What happens when the grid fails
When the utility supply to an AI data center fails, the racks keep running. A chain of machines takes over in a fixed order: batteries at once, diesel generators within 10 seconds, chillers within minutes. Scroll, and the clock runs.
The diagram is a simplified example site. Figures come from the standards and vendor documents listed under Sources. Typical values are marked as typical, and the temperatures are illustrative.
Normal operation
Before the outage
Power comes in from the utility at medium voltage, 13.8 kV in a typical US design. Transformers step it down to 480 V for the building (EC&M). Three diesel generators wait on standby at the top right.
Next comes the transfer switch, which connects the building to either the grid or the generators, and then the UPS. In a double-conversion UPS the rectifier turns AC into DC, the DC bus keeps the batteries charged, and the inverter turns DC back into AC. The racks always run from the inverter, never straight from the grid.
Each rack has two feeds, A and B, from separate UPS units and busways, and either one can carry it (ModulEdge). Inside an Open Compute rack, power shelves convert that to 48 to 50 V DC on a busbar; one current shelf delivers 33 kW (Advanced Energy).
Cooling is a second chain. A coolant distribution unit (CDU) pumps liquid through cold plates on the GPUs. The heat goes into a facility water loop and out through chillers. In this example the CDU pumps run from the UPS and the chillers do not. Keep an eye on both.
The grid fails
t = 0 to 50 ms
A fault on the utility line, and the site's voltage collapses. The racks notice nothing. The inverter was already carrying them, so nothing had to switch: a double-conversion UPS has no transfer time when its input fails (Schneider Electric).
50 ms to 1 s
The rectifiers have no input, so the batteries feed the DC bus. The chillers lose power and coast down. The CDU pumps keep turning, because they run from the UPS.
The transfer switch has seen the voltage drop, but it waits. NFPA 110 asks for a start delay of at least 1 second, so that a short dip does not start the engines for nothing (Nixon Power).
The generators start
t = 1 s to 4 s
The start signal goes out and the diesel engines crank. Each set has to reach 1,800 rpm, the speed that gives 60 Hz (CSE), and build up its voltage before it can take any load.
t = 4 s to 10 s
NFPA 110 calls an emergency system Type 10 when the load has acceptable power back within 10 seconds, with that first second of delay counted inside (ACHR News). This example site is Type 10. A single generator usually reaches rated voltage and frequency in about 10 seconds or less (Schneider Electric). Watch the battery gauge: it falls the whole time.
Transfer and walk-in
t = 10.00 s to 10.10 s
The generators are ready, and the transfer switch moves to the generator side. An open-transition switch breaks one connection before it makes the other, which takes 3 to 5 cycles, about 50 to 80 ms (Cummins). The UPS hides that gap the same way it hid the first one.
t = 10 s to 30 s
The UPS does not drop the full load on the engines in one step. Its rectifiers walk the load in over 10 to 20 seconds, and can be set to take up to about 30 (EC&M). As they take over, the batteries stop discharging. They have used a few percent of their charge.
Cooling rides through
t = 30 s to 85 s
The chillers got power back at 10 seconds, but a chiller cannot simply switch on again. A standard machine can take 10 to 15 minutes to restart (Trane). Quick-start models are faster: Daikin quotes a restart in 35 seconds and 80% load at 75 seconds (Daikin). The diagram uses those quick-start figures.
t = 85 s to 3 min
The coolant never stopped moving, because the CDU pumps are on the UPS. Without chillers the water warms; once they are back it cools down again. Vertiv notes that a high-power chip can reach its thermal limit in under a second if the coolant stops (Vertiv).
The temperatures here are illustrative. How fast the water warms depends on the loop's volume and the load, and we found no published figure for it.
