion of large computational loads beyond the question of whether a site should remain connected through a grid disturbance. On July 22, 2026, a normally cleared transmission fault in northern Virginia triggered an unexpected transfer of data-center load away from the grid, producing a large and rapid change in system demand and voltage conditions. PJM and Dominion subsequently described the event as the largest such large-load transfer that PJM had experienced, with operators reviewing why computational loads responded differently from the behavior expected in planning models.. PJM’s public process does not establish that every site must use one particular recovery architecture, but it does show that the electrical behavior of large computational loads is receiving explicit regulatory and engineering attention.

The event that changed the question

The wording emerging from PJM’s response is important because the operator is examining both existing and future ride-through requirements for large computational loads rather than treating the July event as an isolated operating anomaly. PJM has established a dedicated large-load ride-through education process, with September 2026 Planning Committee materials covering background, proposed requirements and draft language for computational large loads. That work indicates a movement toward specifying expected electrical behavior rather than relying solely on conventional assumptions about how a customer load should respond when the grid experiences abnormal voltage conditions. For an AI site, the resulting technical conversation reaches into controls that govern the transition between grid-connected operation, internal protection, backup operation and eventual restoration of normal supply.

The change becomes clearer when the interconnection point is treated as the place where the combined electrical effect of internal site controls becomes visible to the wider power system. A computational load can contain many independently controlled electrical paths, yet the grid does not necessarily observe those internal boundaries when their combined response appears at the transmission connection. If several paths restore simultaneously, the resulting active-power movement can appear as one larger electrical response even though the site experienced multiple internal control actions. The same issue can apply to reactive power, because changes in voltage conditions can alter the electrical behavior of power-electronic equipment as a site restores demand. PJM’s recent actions indicate that interconnection discussions are giving greater attention to the dynamic behavior of large computational loads rather than treating them only as static blocks of demand.

Recovery becomes part of the interconnection conversation

The deeper engineering question is what happens after a large load has successfully avoided an inappropriate disconnection and the surrounding electrical conditions have returned to a range in which the site can resume normal operation. A ride-through requirement can prevent unnecessary load loss, but preventing the first event does not automatically define the second phase of the response. The site still needs a controlled pathway from protected operation toward normal computation, and that pathway may involve several electrical and mechanical dependencies that do not recover at the same rate. Cooling availability can constrain compute restoration, power-conversion controls can impose their own restart logic, networking systems can determine when workload blocks become usable and internal protection can delay selected equipment until electrical conditions remain stable. Those dependencies mean that a predictable recovery profile can require deliberate coordination across systems that historically operated under different design objectives.

This creates a different kind of engineering request for developers because the grid operator needs more than confirmation that equipment can tolerate an electrical disturbance. Operators increasingly need to understand the shape of the load response that follows the disturbance, including whether the return occurs in blocks, whether those blocks depend on common triggers and whether the site’s controls can hold part of the demand back when system conditions require additional recovery time. Such information does not necessarily require disclosure of proprietary workload logic, because the relevant interface concerns electrical behavior rather than the commercial details of the computing operation. What matters is whether the site’s electrical controls can produce a response that engineers can represent, test and reproduce under the operating conditions considered during interconnection studies.

National Grid’s View — When Recovery Looks Like a Second Event

Great Britain’s current work offers a different but closely related signal because the system operator is examining large transmission-connected data centers before a mature body of operating evidence exists. The DC-RIDE project, launched in September 2026, specifically examines whether large transmission-connected data centers can meet emerging requirements for fault ride-through, voltage disturbances and fast active-power recovery. The project description identifies limited Great Britain-specific evidence about how large data centers respond to faults and voltage dips, creating uncertainty around their effect on system stability and security during network disturbances. The stated objective includes understanding whether data centers can remain connected and recover in a controlled manner without adversely affecting system stability, which directly connects the site’s return behavior to the wider transmission system. Britain’s approach therefore offers an early indication that recovery characteristics may enter future connection requirements through evidence gathering and modeling before a final technical rule exists.

