Data centre power: sizing UPS for an AI training rack

Key Takeaways

• AI GPU racks now routinely draw 40-100 kW, against 5-10 kW for a conventional server rack, a 5-16x density increase that invalidates power-sizing assumptions built for traditional infrastructure.
• A 60 kW rack on a 208V three-phase supply draws roughly 166A per phase, well beyond a standard PDU's rating, requiring 3-phase rack PDUs specifically rated for 60-100A per phase.
• GPU training workloads create synchronized power swings from peak to idle within milliseconds, during checkpointing, communication, and start/stop events across thousands of GPUs acting in lockstep, a load profile UPS systems built for steady-state IT loads weren't designed around.
• Facility-level demand for AI-focused data centres now runs 100-750 MW per site, and rack-level density is projected to climb from roughly 50 kW toward 1 MW as next-generation GPU architectures scale.

Sizing a UPS for an AI training rack by taking its rated power draw and adding a standard safety margin misses the part of the load profile that actually causes problems: AI training doesn't draw power steadily, it swings hard and fast, in a pattern conventional IT load planning never had to account for.

The density jump that breaks old assumptions

Racks running NVIDIA H100 and H200 GPU clusters now routinely require 40-100 kW, compared to 5-10 kW for a traditional server rack, and a single NVIDIA GB200 NVL72 rack draws 120-140 kW (GPUSmith, GPU rack power requirements planning guide, retrieved 2026-09-10). That's a 5-16x density increase over conventional infrastructure planning baselines, which means UPS and distribution sizing rules calibrated for older server generations don't transfer to an AI training deployment without being rebuilt from the actual load, not scaled from the old baseline.

The distribution problem before the UPS problem

Before the UPS sizing question, there's a distribution question: a 60 kW rack on a standard 208V three-phase supply draws approximately 166A per phase, well beyond what a standard PDU handles (GPUSmith, retrieved 2026-09-10). High-density AI deployments require 3-phase rack PDUs rated for 60-100A per phase with per-outlet intelligent metering, and specifying a UPS correctly matters little if the distribution infrastructure feeding the rack can't carry the current in the first place. Model the full rack's projected current draw through the data centre power planning tool before finalising a UPS specification, since the PDU and distribution sizing constrains what the UPS actually needs to deliver.

The swing that steady-state sizing rules miss

The characteristic that most distinguishes AI training load from conventional IT load is speed of change, not just magnitude. Thousands of GPUs operating in lockstep cause power consumption to swing from peak to idle within milliseconds during synchronised communication, checkpointing, and startup or shutdown events (GPUSmith, retrieved 2026-09-10). A UPS sized purely on average or even peak steady-state draw, without accounting for how fast the load actually moves between those states, can be adequately rated on paper and still struggle with the transient response the workload actually demands. This is the AI-training-specific version of a more general UPS sizing failure mode: sizing for the number on the spec sheet rather than the load's actual behaviour over time.

Sizing for where the roadmap is heading, not just today's hardware

Facility-level power demand for AI-focused data centres now runs 100-750 MW per site, and rack-level density is projected to climb from roughly 50 kW toward 1 MW as next-generation GPU architectures (Rubin, Kyber-class) scale over the next several years (GPUSmith, retrieved 2026-09-10). A UPS and distribution architecture sized precisely to today's rack density, with no headroom for the next hardware generation, is a design that gets stranded within a single refresh cycle. Building in headroom against the density roadmap, not just the density of the hardware being commissioned this quarter, is the more expensive choice upfront and the cheaper one over the facility's operating life. That's exactly the planning gap an AI data centre power management solution is built to close, sizing distribution and backup capacity against the multiyear density curve rather than against a single hardware generation.

Frequently asked questions

Can a standard data centre UPS handle an AI training rack's load swings?

Not reliably if it was specified using conventional steady-state sizing rules. AI training load's millisecond-scale swings between peak and idle require UPS and power architecture specifically evaluated against that transient behaviour, not just against average or peak wattage figures.

Is the PDU rating or the UPS rating the bigger constraint for a high-density AI rack?

Both matter, but distribution (PDU and the circuit feeding it) is often the more immediate constraint, since a 60+ kW rack's current draw can exceed standard PDU ratings before the UPS sizing question is even reached. Both need to be sized together against the actual rack load, not independently.

How much headroom should a new AI data centre build plan for future rack density increases?

There's no universal figure, but given the trend from roughly 50 kW toward 1 MW per rack over the coming hardware generations, planning distribution and cooling infrastructure with meaningful multiyear headroom, rather than matching exactly to the first hardware generation installed, avoids a costly mid-life retrofit.

The bottom line

Sizing a UPS for an AI training rack is not the same exercise as sizing one for conventional IT load, scaled up. The density is higher, the distribution requirements are different, and the load itself moves in a way steady-state sizing rules were never built to anticipate. Plan the distribution and UPS together, against the actual transient load profile and the multiyear density roadmap, not against the nameplate rating of the hardware being installed today.

Figures were verified on 10 September 2026 against published data centre and GPU rack power planning research. AI hardware power specifications and density trends are evolving rapidly; confirm current figures against the specific GPU generation being deployed before finalising a power architecture.