Grid-Interactive Behavior: Ride-Through, Reactive/Voltage Support & Frequency Response Toward the POI

A data center's relationship to the grid was long one-directional: the grid delivered clean power, and when it did not, the facility protected itself. A double-conversion UPS carries the load on stored energy, a static-transfer switch moves it to another source, and a conservative undervoltage relay trips it. All three exist to isolate the IT load from a disturbed grid. That instinct, protect my electronics first, was reasonable when a data center was a 10–30 MW rounding error on a feeder. It becomes a planning problem when one campus is 500 MW to a gigawatt of voltage-sensitive load on a 230 kV or 345 kV bus.

That instinct now inverts. The question is no longer "how do I isolate my load from the grid during a fault?" but "how do I keep a gigawatt of load online through a fault the grid expects me to survive?" Get it wrong and the cost is more than your own outage. A block of load that leaves at once is a contingency that transmission planners now study, because it can push voltage and frequency out of range, and NERC has written Essential Actions to prevent it. In 2026 NERC formalized that shift: its highest alert tier in May, and in October the final ballot that passed its first computational-load Reliability Standards. This chapter covers ride-through, reactive/voltage support, and frequency response toward the point of interconnection (POI), the three behaviors that decide whether your facility is modeled as a grid asset or a grid risk. Chapter 4.3 holds the canonical NERC obligations and Chapter 4.5 the transient physics; here they become relay settings, Mvar, and droop curves at the fence line.

On a summer day in 2024 in the Eastern Interconnection, a lightning arrester failed on a 230 kV transmission line. The fault was normally cleared: protection isolated the line in tens of milliseconds, as designed. What had not been anticipated was how much customer load would come off the grid in response to the voltage sags. About 1,500 MW of data-center load left the grid across the six-fault sequence, as multiple facilities reacted to a disturbance the rest of the grid shrugged off. The sequence produced six successive system faults over an 82-second window. Voltage in the load-loss area sagged to 0.25–0.40 pu of nominal during each fault, for 42–66 ms, and after the event system voltage rebounded to 1.07 pu as the suddenly unloaded grid over-swung high (NERC Incident Review: Large Load Loss). The 1,500 MW is the sequence total. About 1,260 MW was the sustained loss at the third voltage depression, and it stayed off the grid for hours.

The over-voltage is what operators fear most. When a gigawatt of load vanishes in milliseconds, the reactive power it was absorbing is left on the system and voltage rises, which can trip other equipment such as generators, capacitor banks and neighboring loads. The grid held, but a voltage-sensitive block that protects itself is not neutral: it turns a contained transmission fault into a swing across the surrounding grid. NERC's response was its rare Level 3 "Essential Actions" Alert, issued May 4, 2026, the highest of its three alert tiers, reserved for immediate risks to the bulk power system. Registered entities had to acknowledge by May 11 and respond by August 3, 2026. The seven Essential Actions turn large computational load from an emerging-risk footnote into a first-order planning input: the NERC May 4, 2026 Level 3 Alert, pp.2–3, recommends that transmission planners study stability margins and corrective actions against loss of firm load in normally cleared faults. Acknowledging and reporting are required; implementing the Essential Actions is not a mandatory Reliability Standard obligation.

It happened again, at about 2.5 times the scale, on 2026-07-22. A single normally cleared 230 kV fault in Ashburn sent roughly 3.8 GW of Data Center Alley load to backup in two waves. The nearer line terminal tripped in Zone 1 and the remote one in Zone 2, and the fault and every reclose attempt cleared in 2–3 cycles, as designed (PJM/Dominion, 2026-08-06). PJM system load fell 3.8%, from 99,984 MW to 96,205 MW, the largest such event in PJM's history. PJM reported no significant reliability impacts, but the voltage disturbance was observed from D.C. to Chicago. On 2026-09-08 PJM proposed voltage and frequency ride-through requirements that would keep large loads connected through normally cleared disturbances, with a FERC filing planned for November 2026 (PJM, 2026-09-10). NERC's 2026 State of Reliability had already counted two data-center-initiated load losses above 1 GW during 2025. These events are a pattern, and they grow with the installed base.

Fault-ride-through (FRT), or low-voltage ride-through (LVRT), is the duty to stay connected through a voltage dip and follow a defined voltage-versus-time and power-recovery curve. The deeper the sag, the shorter the time the load must tolerate before it is permitted to trip. For a normally cleared transmission fault the load must meet the applicable connection, consumption and recovery requirements. Two regional rules now publish that curve for large computational loads.