Britain starts with the evidence gap

The timing of that work matters because the grid is encountering a class of demand whose internal electrical behavior differs from the historical assumptions attached to many conventional large loads. Large data centers rely heavily on power electronic conversion, digitally controlled equipment and tightly coordinated internal systems, creating a pathway through which a network disturbance can affect multiple electrical components at once. The system operator therefore needs to understand not only whether the site remains connected during a voltage disturbance, but also whether its active-power demand returns smoothly, rapidly or in a pattern that correlates with other large sites. A common recovery trigger can create a synchronized response even when the sites have no direct operational relationship with each other, because the external grid condition itself provides the shared trigger.

Britain’s existing grid-code history also shows why recovery cannot be separated neatly from the wider question of how connected equipment behaves around a fault. NESO maintains established processes for fault ride-through compliance, including rules covering situations where a user’s site or network asset trips or de-loads during a fault ride-through occurrence. More recent work continues to examine the technical requirements and compliance process, demonstrating that disturbance response already forms part of the broader grid-code architecture. The new data-center research extends that conversation toward a demand class whose dynamic behavior has received less direct study in the British transmission context. A site that survives the disturbance but then restores active power in a correlated manner can create a new system response that deserves separate analysis from the initial voltage event.

The second event is a recovery problem

The engineering issue is that the system can encounter a second significant change in electrical conditions after the initiating fault has already cleared. When a large computational site restores demand, its active-power trajectory can influence generation-load balance, voltage conditions and the operating margins available elsewhere on the network. A correlated recovery from several sites can therefore produce a system response that requires its own modeling even though no new transmission fault has occurred. That is why the DC-RIDE project places active-power recovery alongside fault ride-through and voltage disturbance behavior in its stated scope. The British approach suggests that recovery should be treated as a measurable characteristic of the connected load, with evidence gathered to determine whether existing requirements adequately represent the behavior of future large data-center connections.

For an AI site, that means the path back to full operation can become part of the electrical design rather than a sequence controlled only by workload orchestration. Compute blocks may become available at different moments, but the electrical controls can determine whether their associated power demand reaches the grid gradually or together. Cooling controls can create another dependency because computational equipment cannot return to sustained operation unless its thermal-management systems can support the restored workload. Power conversion equipment can also impose restart conditions that affect when individual blocks draw meaningful active power, while reactive behavior can change as converters move between protective and normal control modes. These interactions can create a recovery profile that looks simple from inside the site but becomes much more consequential when viewed from the transmission connection.

Three Markets, One Pattern — Correlated Return Is No Longer Theoretical

The emerging pattern across ERCOT, PJM and Great Britain is not a common rulebook, but a common movement in the questions being asked about large computational demand. ERCOT has focused on large-load voltage ride-through and active-power recovery after observed events in which computational loads disconnected in response to transmission disturbances. PJM has responded to unexpected large-load transfers by examining ride-through requirements for computational loads and developing a dedicated process for defining their expected behavior. NESO has launched DC-RIDE to gather evidence about fault ride-through, voltage disturbance response and active-power recovery for large transmission-connected data centers in Great Britain. Yet all three are moving toward the same underlying engineering question: how should a large electronically controlled load behave as the grid moves through a disturbance and then returns toward normal operation?

The common signal sits after the fault

The importance of correlation comes from the fact that grid disturbances can provide a shared external trigger to sites that otherwise operate independently. An AI site does not need to coordinate directly with another site for both facilities to begin changing their electrical demand after the same voltage condition crosses an internal control threshold. If similar equipment and control philosophies exist across a growing population of computational sites, their recovery behavior can become more closely aligned than traditional load models suggest. The grid operator then sees the aggregate electrical response at transmission level, where internal differences between sites may become less visible than the combined change in active and reactive power. This does not mean that every data center will respond identically, because site architecture, protection settings, equipment selection and control strategies can produce materially different responses.

The system-level issue becomes more pronounced as individual sites grow large enough for their internal control decisions to register clearly at the transmission interface. A small change in one building can disappear into the normal variability of a regional load profile, while the coordinated response of several large computational sites can become visible as a discrete electrical movement. That is why the current operator focus extends beyond steady-state demand and into dynamic behavior, particularly during faults and the recovery that follows them. The relevant question is not whether an AI site can operate efficiently under normal conditions, but whether its transition between abnormal and normal conditions remains sufficiently predictable for the surrounding power system to model and manage. ERCOT’s formal recovery requirements, PJM’s response to recent large-load events and NESO’s dedicated recovery research each point toward a future in which this behavior receives greater attention during connection studies.

