1. Data centers and storage in brief

• The traditional job: the battery in a UPS (uninterruptible power supply) takes the load at the moment of an outage or voltage disturbance and carries it for the short time until a standby generator starts. NERC (the North American Electric Reliability Corporation) describes the UPS battery bank as designed not to carry the load for long, but for the brief disturbance, or, in a complete outage, long enough to start a backup generator Sourced [Ref. 1].
• The changing job: AI training load swings sharply because many GPUs ramp up and down together. NVIDIA has published a design with capacitors in the rack power shelf, and a facility-level battery system (BESS) to condition the boundary with the grid Sourced (company claim) [Refs. 6 and 8]. Meanwhile grid operators treat data centers switching to UPS and vanishing from the grid all at once as a reliability risk Sourced [Refs. 1 and 2].
• What this asks of the battery: full output within minutes (high power), a long life while sitting near full charge (long life), and safe installation inside or beside a building (safety), all at once. In our calculation, for a 5 MW UPS with a short hold time, the number of battery racks is set by power, not energy (Chapter 5).

2. What a UPS does, and how it is built

The UPS standard IEC 62040-3:2021 sets performance and test requirements for equipment that takes AC input up to 1,000 V, delivers AC output up to 1,000 V, includes an energy-storage device up to 1,500 V DC, and has continuity of load power as its primary function Sourced [Ref. 12]. The battery is that "energy-storage device".

View the original SVG schematic and exact annotations

In its review of the July 2024 incident, NERC sorted data-center UPS systems into three designs Sourced [Ref. 1].

All Sourced (a summary of the NERC incident review [Ref. 1]).

What matters here is that a UPS moves the load to the battery even for a voltage dip of a few tens of milliseconds. In NERC's review the voltage dips lasted between 42 and 66 milliseconds Sourced [Ref. 1]. With a static UPS the load returns to the grid once the disturbance clears, whereas with a DRUPS the load typically has to be transferred back manually, the review notes. This behavior leads to the grid-side problem described in Chapters 6 and 8.

3. From lead-acid to lithium-ion

UPS batteries in data centers were for a long time valve-regulated lead-acid (VRLA) or flooded lead-acid. In 2019 the chief technical officer of the Uptime Institute listed the advantages of lithium-ion as higher power density, so more power from a smaller footprint, and more charge-discharge cycles without degrading. Lead-acid, he wrote, must be replaced when it reaches about 80% of its original capacity because capacity then falls off quickly, while lithium-ion loses capacity gradually and predictably. He also noted that suppliers claim a lower total cost of ownership despite the higher initial cost, but that not all such claims can be substantiated with long-term data Sourced (expert opinion, 2019) [Ref. 10].

On adoption, a January 2023 Uptime Institute article says that in its Global Data Center Survey 2021 nearly half of operators had adopted lithium-ion for centralized UPS plants, up from about a quarter three years earlier, and that the 2022 survey showed a further rise Sourced (survey responses) [Ref. 11]. The respondent pool and the definition of "adopted" are not given, so we do not use the figure as a market share.

All Sourced (expert opinion and survey statements). Because the conditions are not given, we make no judgement on cost.

4. What the battery has to deliver: power, life, safety

Here is one real specification. Samsung SDI's end-customer product specification for its lithium-ion battery system for UPS (U6A4) describes a rack built from 17 modules of 8 series-connected 67 Ah cells (136 cells in series). However, the document's revision history shows only one entry, "2025.04.28, First draft (0.0)", and we could not confirm that a final version exists Sourced (manufacturer's stated values) [Ref. 9].

All Sourced (Samsung SDI product specification, first draft [Ref. 9]). Cycle life, warranty period and cell chemistry could not be confirmed from this document.

From these numbers we calculated the following Our calculation.

• 183.6 kW continuous ÷ 34.6 kWh nominal = about 5.3C equivalent. Time to use all the energy: 34.6 ÷ 183.6 × 60 = about 11.3 minutes.
• 450 A ÷ 67 Ah = about 6.7C (peak current rate per cell).
• These rates are relative to the nominal capacity defined at 1/3C; at high rates the capacity that can be extracted can be smaller than nominal (Commentary). The specification does not quantify the difference.

