News

June 2026

AI is turning the data center battery into a power product

AI is turning the data center battery into a power product

For decades, the data center battery had one dominant selection question in classic UPS (uninterruptible power supply) duty: how many minutes it could support the load before the generator or another backup source took over.

Engineers sized it by load, judged it by runtime, and swapped it on a schedule. AI is changing what the battery has to do, and the spec sheet that selects it is changing with the work.

The load changed first

A traditional data center rack often draws 5 to 10 kW, and its load stays steady enough that backup design could focus more on ride-through than on repeated load movement.

AI racks break that assumption.

High-density AI racks now exceed 50 kW, some rack-scale systems are already above 100 kW, and the load no longer sits still. One industry estimate, from power-protection supplier Capacitech, puts AI processing surges at 30 to 50% above baseline lasting 0.2 to 2 seconds.

Joint research cited by NVIDIA and published by Microsoft describes large AI training jobs spanning tens of thousands of GPUs, moving through compute-heavy and communication-heavy phases in a synchronous pattern.

Those swings can synchronize across the cluster and become large enough to create a facility-level power-quality problem. A cell tuned for long, slow energy storage is not typically built to absorb that kind of repeated, high-rate duty.

The old UPS question is too narrow

Buyers used to judge a backup battery by one number: how many minutes it could run.

Runtime still matters, but it no longer settles the decision on its own. The duty now asks how quickly the battery responds, how hard it can discharge, how many high-rate cycles it can survive, how it behaves thermally under that load, and how much power it delivers per square meter of floor space.

A 2026 Endeavor Business Intelligence survey of 150 data center professionals tracks the shift, with 52% naming AI power swings and power quality as a top impact on their infrastructure, up from 37% a year earlier, and 57% saying AI workloads push them toward higher density in a smaller footprint.

Two-thirds say a battery that smooths those swings would help.

Two battery layers are emerging

The pressure appears at two different timescales, and each needs its own product rather than a resized version of the same cell.

Close to the rack, in roughly the 200 millisecond band, the job is ultra-high-power buffering, where a battery sidecar buffers the fastest transients before too much of that movement reaches the wider power architecture.

Further back, in the power room, the job runs for about 60 seconds to 5 minutes, with a high-power battery rack carrying the load through disturbances and supporting the UPS.

They are different products. The sidecar is engineered for sub-second power delivery, the power-room rack for short-duration ride-through, so cell design, pack design, thermal design, and controls have to match the duty.

Why standard energy-storage LFP is not enough

A grid-scale BESS (battery energy storage system) is usually designed around energy capacity, cost per kilowatt-hour, degradation over cycles, and longer discharge windows.

When it uses standard grid-storage LFP (lithium iron phosphate) cells, they are built to meet that duty. That design can win grid-storage projects, but it answers the wrong question for AI data center duty.

This work needs high-power LFP built for repeated shallow cycling, fast response, thermal stability under sustained high-rate pulses, and tight integration with the power electronics around it. Price per kilowatt-hour, the metric that ruled grid-storage buying, says almost nothing about which cell can tolerate that pulse load.

What buyers should ask now

The selection question is expanding beyond “how many minutes?” to a harder set of questions.

How fast does the battery respond, how much power can it deliver, how many shallow high-rate cycles can it tolerate, how much floor space does it take, how does it behave thermally under that duty, and how cleanly does it integrate with the UPS, power electronics, and controls.

Buyers who get this right can match the battery to the timescale imposed by the site, rather than buying a single system and hoping it covers every job.

See us at Intersolar

ElevenEs will be at Intersolar Europe in Munich next week, presenting high-power battery solutions built for AI data center duty cycles. Come find us there at booth C3.115.

Share this article