When the battery unit is the whole building

EPRI's new battery-failure paper made headlines last week for a 98 percent drop in the failure rate since 2018. That number is real. But the same paper has a finding most coverage gave one sentence, and it explains why one California fire destroyed more than half a battery plant while a fire at the plant next door took out less than one percent.

The paper is "Risk Factors and Failure Statistics of Lithium Ion BESS," an EPRI white paper dated October 2026, built on EPRI's own BESS Failure Incident Database. pv magazine USA covered it on 14 October under the headline "Battery energy storage failures are down 98% since 2018 and continue to decline," and the indoor finding appears there as a quoted line about fire walls. That is fair as far as it goes. The data behind that line is worth a closer look, because it is the part that tells you how to build one.

First, the headline number, briefly

EPRI counts a failure as a large thermal event (arc flash, fire or explosion) caused by something inside the battery system, at utility-scale or commercial and industrial sites. The paper reports that the annual failure rate per deployed gigawatt-hour fell 98 percent between 2018 and 2025. In 2024 it counted 26 incidents in a global fleet of 343 GWh; in 2025, 24 incidents in 634 GWh. Notice that the incident count barely moved. The rate halved mostly because the fleet nearly doubled. That is how a per-GWh rate works, and it is still good news, but it is a statement about exposure, not about any single site getting safer. (Report 083 went through what that rate is made of, using EPRI's 2024 edition, which put the drop at 97 percent.)

EPRI is also candid about the gaps. It lists 33 "non-public" incidents, confirmed by trusted safety experts but never reported in the media, which make up 22 percent of the total. It notes there is no mandatory failure reporting in the U.S. or Europe, and that 22 of the 25 Chinese incidents in the set are non-public. In the paper's words, "Limited transparency and data gaps remain a major challenge to a complete database."

The unit is what burns

Here is the idea the rest of the paper hangs on. In an outdoor containerized plant, the smallest "unit" is one container. In a building-based plant, the smallest unit is the building itself, unless someone has installed fire barriers inside it. When a fire starts, EPRI finds it usually consumes at least one whole unit. So the size of the unit sets the size of the loss.

The paper puts two incidents side by side, at neighboring plants in Moss Landing, California, and they are about as close to a controlled comparison as real accidents get.

Elkhorn, September 2022: an outdoor system of 730 MWh. One container caught fire and was allowed to burn itself out while firefighters cooled the containers around it. About 3 MWh was damaged, 0.4 percent of the system, and the incident lasted roughly 24 hours.

Vistra Moss Landing, January 2025: an indoor system of 1,200 MWh, with over 1 GWh of capacity installed in a single building. About 660 MWh was damaged, 55 percent of the installed capacity, and the incident lasted several days. EPRI calls its magnitude anomalous, driven by the system's size and "the large proportion of damage due to the indoor architecture."

Similar scale, neighboring sites, more than a hundredfold difference in the share that burned. EPRI's summary of the pair: "the architecture limited the damage to less than 1% of the system's energy capacity."

It isn't just one pair

Two incidents can be a coincidence, so EPRI ran two broader checks.

How often indoor sites fail. Using U.S. data only (the only country where EPRI had architecture-specific fleet data, from the Energy Information Administration, covering systems of at least 1 MW), it compared the share of indoor sites among failures to their share of the fleet. Indoor sites were overrepresented. The paper is careful about the limits: of 36 U.S. incidents, 30 could be placed in the comparison, 4 of them with unknown architecture; the fleet data is a 2024 snapshot while failures span 2018 to 2025. But it argues the errors can only push the indoor fleet count too high, not too low, so if anything the finding is conservative.

How much burns when they do. Across 68 global incidents with enough information (45 percent of the 150 in the database), outdoor systems typically lost under 10 MWh regardless of how big the plant was. Indoor losses were spread widely. Leaving out systems of one container or less, where nothing could spread anyway, the median share of capacity damaged was 100 percent for indoor systems and 6 percent for outdoor ones.

The paper also shows the fix works. Two indoor incidents lost under a quarter of their capacity, at Otay Mesa (2024) and Yeongam (2022), and in both the plant was split across multiple buildings, so separation stopped the fire short of the whole site.