What the standards look at

All Sourced (official pages of the standards bodies and certifier). The standards' texts are paid, so we give only titles and scope.

5. Our calculation 1: how many battery racks does a 5 MW UPS need?

Taking 5 MW, the upper end of the centralized UPS range NERC cites (2–5 MW), we count how many racks of the specification above would be needed Our calculation.

All Our calculation. Racks for power = 5,000 ÷ 0.95 ÷ 183.6 = 28.7, rounded up to 29. Racks for energy = nameplate ÷ 34.6, rounded up.

The result: at a 5-minute hold, energy alone would need 16 racks, but power requires 29. At 10 minutes the two roughly balance, and at 15 minutes energy dominates at 48 racks Our calculation. In other words, for holds of a few minutes the battery is sized by buying power, not capacity, and about half the stored energy is never used.

These assumptions (95% efficiency, 80% at EOL, 183.6 kW per rack) are not a design. In practice internal resistance rises toward end of life, so deliverable power falls and more racks are needed, and redundancy adds more. The calculation shows what sets the count (power or energy), not an actual count.

6. AI load swings and storage

(1) What is swinging

A paper by NVIDIA, Microsoft and OpenAI (arXiv, August 2025) explains that AI training spans tens of thousands of GPUs and that in each iteration compute-heavy and communication-heavy phases alternate in lockstep, so power swings widely. It adds that if the frequency of the swings coincides with critical frequencies of utilities, physical damage to grid infrastructure can result Sourced (from the abstract) [Ref. 7]. NVIDIA's technical blog says that where ordinary data-center workloads are asynchronous and average out, in AI training many GPUs move in step, so grid-level power mirrors and amplifies the pattern of a single node Sourced (company claim) [Ref. 6].

(2) Which layer takes which swing

The NVIDIA GB300 NVL72 smooths with three mechanisms: a power cap on ramp-up, short-term storage in electrolytic capacitors, and a "burner" that dissipates power in the GPUs during ramp-down. About half the volume of the power supply is capacitors, with 65 J per GPU of storage, optimized with power-supply maker LITEON. From measurements on a GB200 rack power shelf, NVIDIA reports that the peak power seen by the grid fell 30% when training Megatron LLM Sourced (company claim, NVIDIA's measurement) [Ref. 6].

65 J covers a 65 W difference for one second Our calculation (65 J = 65 W × 1 s). So storage in the rack is for swings of a second or less up to a few seconds, a different time layer from the batteries that bridge outages (minutes and up).

In a June 2026 blog, NVIDIA says facility-level BESS takes the residual transients that rack-level smoothing leaves, and protects the generators and the grid interface. It also stresses that BESS is a control system that exchanges power with the grid through power converters and controls, not just a store Sourced (company claim) [Ref. 8]. The same blog says that existing interconnection standards do not yet cover the behaviors AI factories need (load smoothing, switching between grid-connected and islanded modes, coordination with on-site generation), and it presents its own self-qualification guidelines Sourced (company claim) [Ref. 8]. That the standards lag is NVIDIA's view, not that of a standards body.

7. Our calculation 2: how many cycles a year does smoothing add?

When a battery smooths load swings, it is used in a completely different way from one full charge-discharge a day. We computed the scale for a hypothetical load Our calculation. It is not the shape, period or amplitude of any real AI training load.

All Our calculation (hypothetical load). Equivalent full cycles = total discharged energy ÷ battery capacity.

Three things stand out Our calculation.

• The count is orders of magnitude higher. One cycle a day is 365 a year. Using the same battery to smooth swings gives thousands to over ten thousand equivalent full cycles a year (4,380 to 17,520 in this example).
• Vast numbers of shallow cycles. DoD per period is under 1%. How shallow cycling affects ageing depends on the materials and conditions, and we do not generalize.
• Heat. At an assumed 90% round-trip efficiency, daily charging is about 267 MWh (240 ÷ 0.9), of which about 27 MWh becomes heat, or about 1.1 MW on average (assumed round-trip efficiency: 90%). Cooling must be designed for that amount.

