In short. AI and HPC halls put far more power into a small floor area than a conventional server room, and the load moves fast. Generator backup for these facilities must be sized from the UPS input behaviour, the cooling restart and the growth plan, not from a fixed multiple of IT kW. The deciding checks are load-step acceptance, UPS rectifier walk-in and battery recharge, and whether the cooling plant can be restored before the racks overheat.
Why AI and HPC halls behave differently at the generator
A conventional enterprise hall might run 5 to 10 kW per rack with a fairly flat load. GPU clusters are routinely specified at 30 to 100 kW per rack and higher, with direct-to-chip liquid cooling or rear-door heat exchangers removing the heat. Two things change for the standby plant. The mechanical load becomes a large, partly motor-driven block of coolant distribution unit pumps, chilled-water pumps, dry coolers and fans that restart together after a transfer. And the IT demand itself swings, because a synchronised training job can move a cluster from a fraction of its power to near full in well under a second. Reddy Generators treats such sites in Hyderabad as a system design problem first and a kVA selection second.
Building the generator-facing load from real data
The number that matters is what the generator terminals see after the UPS, cooling plant and auxiliaries are added. Start from a load inventory that separates the blocks, then confirm each with equipment data rather than multipliers.
P_gen = (P_IT / η_UPS) + P_recharge + P_mech + P_aux
P_gen is the running kW the generator must carry, P_IT is the expected IT load rather than installed capacity, η_UPS is the UPS efficiency on generator, P_recharge is the battery recharge power drawn after an outage, P_mech is the cooling and pumping load on generator, and P_aux covers lighting, controls and building services. A worked example (illustrative): a 2,000 kW IT block on a double-conversion UPS at 96 per cent efficiency draws about 2,083 kW at the UPS input. Recharge limited to 10 per cent of UPS rating adds about 200 kW, a liquid-cooled hall at a mechanical share of 0.25 adds roughly 500 kW, and auxiliaries add 100 kW. The running load is close to 2,900 kW before load-step margin, derating or growth.
| Load block | How it appears at the generator | Data to obtain |
|---|---|---|
| GPU and CPU racks via UPS | Filtered by the UPS, but fast swings pass through in some modes | UPS topology, eco or double-conversion mode, input power factor, input THD |
| UPS battery recharge | Extra block load for tens of minutes after transfer | Recharge current limit, whether recharge can be delayed on generator |
| CDUs, chilled-water pumps, dry coolers | Motor starts in a tight sequence with a large reactive component | Motor kW, starter type, VFD ramp, restart timers |
| Auxiliaries | Steady small load | Connected kW and diversity |
Load steps, UPS walk-in and cooling restart
A turbocharged diesel engine does not accept its full rating in one step. Typical sets take a first step of roughly 50 to 70 per cent of rated kW within the limits of ISO 8528-5 class G2 or G3, and the exact figure must come from the manufacturer’s load acceptance curve. Modern UPS systems help by walking the rectifier in over several seconds and by staging recharge. Poorly configured units do the opposite, so the UPS settings are part of the generator design and should be recorded and locked after commissioning. The background is in transient response and UPS and non-linear loads.
Repetitive swings from the workload
Training clusters can impose load changes of hundreds of kW every few seconds. In double-conversion mode the UPS smooths most of this; in eco mode the swings reach the governor and AVR almost directly, so ask for measured power traces from a comparable deployment before the generator is chosen.
Cooling restart is a time problem
High-density racks have very little thermal headroom. If coolant stops circulating, inlet temperatures can exceed limits within tens of seconds to a few minutes, so pumps and CDUs normally sit on UPS and the heat rejection plant must return within the time the thermal mass allows. The restart sequence is designed against that budget and proven in an integrated test.
Architecture choices for AI halls in India
Most sites in Telangana take 11 kV or 33 kV from the DISCOM and distribute at 415 V, with generators at LV for smaller blocks or at 11 kV with paralleling switchgear for multi-megawatt plants. N+1 is the usual minimum, with 2N where availability targets demand it. The number of sets, the bus arrangement and the paralleling controls are decided together, because one set failing to synchronise during a block transfer is a common cause of a lost hall. See N+1 systems and paralleling.
New diesel sets up to 800 kW must meet CPCB IV+, and larger engines fall under the norms for their class, so the choice between many sets below 800 kW and fewer large sets has an aftertreatment dimension as well as a redundancy one. Fuel autonomy is specified in hours at the real running load, and pre-monsoon ambient temperatures belong in the derating basis.
Proving it before go-live
A generator-only load bank test proves the engine, alternator, fuel and cooling systems. It does not prove the hall. The acceptance plan should include a full-rating generator test at site, a UPS and generator step-load test, a transfer sequence test from utility failure to steady operation, and an integrated test with the cooling plant restarting on its real timers. Resistive load banks suit engine commissioning, but a reactive component is needed at the design power factor; see load-bank sizing.
Record the current deployment, committed expansion and the infrastructure allowance for each stage, because a change in cooling method or rack density alters both electrical and mechanical interfaces and the original sizing is valid only for the stage it served. A stage that outruns the permanent plant can be bridged with data centre rental power.
Common mistakes
- Sizing the generator as a fixed multiple of the UPS rating without reading the UPS input settings. The set is either oversized and lightly loaded for years, or undersized for recharge plus cooling restart.
- Leaving the UPS recharge current at its factory default. The recharge block arrives as the chillers and pumps restart, and the combined step exceeds what the engine can accept.
- Testing on a resistive load bank and calling the hall commissioned. The UPS interaction at real power factor has never been seen, and the first outage becomes the first test.
- Treating the mechanical plant as a secondary load. In a liquid-cooled hall the pumps and CDUs are as critical as the racks, and on the wrong supply the IT load shuts down within minutes.
- Freezing the generator design for phase one and never revisiting it. Density grows and the standby plant falls behind the hall it is meant to protect.
What to send us
Send the planned IT load by phase, the rack density and cooling method, the UPS make, topology and input settings, a single-line diagram showing the utility intake and proposed generator bus, and the mechanical load list with motor ratings and starter types. If the hall is already running, a week of measured power traces at the UPS input is worth more than any nameplate list, and photographs of the generator yard help us check access and exhaust routing. Send your application details and we will reply with the questions that remain open.
Questions engineers ask
How much generator capacity do I need per MW of GPU load?
There is no fixed ratio. The generator must carry the UPS input load, battery recharge, the cooling plant and auxiliaries, and then satisfy the largest load step. For a liquid-cooled hall the running figure often lands between 1.3 and 1.6 times the IT kW before derating and redundancy, but the reliable number comes from the actual UPS, cooling and growth data.
Can a diesel generator follow the rapid load swings of a training cluster?
The engine and alternator follow steps within their load acceptance limits, but repeated swings of hundreds of kW every few seconds are better absorbed by the UPS or an energy buffer in front of the generator. The design should decide where that smoothing happens and prove it with a step-load test.
Do the coolant pumps need to be on UPS as well as generator?
Usually yes for direct-to-chip liquid cooling, because the thermal margin is measured in seconds. Pumps and CDUs go on UPS or a ride-through supply, while the heat rejection plant restarts on generator within the time the thermal mass allows.
How often should the integrated test be repeated?
At first commissioning, whenever a new IT block or cooling stage is energised, and annually. Only the integrated test shows whether the UPS settings, timers and generator still work together.
Technical content reviewed by Srinivas Reddy, Managing Partner, Reddy Generators. Last updated 19 October 2026.