Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

Servers & Power

Your UPS Was Not Designed for This: How AI Loads Break Old Power Assumptions

For decades, sizing a UPS was simple arithmetic. Add up the equipment, allow some headroom, choose a runtime. It worked because ordinary servers draw power the way a fridge does: steadily, with gentle changes.

A GPU cluster in training does not behave like a fridge. It behaves like a factory press. The load jumps up and down many times a minute, and all the machines do it together. That one change undermines most of the old assumptions.

What is different about an AI load

Compare 🗄️ Ordinary servers 🧠 GPU cluster in training
Shape of the load Steady
Changes slowly through the day
Pulsing
Rises and falls in seconds or less
Do machines move together? No
Peaks rarely line up
Yes
GPUs run in step, so peaks stack
Typical draw against nameplate Well below
Close to it under load
What the UPS and generator see A calm load Repeated steps that upstream equipment must ride through

The bars are illustrative, not measurements. Sources for the behaviour described: Schneider Electric: AI-tolerant UPSs, Vertiv: Adapting data center power infrastructure to AI workloads

What the vendors' own tests show

Schneider Electric, which owns APC, describes the cause plainly: training algorithms run in sync across the GPUs, which produces an oscillating load, and a UPS that is not designed for those frequent swings may drop the load.

Vertiv published a test using a load modelled on NVIDIA H100 training: two seconds at 90% of the UPS rating, then two seconds at 50%, over and over. Its large UPS held the output steady. A second Vertiv paper found a threshold worth knowing about. When AI made up roughly 75% of the total IT load, generator voltage and frequency could move outside the range the UPS will accept.

Two cautions. These are vendor tests of the vendors' own products. And they concern large facility UPS systems, not the unit under a desk. The lesson that transfers is the principle: ask how a UPS behaves under a pulsing load, not just whether the kilowatts add up.

Source: Vertiv: Advanced UPS controls for AI workloads

The battery has a new job

Traditionally the battery waits for a power cut. In AI facilities it is increasingly asked to work all the time, soaking up the peaks so the grid and the generator see a smoother load. Vertiv calls its version Input Power Smoothing. The newest NVIDIA racks go further and build energy storage into the rack's own power shelves, which we covered in AI Rack Power: Busbars, 54 V and 800 V DC Explained.

The trade-off is battery wear. A battery that cycles all day ages differently from one that sits idle, so battery type and replacement plan belong in the conversation.

What this means at smaller scale

Most SourceIT customers are not building a training campus. They are adding one to four GPU servers to an existing room. The risks there are more ordinary, and more common:

Common mistake Why it bites with GPUs
📉 Sizing on average draw A GPU server under load sits near its peak for hours. Size on the configured maximum.
🔌 Adding it to the existing UPS One integrator puts an 8-GPU RTX PRO 6000 server at roughly 5,400 to 6,700 W. That can exceed everything else in a small room combined.
⏱️ Asking for long runtime Hours of battery for a multi-kilowatt GPU load is costly. Enough time to save work and shut down cleanly is usually the sensible target.
🌡️ Forgetting the cooling If the UPS keeps the server alive but the air conditioning is off, the room overheats in minutes.

Source for the server figure: VRLA Tech: 8-GPU Blackwell server facility requirements

Six questions for your UPS supplier

  1. Has this model been tested with a pulsing or step load, and at what swing?
  2. What is the rated output in watts, not just VA, and what is my configured peak?
  3. How does it behave on generator power when the load is swinging?
  4. Is the battery expected to cycle in normal use, and what does that do to its life?
  5. What runtime do I get at my real load, and is that enough for an orderly shutdown?
  6. Can it signal the servers to shut down automatically?

The slightly controversial view

"AI-ready UPS" is becoming a label, and labels are cheap. The underlying issue is real, and the published tests are useful. But no label replaces the questions above, and for a small deployment a correctly sized conventional online UPS with honest headroom is still the answer. The specialised features matter when AI is most of the load in the building.

Related on SourceIT

SourceIT Approved

SourceIT Approved

Researched and verified by the SourceIT team for Singapore buyers.

Share This Article
Facebook

More from our blog


← Older Post Newer Post →