The scale changes the nature of the problem

At large computational sites, recovery cannot be reduced to a single switch returning to its previous state because modern AI infrastructure contains multiple electrical and operational layers that determine when demand can resume. Power conversion, cooling, networking and compute systems can each impose conditions on the next stage of restoration, while protection controls can interrupt that sequence whenever electrical conditions move outside their permitted range. A site can therefore recover through several coordinated stages even when its operators describe the process simply as returning to normal operation. From the grid perspective, those stages matter because each stage can change the active and reactive power seen at the interconnection point. If several sites follow comparable sequences, the aggregate response can become sufficiently correlated to affect the wider system even though every individual site remains within its own engineering limits.

The emerging operator interest also changes what counts as useful evidence during an interconnection study. A static load value describes how much power a site consumes at a selected operating point, but it cannot describe how that demand changes when voltage falls, when protective controls act or when the site restores normal operation. A dynamic model can represent those behaviors, yet the quality of the result depends on whether the model reflects the actual controls installed at the site and whether those controls remain consistent as equipment configurations evolve. Operational telemetry can provide a separate validation layer by showing how the real site responds during controlled testing or actual disturbances. That evidence becomes especially important for AI infrastructure because changes in computing equipment, power architecture and cooling arrangements can alter electrical behavior without changing the site’s headline interconnection capacity.

What operators are beginning to see at the connection point

The most useful signals at the interconnection point are increasingly the shape and sequence of the response rather than a single measurement taken after conditions normalize. A staggered active-power return can show that the site is releasing demand through controlled blocks rather than restoring everything against one common trigger. Reactive-power movement can reveal how converters and other electronically controlled equipment respond while voltage recovers, particularly when several internal systems transition between operating modes. Sequencing information can show whether cooling, power conversion and compute blocks return through a defined order or whether multiple systems restart in parallel without a coordinated electrical limit. These signals can help an operator distinguish a site that simply resumes demand from one whose recovery follows a known control strategy.

Recovery telemetry also gives operators a way to compare the behavior represented in an interconnection model with the behavior that the site actually produces. That comparison can identify whether a model captures the timing of active-power restoration, whether reactive behavior follows the expected trajectory and whether internal controls create a synchronized response that the study did not anticipate. The purpose is not to expose the internal operation of the computing workload, because the grid does not need to know what an AI model is processing to understand its electrical response. What matters is the electrical envelope created by the site and the controls that determine how quickly demand returns after the grid recovers. This approach can also help site owners because a clear recovery profile provides a more defensible basis for explaining how the site will behave under unusual network conditions.

From Holding On to Letting Go Slowly — How Operating Dialogue Is Changing

The emerging operating dialogue does not require an AI site to remain at reduced output indefinitely after a disturbance, because the objective remains reliable restoration of normal operation. The change concerns how that restoration occurs and whether the site can control its return instead of allowing every internal demand block to respond to the same external signal. A managed re-entry sequence can hold selected compute blocks while electrical conditions stabilize, restore cooling capacity in stages and then release additional demand as the site confirms that the connection can support it. Such an approach gives the grid operator a more predictable electrical trajectory without requiring the site to abandon its own operational resilience strategy. It also creates a direct link between internal controls and interconnection performance, because the sequence determines what the transmission system sees as the site moves from disturbance conditions toward normal operation.

Managed re-entry replaces the simple return-to-service idea

The concept becomes especially relevant when the site operates through several independent electrical and computational blocks that can return at different times. Instead of treating the whole site as one demand element, the control system can recognize separate recovery states and release power according to conditions established at the interconnection point. Cooling systems can support selected compute blocks first, networking infrastructure can restore the paths required for those blocks and power-conversion systems can manage the associated electrical transition without forcing the entire site to return simultaneously. The approach creates an operating sequence that can be represented in a dynamic model and tested against telemetry from the actual connection. It also gives system operators a clearer explanation for why the site is increasing demand at a particular stage rather than simply observing a sudden restoration with no visibility into the underlying sequence.

