AI workloads have roughly doubled the average watts per rack at most new deployments, and the most common buyer mistake is reusing the UPS capacity plan that was built for the old enterprise hall. That plan quietly under-specifies protection exactly where the new load is densest.
A conventional enterprise rack of general-purpose servers drew perhaps six to ten kilowatts. Mainstream AI inference and fine-tuning deployments now land in the twelve to twenty kilowatt band, and training rows push well beyond that. The average across a new hall has roughly doubled, which means the aggregate the UPS must support has doubled even when the number of racks stayed the same.
This is not a headline efficiency story; it is a capacity story. The building's power chain was sized for the old peak, and the new peak arrives in the same floor space.
Legacy UPS sizing started from nameplate IT load and added a margin for growth. That method assumed a slowly varying, predictable draw. AI clusters execute synchronised jobs, so thousands of accelerators can swing load across a row within a fraction of a second as a job starts or checkpoints. A unit sized comfortably for the average now sees step loads that approach its instantaneous rating far more often.
The second failure is heat. Doubling watts per rack doubles the heat the power and cooling chain must remove, and the UPS conversion loss is part of that heat. A plan that ignored the extra thermal load leaves the room short of air or floor space.
Start from measured peak watts per rack, not the average nameplate, and multiply by the rack count for the affected zone. Add the loss of the UPS itself, typically three to four percent in double conversion, plus the distribution and cooling overhead. Then apply headroom of twenty to thirty percent so a single job burst or a future capacity step does not breach the rating.
Where a generator is present, battery runtime only has to cover start and stabilisation, usually five to ten minutes. Where there is no generator, runtime is a business decision about how long the zone must keep serving guests or customers.
Centralised UPS at the room or row remains the simpler option to monitor and service, and it still fits many AI inference halls. Rack-level or in-rack modules trade that simplicity for shorter distribution runs and the ability to scale a single dense row without oversizing the whole room.
The buyer's choice depends on how concentrated the density is. A hall with a few very hot rows benefits from localising capacity there; a uniformly dense hall is often cheaper to protect centrally. Either way, the unit must support the intended distribution voltage natively, without an added transformer stage that wastes floor and energy.
Liquid-cooled AI racks move most of the heat off the air path, but they add pumps, controls and a coolant distribution unit that also need clean power. The UPS plan has to cover that support load, not just the compute. A common omission is protecting the cooling plant itself, which defeats the purpose if the room coasts to thermal shutdown during a short outage.
Under-specifying is the visible failure: alarms during every job burst and a real risk of transfer to bypass. Over-specifying is the quiet one: a doubled frame that idles at low load runs less efficiently, costs more to buy and cool, and wastes floor. The aim is to size against measured peak with deliberate headroom, then let monitoring drive the next step rather than guessing it at purchase. Our data-centre power configurations are outlined at https://www.upsboss.com/data-center/.
If you are buying capacity for a new or retrofitted AI hall, send us the rack layout and the vendor's peak draw, and our engineers will return a UPS and battery plan sized to the real load. The current product range is at https://www.upsboss.com/products/.
Key takeaway: the watts-per-rack number has roughly doubled, so resize from measured peak, not the old average plan. Cover the cooling plant too, demand step-load data in writing, and leave room for the next capacity step instead of guessing it at purchase.
Can I just add more racks to my existing UPS?
Only if the existing unit has headroom against the new peak, not just the old average. AI density usually removes that headroom, so the plan needs revisiting rather than extending.
Does liquid cooling mean I need less UPS?
No. It reduces air-side heat but adds pump and distribution loads that also need protection, and the compute draw is unchanged. The cooling plant must be on the UPS too.
Why ask for step-load response in writing?
Because average efficiency hides the behaviour under a synchronised job burst. Measured step-load and overload figures tell you whether the unit will hold voltage when the cluster swings, which is the real test.
Contact: Frank Zhang
Phone: +86-135 5688 8641
Email: frank@upsboss.com
Add: Jufeng Road, Guangming Street, Guangming District, Shenzhen City, Guangdong Province, China