SPP’s updated High Impact Large Load (HILL) ride-through table (revision of 2025-09-23) defines a HILL as a new commercial or industrial load, or an increase, at one site behind shared delivery points: at least 10 MW at 69 kV or below, or at least 50 MW above 69 kV. It applies to requests on or after RR696’s effective date of 2026-01-15; earlier requests need reasonable efforts within the facility’s capabilities. At the POI, each voltage has one duty. From 0.90 to 1.10 pu inclusive the load runs continuously. From 0.80 pu up to, but not including, 0.90 pu it must ride through 2.00 s. From 0.50 pu up to, but not including, 0.80 pu the duty is 0.50 s, and below 0.50 pu it is 0.15 s. Between 1.10 pu (exclusive) and 1.20 pu (inclusive) the load has to stay connected for 0.50 s, and past 1.20 pu the choice to ride through or trip is its own. Power-electronic and variable-speed-drive HILLs must keep drawing from the grid, cutting power in proportion to voltage in the 0.5–0.8 pu region under constant-current control with current limiting. Constant-power control is prohibited during disturbances. Below 0.50 pu the load may go fully onto UPS or trip. Once voltage is back at 0.9 pu, at least 90% of its grid consumption must return within 1 s. Where a facility counts disturbances to decide when to trip, the setting must tolerate at least six fault-clearing attempts in 90 s.

ERCOT’s NOGRR 282 / NPRR 1308 was PUCT-approved 2026-07-09 and is effective 2026-08-01 per the issue record. It sets voltage (Operating Guides §2.15) and frequency (§2.6.4) ride-through for non-exempt Large Computational Loads (LCLs): sites of ≥75 MW aggregate peak demand where ≥50% of demand is power-electronic computational load. A load that was operating, or had ERCOT energization approval, on or before 2025-11-14 is grandfathered. Other routes to exemption depend on the agreements, construction authorization, funding and study or attestation conditions in §2.15(1), and a qualifying later modification removes the exemption. Below 0.8 pu the load may reduce consumption but must return to at least 90% of its pre-disturbance draw within two seconds of voltage recovering above 0.90 pu. A load that rides a sag below 0.8 pu by transferring to another source or stabilizing its controls is on a faster schedule. Once voltage climbs back above 0.9 pu, it has 0.25 s to start restoring demand and 0.5 s to reach 90% of its pre-disturbance draw, both counted from that recovery. Chillers and other mechanical loads may trip below 0.35 pu, and if they do they are left out of the recovery tally. A stricter 0.5–0.8 pu profile applies once a load satisfies Planning Guide §9.5 after 2028-01-01, regardless of its energization date. Outside the US, Ireland’s Grid Code modification MPID345 (effective October 27, 2026) requires transmission-connected demand to ride through specified faults and restore at least 90% of pre-fault active demand within 500 ms of system voltage recovering to 90% of nominal (Chapter 3.2). None of these rules is your acceptance curve until the edition your connection agreement binds you to is in the design file. Set the sag, recovery and mandatory-setting values from that text.

The deep engineering problem is that AI load is a near-ideal constant-power load (CPL). The front-end rectifiers regulate output power against voltage, so when grid voltage sags, current rises to hold power constant. That is the wrong sign: it drags voltage down further and looks to the grid like negative incremental resistance. With limited DC-link energy, a CPL tends to trip on deep, short sags unless storage or coordinated controls buffer it. Ride-through is therefore a coordinated chain of settings: the protective relays (SEL or equivalent) at the POI and on the MV distribution; the UPS transfer logic, which decides when to go to battery and when to hold on grid; and the undervoltage load-retention settings at the rack and rectifier. Each must keep equipment connected through the curve and deliver the required POI reduction and recovery without an early trip.

Ride-through keeps you online; reactive support keeps the bus healthy. A large AI facility’s net reactive demand includes its active-front-end operating point, transformers, motors and cable charging. Inductive and capacitive contributions can oppose each other, and a near-unity-PF rectifier need not draw much Mvar itself, so model the net POI exchange before sizing voltage support. Grid codes and large-load tariffs increasingly require the facility to manage its power factor at the POI, and your host utility’s requirements and tariff set the duty. One example is East Kentucky Power Cooperative’s requirement of 0.95 lagging to 0.98 leading for connected loads (Facility Connection Requirements Rev. 16, §9.0, effective 2025-08-15). The tariff and site agreement also set the exact leading and lagging P/Q range. In the strictest jurisdictions the facility must also provide dynamic voltage support: fast reactive injection during disturbances, distinct from steady-state correction.

The provision ladder runs from cheap and static to expensive and fast. Switched capacitor and reactor banks correct steady-state power factor at low cost but respond in cycles to seconds, so they cannot help a millisecond sag. STATCOMs (static synchronous compensators) inject or absorb Mvar continuously and respond in milliseconds. They are the workhorse for dynamic-VAR duties, and gigawatt campuses increasingly specify them for the on-site substation. A BESS with qualified inverter controls can ride through the fault, and its controls set whether active or reactive power takes priority once the inverter reaches its current limit (NVIDIA, Designing Production-Ready Battery Energy Storage Systems for AI Factories, 2026-06-10). To do several duties at once, it needs validated headroom in MVA, current, energy and state of charge. Grid-forming control sets a voltage and frequency reference, including in an island, but reactive support and frequency response do not require it. Grid primary frequency response is a separate duty (NERC Primary Frequency Control guideline, September 2023), and a facility can take part only where the applicable tariff allows. The choice runs from meeting a steady-state PF target with banks (cheap, passive) to committing to dynamic support with a STATCOM or qualified BESS. That costs more, but it turns you from a voltage problem into a voltage solution and may earn grid-services revenue (Chapter 15.8).