The same logic changes the conversation between a site developer and a system operator during the interconnection process. Instead of discussing only the maximum demand that the site intends to draw, engineers can discuss how that demand will behave under abnormal voltage conditions and how it will return when those conditions clear. The site can describe the control states that govern recovery, the conditions that permit additional blocks to restart and the telemetry available to verify the resulting response. The operator can then evaluate that behavior against the characteristics of the surrounding network rather than relying on an assumed generic large-load response. Such dialogue can also expose design choices early, including whether a common recovery trigger exists across multiple electrical blocks and whether that trigger can produce an avoidable synchronized demand increase.

Cooling, network and compute blocks become part of the electrical sequence

A controlled recovery architecture also changes how cooling should be viewed within the electrical response of an AI site. Cooling is normally treated as a support system for computing equipment, yet during recovery it can determine how quickly additional computational blocks can return to service and therefore how quickly electrical demand can increase. If cooling capacity returns in stages, the associated compute load may naturally follow a staggered path, creating a more gradual electrical restoration. If cooling and compute controls share a common restart trigger, the same architecture can instead concentrate demand into a narrower recovery window. The electrical design therefore cannot be separated completely from the thermal and compute control layers when the objective is to understand the site’s response at the grid connection.

Networking introduces another layer because compute blocks do not become useful simply because electrical power becomes available. AI workloads depend on communication paths, storage access and coordinated compute resources, so restoring one electrical block does not necessarily mean that the associated workload can immediately resume at full demand. That creates an opportunity for site controls to coordinate electrical restoration with the actual readiness of the computational system rather than allowing power to return before the workload can use it. Such coordination can help prevent unnecessary electrical demand during intermediate recovery states while also giving the site greater control over the order in which computational resources become active. From the grid perspective, the relevant outcome remains the electrical response, but the internal dependencies explain why a carefully sequenced recovery can be more predictable than a single command that restores every available block.

The interconnection point becomes the place where recovery is proven

The next stage of the discussion is likely to center on what a site can demonstrate rather than what it can simply state in an application. A recovery philosophy can describe intended behavior, but operators need confidence that the controls will produce that behavior under the electrical conditions represented in the study. Testing, dynamic modeling and telemetry can connect the design intent to the actual response and identify differences before they become a system operating problem. That evidence can include the active-power trajectory, reactive-power behavior and sequencing of major demand blocks without exposing the details of the AI workloads themselves. It can also show whether the site’s response changes as different operating configurations come online, which matters because a partially commissioned site may behave differently from a fully populated one.

The operating model that emerges is therefore less about keeping every megawatt connected at all costs and more about controlling how the site changes its electrical state as the grid changes around it. Holding on during a disturbance remains important, but holding on without a predictable recovery pathway leaves the system operator with only half of the information needed to manage the load. A controlled release of cooling, networking and compute blocks can turn recovery into a sequence that the site understands and the grid can represent. ERCOT’s recovery requirements, PJM’s current ride-through work and NESO’s DC-RIDE research all point toward greater attention to this complete response rather than to the fault alone. The resulting interconnection question is increasingly straightforward even if the technical answer remains complex: when an AI site comes back, can the grid see exactly how it is coming back?

Recovery Will Decide How AI Factories Earn Their Place on the System

Sites in this emerging environment will need to consider not only how quickly they restore internal systems, but also whether the resulting electrical response can be accommodated and understood by the grid. A predictable recovery can give operators a clearer basis for planning, while a poorly understood recovery can create uncertainty even when the site’s equipment performs exactly as its designers intended. The difference lies in the shape of the electrical response, the coordination of internal blocks and the ability to demonstrate that behavior through models, controls and telemetry. That places the interconnection point at the center of a larger engineering conversation in which electrical, thermal, networking and compute systems increasingly need to operate as one coordinated recovery sequence.

The implication for future AI sites is not that every project needs to adopt the same recovery sequence, but that every project increasingly needs a recovery sequence that can be explained in system terms. The site must know which blocks can return first, which conditions permit the next release of demand and how its controls prevent a common external trigger from producing an unnecessarily synchronized response. It must also know how those decisions change as the site expands, because a recovery strategy that works for an initial configuration may behave differently after additional computational capacity enters service. Operators, meanwhile, need enough information to represent the site’s response accurately without requiring access to the commercial details of the workloads running behind the electrical boundary.