TL;DR:

  • AI workloads are pushing rack power densities from roughly 12 kW in 2024 to 26–27 kW today, with some AI racks already running 50-100+ kW.
  • Unlike steady-state IT loads, AI training and inference create rapid, repeated power swings – not just backup events.
  • Traditional backup batteries and capacitor-based systems each address different requirements: one prioritizes runtime, while the other prioritizes rapid response.
  • Rack-level energy storage is emerging as its own design category, distinct from centralized UPS systems.
  • Battery chemistry choices at the rack are now a safety and reliability variable, not just a capacity spec.

# # #

For most of the last decade, data center power design centered on one question: How long can the UPS ride through an outage? That question hasn’t gone away, but now it has company. ZincFive®, known for nickel-zinc (NiZn) battery technology in UPS systems, announced an expansion into in-rack energy storage with its new NiZn In-Rack Power Solutions, which includes dedicated Battery Backup Unit (BBU) and AI Dynamic Power Module (DPM) configurations built for AI’s faster, more dynamic power draw. The move points to a broader shift underway across the industry: Rack-level power is becoming its own design discipline, separate from the UPS layer.

Average rack densities have more than doubled in two years, climbing from around 12 kW in 2024 to roughly 26-27 kW in 2026, according to recent industry benchmarking. Some AI training racks already exceed 50 kW, and next-generation GPU architectures are being designed around 100 kW or more per rack. That’s the headline most operators already know.

What’s less understood is how that power behaves. AI training and inference don’t draw a steady current; they spike and drop in fast, repeated bursts as GPU clusters ramp up, pause, and resume. These transient loads can stress busbars, power supplies, and upstream infrastructure in ways a traditional steady-state design never had to account for.

Backup Runtime and Rapid Response Are Different Problems

This is where the engineering gets interesting. Traditional battery backup units are built to provide runtime, with enough stored energy to ride through a utility disturbance for seconds to minutes. Capacitor-based systems, by contrast, respond in milliseconds but hold very little energy, so they smooth short spikes without offering meaningful backup duration.

Most facilities have defaulted to choosing one or blending both, accepting the tradeoff. A capacitor bank might absorb a millisecond-scale current step from a GPU cluster starting up, but it can’t cover a multi-second grid disturbance. A traditional BBU can cover that disturbance but may not be designed to manage rapid, repeated load transitions.

“AI is changing not only how much power data centers require, but how that power is consumed,” said Tod Higinbotham, CEO of ZincFive. “As power densities increase and workloads become more dynamic, the industry needs new approaches that can deliver rapid power response without compromising safety, reliability, or flexibility.”

As racks get denser and workloads more dynamic, that tradeoff becomes harder to accept. Operators need infrastructure that can do both jobs without over-provisioning space or cooling capacity at the rack, which is already at a premium in AI deployments.

Why Battery Chemistry Matters at the Rack Level

As energy storage moves physically closer to compute, the chemistry inside that storage becomes a bigger factor in facility risk management. Lithium-ion remains dominant across the industry, but proximity to high-density, high-heat compute racks raises the stakes around thermal behavior.

Nickel-zinc is one proven chemistry gaining attention specifically for this reason, already validated at scale in UPS-level deployments before moving closer to the rack. It supports high-rate charge and discharge capabilities, which suit repeated transient events, while carrying no risk of thermal runaway. For rack-level deployments, where space constraints can limit fire suppression and containment options, that safety profile is becoming as relevant as energy density or cycle life.

What This Means for AI Infrastructure Planning

The practical takeaway for facility planners and OEMs is that rack-level power is no longer a single-spec decision. It requires evaluating backup runtime, transient response speed, thermal safety, and physical footprint together, not in isolation.

ZincFive is now developing dedicated in-rack solutions – BBU configurations for established backup applications and AI DPM configurations for transient and pulse power mitigation – rather than a single configuration meant to address every requirement. That approach reflects a broader industry shift: As AI reshapes how power is consumed, the tools used to manage it are having to specialize rather than generalize.

For an industry accustomed to sizing infrastructure around steady-state assumptions, that’s a meaningful adjustment, and one that’s likely to define rack design conversations as AI infrastructure continues to evolve.

To read the full press release, visit here.