Battery cabinets are becoming the awkward guest in the AI data-center boom: absolutely necessary, expensive to install, and nobody wants to discover their failure behavior during a real emergency. Ampace’s newly presented PU100 large-scale fire-test results deserve attention for that reason. This is not a magic safety stamp, and the company says the findings apply only to the tested configuration and conditions. Still, the numbers provide the kind of physical evidence that operators, insurers, engineers, and investors ought to prefer over polished assurances.
Why This Test Carries More Weight Than a Brochure
Ampace received a UL certificate at RE+ 2026 after testing its PU100 UPS battery system under UL 9540A, Sixth Edition. The point of a large-scale fire test is not to prove a battery can never fail. Batteries can fail; pretending otherwise is how bad procurement decisions get made. The practical question is whether a failure stays contained, how much heat reaches nearby structures, whether adjacent equipment becomes involved, and whether the system reignites after the visible event appears finished.
That is particularly relevant as AI data centers pack more electrical infrastructure into finite real estate. Operators want backup power without handing away half the building to safety spacing. At the same time, compact deployment raises the stakes. A battery system that performs sensibly only in a generous, laboratory-like layout is not much help to a facility team wrestling with actual floor plans and hard construction budgets.
Propagation Control Is the Result That Matters Most
The headline result is that thermal runaway did not propagate from the initiating module into adjacent modules, despite simultaneous heating of 50% of the cells in that initiating module. Ampace said adjacent modules showed normal post-test voltage readings of 53.2 V, while the adjacent cabinet retained mechanical integrity and remained operational during inspection.
That is a meaningful outcome because cascade risk is the ugly part of battery-fire discussions. One damaged module is bad enough. Multiple cabinets feeding one another turns a contained equipment failure into a facility-level problem. The test involved 10 cabinets at 100% state of charge, set with zero side-by-side clearance, 20 mm back-to-back spacing, and 850 mm face-to-face spacing. No fire spread to adjacent or target cabinets was observed.
- Wall temperature rise stayed below 80°C, versus an assessed 97°C limit.
- Ceiling steel-beam temperature remained below 250°C, against an assessed 538°C limit.
- The event self-extinguished after 43 minutes with sprinklers armed but not activated.
- No re-ignition was observed during the following 24 hours.
Margins Need Context, Not Victory Laps
The wall result gives a margin of more than 17°C, while the steel-beam result carries more than 288°C of room versus the assessed limit. The larger structural margin is reassuring. The wall margin is also below the stated threshold, but it is not enormous enough to justify casual thinking. Construction materials, room geometry, ventilation, installation quality, fire detection, and the surrounding equipment mix can all alter a real-world risk picture. That is why Ampace’s own caveat matters: this test supports project-specific engineering assessment; it does not replace it.
Measured fire behavior is more useful than broad safety language, but one successful tested setup is not a universal permission slip.
The post-fire portion may be just as valuable as the dramatic flames-and-heat portion. Ampace reported no flaming debris escaped the cabinet, exceeding the assessed criterion of no flaming debris beyond 1.5 meters. It also reported no observed reignition for 24 hours following self-extinguishment. Recovery planning is where operators earn their keep. If an event is limited, nearby assets remain viable, and the site does not face a delayed second incident, the operational and financial damage can look very different.
What Sector Watchers Should Take From It
Ampace, established in 2021, says it has supplied battery packs and cells at scale across several markets and has recorded zero safety incidents in company records. That claim should be read exactly as stated: it is the company’s record, not an outside audit described in these notes. Even so, the PU100 test gives prospective customers a much firmer basis for comparison than a generic reliability claim.
For the broader data-center supply chain, this is another sign that battery safety is moving from a compliance checkbox into a design and capital-allocation issue. Buyers should ask for the exact tested arrangement, state of charge, spacing, suppression assumptions, thermal measurements, and post-event observations. Ampace has put several of those details on the table here, which is a good start. The sensible next step is seeing how closely a proposed installation matches them. In this business, the gap between “tested” and “installed” is where trouble usually hides.