The third grid-interactive behavior is frequency response, where the data center is mostly an undeveloped asset. A passive constant-power load contributes nothing to frequency regulation. After a generation loss pulls frequency down, that load keeps drawing its full demand. The balancing authority has to cover the gap from its reserves and from generators already providing primary frequency response. Governors on online generators act without being told. Inverter-based plants can help too, but only if they have controls that respond and spare active power to give. A grid-interactive load sheds or curtails demand on under-frequency and acts alongside those resources. Whether it helps depends on its measured response time, the megawatts it delivers and how long it can hold the reduction.

AI load may suit this, but each site has to prove it. Training is checkpoint-tolerant and batch inference is interruptible (Chapter 1.1), yet the number that counts is how much of your load is power-capped or paused on command. Pin down the eligible jobs, the control latency, the deliverable MW and its duration, the recovery and rebound, and the checkpoint and replay cost. The engineering is a droop characteristic measured at the facility: define a frequency deadband and a slope (MW shed per 0.1 Hz below nominal), wire it to a fast load-control layer that throttles GPU power caps or pauses curtailable jobs, and represent that behavior in the utility's model. This is a real strategic fork. You can stay a passive load, which is simplest but adds to the reserve burden you are increasingly asked to fund, or engineer load-side response, which adds control complexity but makes the facility a grid resource with a revenue line. Chapter 4.8 treats the islanded-microgrid version, where the facility's own inertia and grid-forming inverters set frequency.

In ERCOT part of that choice is already made. Texas SB 6 (2025) requires the serving utility, before it connects a transmission-voltage load after December 31, 2025, to have a protocol and any equipment it needs to curtail that load during firm load shed (Chapter 3.2). Your transmission owner may also put the site’s feeders in its region’s under-frequency load-shedding (UFLS) scheme. Run the mandated curtailment, any UFLS trip and any voluntary droop response through one tested load-control path; separate schemes can act against each other or against your ride-through settings.

Every decision in this chapter comes back to one question: is your facility a grid asset or a grid risk? The risk is a voltage-sensitive constant-power block that trips or transfers to backup. It can protect its electronics while causing a demand-loss contingency. Your actual settings, the study result and the agreed reinforcement allocation determine the cost, and the submitted dynamic model and the operative regional curve determine acceptance. The asset posture is the engineered alternative: a load that rides through faults, holds local voltage with dynamic VARs, sheds on under-frequency, and exports flexibility the balancing authority will pay for. The capital delta between the two (a STATCOM or qualified BESS inverter controls, retuned protection, a fast load-control layer, a verified dynamic model) is real but bounded. It can change the study result, required reinforcement, operating obligations and energization milestone under the host utility's process.

In 2026 the grid stopped asking informally and started writing it down: a Level 3 alert, annual planning studies, ride-through rules in ERCOT (NOGRR 282) and SPP (the HILL table), and FERC's July 16, 2026 order in Docket RD26-7-000. That order directs NERC to file Phase I computational-load Reliability Standards and registry criteria by December 31, 2026 and a Phase II workplan by March 1, 2027. Those are filing deadlines for NERC, not operator compliance dates. The Phase I standards (CLO-001-1, CLO-002-1 and CLO-003-1) passed NERC's final ballot (announced 2026-10-15) and take effect only after NERC Board adoption and FERC approval (Chapter 4.3). They reach Computational Load Sites of 50 MW or more supplied at 100 kV or above. Planners keep the interconnection studies, but the site’s owner supplies the model data they use and verifies it, coordinates its protection settings with the equipment around it, and records sequence-of-events data at its breakers; the operator provides the operational data its grid operators specify and follows their operating instructions. A site already in service when CLO-003-1 takes effect gets 36 months to meet its disturbance-monitoring requirements, so put the recorders and protection-coordination studies in today’s design rather than a retrofit. The standards set no ride-through curve; check the filing in force before you fix settings.

The liability can also cap usable power after you connect. ERCOT's Q4 assessment update of September 17, 2026 named three groups of loads that could trip together. Each has a mitigation plan that switches on near 2,500 MW of combined consumption, and none was active at that date. When ERCOT activates a plan, TSPs and QSEs must keep each group under its allocated not-to-exceed limit, derived from a 2,720 MW control target that sits below the 3,200 MW interconnection reliability operating limit (IROL) ERCOT studied. Better ride-through models plus corrected approved capacities removed the WESTTX_LL constraint. A passive load is no longer only leaving grid-services revenue on the table (Chapter 15.8); it carries a reliability liability that shows up as interconnection delay, study cost and reportable exposure. Chapter 12.2 develops the facility-as-grid-risk framing and how it interacts with cluster goodput when the grid asks your load to flex. Design the building as an asset from the first single-line, and the grid-interactive behaviors follow from decisions you were making anyway about UPS topology, BESS sizing and substation reactive plant. Bolt them on later and each one is a retrofit.