Choosing the Right Batteries for AI Facilities - Lorbel

Choosing the Right Batteries for AI Facilities

AI facilities run twenty-four hours a day, seven days a week. They don’t sleep, take breaks, or pause for maintenance, making reliable power more than essential. You also have your servers running non-stop, GPUs maxed out, and fans and cooling systems working overtime just to keep up.

In situations like this, a backup power supply can’t be an afterthought. It needs to be built for an infrastructure that never really rests. Battery selection may seem like a small detail when building this kind of facility, but overlooking it can create major challenges down the line.

The wrong battery choice can lead to problems like shorter runtimes, monthly replacements, and downtime you just can’t afford. But why does this happen so often? Many engineers make this mistake because of the misconception that traditional data centers have the same power requirements as AI facilities.

Here’s what makes them different:

• Higher power density: GPU clusters pull much more current per rack than standard servers.
• Less tolerance for interruption: A dropped training run can cost days of compute time, not just a few minutes of inconvenience.
• Constant load, not peak load: AI hardware often runs at full tilt for hours or days straight, so batteries get less of a break between demand spikes.

Standard UPS setups built for general IT loads just weren’t designed with this in mind. That’s why battery choice deserves its own conversation, as older designs may not account for today’s AI environments.

The Main Contenders

Most facilities choose between a few battery types. Each comes with tradeoffs.

1. VRLA (Valve-Regulated Lead-Acid): Common and affordable. VRLA batteries are a known quantity, easy to source, and simple to maintain. The downside is that they’re bulkier, heavier, and typically only last three to five years before needing replacement.
2. Lithium-ion: A lot of AI facilities are heading in the lithium-ion direction. Lithium-ion batteries pack more energy into a smaller footprint, handle higher temperatures better, and often last two to three times longer than VRLA. They cost more upfront, but the lifecycle savings can even things out.
3. Nickel-based options: Less common, but worth considering for extreme environments. These handle temperature swings well and have a long service life, though cost and availability can be limiting factors.

There’s no universal winner here. The right pick depends on your space, your budget, and how much runtime you need.

What Actually Matters When You’re Deciding

Rather than evaluating specifications alone, focus on these factors first:

• Runtime requirements: How long does your facility need to stay powered during an outage before generators kick in or systems can be shut down safely?
• Footprint: Space is expensive, especially in facilities packed with racks. Smaller battery footprints free up room for what actually generates revenue.
• Thermal performance: AI facilities run hot, so your battery choice needs to handle that heat without degrading faster than expected.
• Total cost of ownership: The upfront price is one number. Replacement cycles, maintenance, and cooling costs are the ones that sneak up on you.
• Scalability: If your compute needs are growing, your power backup should be able to grow with it, not force a full redesign in two years.

Maintenance Isn’t Optional

Even the best battery underperforms without regular checkups. Voltage checks, temperature monitoring, and capacity testing all catch problems before they become outages. This matters even more with lithium-ion systems, where battery management systems need to be monitored, not just installed and forgotten.

A facility running 24/7 AI workloads needs a maintenance plan that matches that intensity. Quarterly checks might have worked for a lighter load, but may not provide enough protection for these new AI environments.

Getting It Right the First Time

Battery selection is the difference between a facility that handles a power event without anyone noticing and one that goes dark at the worst possible moment.

Lorbel works with engineers and facility managers across California and neighboring states to design power protection that actually fits the workload, not just the budget. From installation through ongoing service and maintenance, the goal is to keep critical systems running when it counts.

If your AI facility’s power backup hasn’t been reevaluated recently, now’s a good time to review whether it still meets your operational needs.