Most reporting on AI data centres asks how many megawatts they will draw. Grid operators also ask what that load does in the milliseconds after a fault on the transmission system, because a large load that disconnects itself is a sudden loss of demand, the mirror image of a power plant tripping. This article follows one documented incident, explains the equipment behind it, and then reads the two 2026 FERC orders that turn the issue into regulation. Everything below comes from the regulators' own documents; where we interpret, we say so.
What happened on 10 July 2024
NERC, the North American Electric Reliability Corporation, published an incident review on 8 January 2025. At about 7:00 p.m. Eastern on 10 July 2024 a lightning arrestor failed on a 230 kV line in the Eastern Interconnection, causing a permanent fault. The line's auto-reclosing control was set for three reclose attempts staggered at each end, which produced six successive faults in an 82-second period. The protection system cleared each properly; the shortest fault lasted 42 milliseconds and the longest 66.[3]
In the same area about 1,500 MW of load disappeared. NERC states that none of it was disconnected by utility equipment: "the load was disconnected on the customer side by customer protection and controls", and all of it was data-centre load.[3] Frequency peaked at 60.047 Hz and settled back in about four minutes; voltage reached 1.07 per unit, and operators removed shunt capacitor banks to bring it down.[3] NERC's summary is measured: neither frequency nor voltage reached a level that posed a reliability risk, but the grid "has not historically experienced simultaneous load losses of this magnitude in response to a fault".[3]
Why a data centre leaves the grid on purpose
A data centre's electrical design aims to keep servers and cooling running through grid trouble. NERC describes the standard tool, the uninterruptible power supply (UPS), and says the design differences change what the grid sees.[3]
| Design | What the UPS does | What the grid sees |
|---|---|---|
| Static UPS (centralised, typically 2 to 5 MW, or rack-level, typically 3 to 4 kW) | Switches the load to batteries during the disturbance, then back | A short loss of load that returns quickly |
| Rotary UPS with diesel (DRUPS) | Moves the load to a flywheel and generator, starts the engine | A loss of load that does not return quickly; transfer back is typically manual |
| Disturbance-counting scheme (on centralised or rack-level UPS) | Moves the load to backup after a set number of disturbances in a set time | Load stays off until manually reconnected |
The third row decided the 2024 event. Owners told NERC the typical trigger is three voltage disturbances within one minute.[3] NERC's timeline lists voltage depressions at 19:00:23, 19:00:23 (the simultaneous reclose at both terminals) and 19:00:39, so a third disturbance arrived about 16 seconds after the first, well inside a one-minute window. NERC reports that about 1,260 MW dropped at the third dip "and did not return for hours", and attributes most of the loss to "the interaction between the automatic reclosing sequence on the faulted transmission line and the data center's protection/control scheme".[3]
Our reading: nobody designed the grid to lose 1,260 MW of demand because a protection relay was counting. Each piece worked as intended; the problem is the combination, and it is only visible if the data-centre settings are known to the people configuring the line's reclosing.
Crypto mining is a different case
NERC's January 2026 review of Texas shows that "data centre" is not one behaviour. Crypto-mining facilities, it says, typically have no UPS or backup generators and rely on constant-power supplies that shut down almost instantly when voltage dips. ERCOT identified 26 events with indirect load loss above 100 MW from crypto facilities between January 2023 and September 2025, with reductions of roughly 17% to 95% of pre-event consumption.[4] NERC also says these facilities return gradually, which limits sudden demand spikes on recovery.[4] The 2024 event shows the other half of the picture: sites with UPS systems ride through with batteries, but whether and how fast they come back depends on equipment and settings, as the table above shows.
What FERC has now ordered
The regulatory path starts with the US Energy Secretary's advance notice of proposed rulemaking of 23 October 2025 (docket RM26-4-000), which asked FERC to consider reforms for large loads.[1][2] Two 2026 orders followed. Neither is a final national interconnection rule.
16 July 2026: reliability standards. FERC directed NERC to file, by 31 December 2026, "one or more new or modified Reliability Standards" for integrating computational loads, with glossary definitions, and revisions to NERC's Rules of Procedure including registry criteria for computational load entities. NERC must also file an informational Phase II workplan by 1 March 2027.[1] FERC says "NERC has documented multiple grid disturbances in which computational loads have caused or contributed to the instability of the Bulk-Power System." NERC's draft definition, quoted in the order, is load "comprised of power demand from information technology equipment, such as servers, storage, and networking hardware."[1]
18 June 2026: MISO show-cause. FERC found MISO's tariff "appears to be unjust, unreasonable, or unduly discriminatory or preferential" and gave MISO and its transmission owners 60 days to show cause or explain a remedy, with a 30-day informational report on generation adequacy.[2] The order preliminarily suggests defining a large load as a new commercial or industrial customer at a single site with peak load of 50 MW or more, connecting above 69 kV, and not part of a co-location arrangement.[2] It also proposes new services for "flexible" large loads able to limit their draw under certain conditions, and says it is not intended to disrupt existing commercial arrangements.[2]
What to ask about a data-centre grid claim
The following is our own checklist, built from the NERC findings; it is not from the orders.
- Ride-through settings. Does the site stay on the grid through faults, and what does the UPS do after one?
- Disturbance counting. Is there a count-and-transfer scheme, and what are its thresholds?
- Return behaviour. Does load come back automatically, manually, or at a controlled ramp? NERC lists ramp rates in operating agreements as an action for operators.[3]
- Model sharing. Does the operator have a dynamic response model? NERC recommends requiring one in interconnection requirements.[3]
What remains uncertain
We read the two FERC orders and two NERC reviews. We did not read the responses MISO and its transmission owners filed after the 60-day deadline, NERC's draft standards, or any order after 16 July 2026, so the current state of those proceedings may differ. The 2024 numbers describe one event; they are not a measure of how often it happens. For how much electricity AI facilities use overall, see our reading of US official data, and for the cooling side of the same sites, water versus energy in data centre cooling. Inside the building, AI cluster networking covers the other constraint on scale. Topic hubs: infrastructure and policy.