Outdoor isn't automatically safe

The reverse case matters too. Only 11 of the 68 incidents spread from one unit to another, and 10 of those 11 were outdoor systems. EPRI notes that nine of them were installed before 2024, and half before 2022, "when propagation mitigation designs and testing were more limited." Spacing and wall design decide whether a container fire stays a container fire, and the paper could not separate those details in the data.

That connects to the standards. EPRI writes that the latest UL 9540A and NFPA 855 requirements mandate large-scale fire testing that specifically considers unit-to-unit propagation, and that these updates arrived in the 2025 to 2026 cycle, "so older systems may not have been evaluated against this criterion." (Report 101 covers what a UL 9540A test is and isn't.) In practice, a site's vintage tells you which questions were asked of it.

EPRI also reports that the U.S. industry has converged on outdoor containers, with dedicated battery buildings becoming increasingly uncommon. So the architecture problem is partly a legacy problem, though the paper flags that future batteries inside substation buildings or data centers may not show up in the fleet data at all if they are under 1 MW.

What it does not tell you about the battery in your garage

Be careful carrying this home, literally. EPRI's database does not track residential batteries or electric vehicles at all, so none of these rates or medians describe a home system. The paper's "indoor" means mainly dedicated battery buildings, and it says plainly that risks and emergency response "change significantly for batteries installed in human-occupied buildings," where measures "should go beyond fire barriers and consider life-safety impacts."

What does carry over is the way of thinking. For any battery, ask what the smallest unit is that a fire is expected to consume, and what stops it there. For a grid plant, that is container spacing or interior fire walls. For a home, it is the product's own enclosure, the installer's clearances and the local code, which this paper does not evaluate. (Report 138 covers where a home battery is allowed to go.)

One more finding worth knowing, because it cuts against a popular talking point: EPRI says that because chemistry was hard to separate from improvements in system design and operation, "it could not be concluded that cell-level chemistry in isolation was a risk factor" for system-level failures. Architecture and vintage showed up in the data. Chemistry on its own did not.

The signal

A failure rate tells you how often. Architecture tells you how much. The 98 percent drop is a fleet-level fact about exposure, while the Moss Landing and Elkhorn comparison is a design-level fact about consequences, and only the second is something a person siting a battery can act on. If I were reviewing a site, I would ask how big the unit is before I asked anything about the cells.

I care about this because I help design the AI battery-cycling systems for a veteran-owned (HUBZone) energy-storage integrator, run by fellow veterans. I do not own that company and earn nothing from this link. Full policy here. Nothing in this report involves them or any system they have built.

Sources

EPRI (L. Srinivasan, E. Billaut, S. Shaw, principal investigators), "Risk Factors and Failure Statistics of Lithium Ion BESS," EPRI White Paper 3002034920, Palo Alto, CA, October 2026. (Primary source, full 85-page PDF downloaded and text-extracted; all quotes string-matched. Source of the failure definition and database scope, including the exclusion of residential and EV systems; the 98% rate decline; the 2024 and 2025 incident and GWh counts; the 33 non-public incidents (22%), the reporting gaps and the China figure; the unit concept; the U.S. indoor-versus-outdoor comparison and its stated limits; the 68-of-150 damage analysis, the under-10 MWh and 100% / 6% median findings; Otay Mesa and Yeongam; the 11-of-68 propagation finding and installation years; Table 13 comparing Moss Landing and Elkhorn; the UL 9540A / NFPA 855 note; the human-occupied buildings note; the trend away from dedicated buildings; and the chemistry conclusion. Figures 16 and 17 are images and their exact proportions are not quoted here.)
William Driscoll, "Battery energy storage failures are down 98% since 2018 and continue to decline," pv magazine USA, 14 October 2026. (Opened. Coverage, showing how the paper was presented; it reports the indoor finding and the fire-wall recommendation but not the Moss Landing / Elkhorn comparison. No figure here is taken from it.)
Prior reporting in this publication: Report 083, the battery fire rate fell 97 percent; Report 180, why a burned battery can reignite; Report 101, there is no such thing as UL 9540A certified; Report 138, where a home battery can go. (Context only.)