Running on generators
t = 3 min to 1 h
The site now runs on diesel. The rectifiers recharge the batteries, and the operators watch the fuel. Published summaries of the Uptime Institute Tier standard give 12 hours of on-site fuel at design load as the minimum (POWER). ANSI/BICSI 002 uses 96 hours, and many operators design for 24 to 96 hours.
t = 1 h to 4 h
Large sites run many generators in parallel. Sets built for data centers start at about 1.5 MW, the most common size is 2 to 2.5 MW, and the largest reach about 4 MW (TechTarget). A 10 MW load needs five 2 MW sets, or six for N+1 (CSE). That sixth set is the subject of the next chapter.
What if a generator fails to start?
Again from t = 0, with N
Run the outage again, and this time generator 1 fails to start. With N redundancy the site has exactly as many generators as the load needs: two here, so generator 3 is not installed. One set alone cannot carry the hall, so the transfer switch stays put and the UPS stays on battery.
t = 12 s to 5 min 30 s
UPS batteries are sized for minutes. The Green Grid gives design runtimes of 1 to 2 minutes for hyperscale sites, about 5 minutes for colocation and 10 to 15 minutes for financial users (The Green Grid). This example has 5 minutes. When the battery is empty, the racks go dark. A real site might shed part of the load to keep the rest running.
Again from t = 0, with N+1
Now the same failure with N+1: three generators where two are needed. Generator 1 fails, generators 2 and 3 start, and the transfer happens at 10 seconds as before. The racks never see the failure. Uptime Institute's 2025 outage analysis still finds power the most common cause of impactful outages (Uptime Institute), with UPS failures a frequent source (DCK).
2N
2N doubles the whole chain: two independent systems, A and B, each able to carry the full load. Every rack has an A and a B feed, so a whole side can fail and the other side carries the racks, as B does here when UPS A fails. Tier III asks for concurrent maintainability, Tier IV for fault tolerance (POWER).
What AI loads change
Inside an AI rack
AI racks carry energy storage of their own. Some Open Compute racks add battery backup units in the power shelf format, rated for a few minutes; published ratings conflict, so take about 4 minutes as an example (Analog Devices). NVIDIA's GB300 NVL72 puts 65 joules of capacitors per GPU in its power shelves (NVIDIA).
Training load swings
A training job makes thousands of GPUs compute and then exchange data in step, so the whole hall's draw rises and falls together. NVIDIA smooths this with those capacitors, a power cap at start-up and a "power burner" that keeps GPUs drawing power as a job winds down, and reports up to 30% lower peak demand from the grid (NVIDIA). The storage smooths the load. It does not back it up.
The swing in the diagram is illustrative: first without smoothing, then with it.
10 July 2024, Northern Virginia
A lightning arrester failed on a 230 kV line, and repeated reclosing caused a series of voltage dips. About 60 data center load points at 25 substations saw the dips and switched to backup, most of them at the third dip. About 1,500 MW of load left the grid in one event (NERC).
What it did to the grid
Each site did what this page describes, and the grid suddenly carried 1,500 MW less. Operators removed shunt capacitors to keep the voltage from rising too far, and the load did not come back for hours (APPA). It happened again on 17 February 2025, about 1,500 MW, and on 22 July 2026, about 3,800 MW (APPA).
NERC's response
NERC has issued a Level 3 alert on large data center loads (Renewable Energy World). Under a FERC order of 16 July 2026, NERC must file the first phase of enforceable standards by 31 December 2026, and ride-through requirements are expected in 2027 (APPA). Ride-through means a site stays on the grid through a short dip instead of leaving it.
The grid comes back
t = 4 h: utility back
Back to the first outage. Hours later the utility returns, and the transfer switch does not jump straight over. It waits until the supply has been stable for a set time, and if the supply wobbles, the timer starts again (Cummins). Operators choose that time; this example uses 5 minutes.
t = 4 h 05 min: retransfer
Then it retransfers. A closed-transition switch connects the utility before it lets go of the generators, with less than 100 ms of overlap, so even the UPS sees no break (Siemens). The generators run unloaded for a few minutes to cool down, then stop; this example uses 5 minutes. The batteries finish charging, and the site is back where it started.