8. Backup power, grid connection and demand response: examples, and who is making each claim

(1) The grid's problem: when the UPS works as designed, the load vanishes at once

NERC's incident review (published 8 January 2025) reports an event in the U.S. Eastern Interconnection at about 7 p.m. on 10 July 2024 Sourced [Ref. 1].

• A lightning arrester failed on a 230 kV line, causing a permanent fault. The line made three automatic reclosing attempts, producing six successive faults (voltage dips) within 82 seconds, each lasting 42–66 ms.
• At the same time, about 1,500 MW of load disappeared. It was not disconnected by utility equipment but by customers' own protection and controls, and all of it was data-center-type load.
• The data centers had transferred their load to backup power in response to the disturbance. Many used a scheme that counts voltage dips within a set time and stays off the grid until manually reconnected (typically three dips within one minute); about 1,260 MW dropped at the third dip and stayed off for hours.
• Frequency rose to a maximum of 60.047 Hz and returned to 60.0 Hz in about four minutes. Voltage peaked at 1.07 per unit, and operators removed shunt capacitor banks to bring it back.

NERC's September 2026 document describes what followed Sourced [Ref. 2]. Some data centers changed the time over which voltage dips are "counted" and the number that triggers disconnection. On 18 April 2026, a fault with similar characteristics in the same area led to only about 500 MW of sustained load loss. In July 2026, however, Virginia saw a customer-initiated load reduction of nearly 4 GW, and NERC says the risk "is increasing, not abating" (that event was still being analyzed as of September 2026). At the direction of FERC, NERC plans to finish registry criteria and foundational Reliability Standards for large computational loads by the end of 2026 Not yet confirmed.

(2) The examples: who says what, and how far

All Sourced (a summary of company announcements; these are company claims, not independently verified values).

(3) Three jobs compete for the same battery

NVIDIA lists the jobs expected of a data-center BESS as transient smoothing, keeping reserve for ride-through, and demand response or generator coordination, notes that "those missions can compete with each other", and says the design needs explicit priorities and a strategy to prevent uncontrolled state-of-charge (SOC) drift Sourced (company claim) [Ref. 8].

In the Belgian case, only 2.75 MWh of the 5.5 MWh was optimized for demand response Sourced [Ref. 4]. It is natural to read the rest as held back as backup reserve, but the announcement does not say so Inference. How to divide capacity between insurance for emergencies and everyday revenue operation is at the center of this application's design (Commentary).

• Backup: keep supplying power at the moment of an outage or disturbance. The battery must always hold reserve near full charge.
• Grid connection: raise and lower load at a speed the grid allows, and ride through. The battery needs power and fast response.
• Demand response: cut load or discharge in response to grid supply and demand. The battery must withstand repeated charge and discharge and restore the reserve used in daily operation.

Every role demands high power, long life and safety at the same time. For cost evaluation (how levelized cost is built), see our explainer on the economics of storage. For the grid-scale side, see our explainers on grid-scale storage systems and on virtual power plants and distributed energy resources.

(4) Safety: treating fire as a design premise

The Uptime Institute discusses the fire at a data center in Pangyo, South Korea, in October 2022. It says the fire started in a battery room, took about eight hours to bring under control, and took down services including KakaoTalk. At the time of the article the root cause had not been disclosed, and claims about the battery maker's battery-management-system records conflicted. We therefore do not attribute a cause to that fire. The design lessons the article draws are choice of suppression, spacing of battery cabinets, and fire-resistant compartments in battery rooms Sourced (Uptime Institute analysis) [Ref. 11].

On installation, NFPA 855 (2026 edition) sets minimum requirements for stationary energy storage systems, and UL 9540A (6th edition) defines a test method for thermal-runaway fire propagation [Refs. 14 and 15]. For batteries inside buildings, conformity to these standards and local fire regulations is a precondition. We make no judgement on any individual facility. The mechanism of thermal runaway is covered in our explainer on thermal runaway and battery safety.

9. How supercapacitors fit in

Our explainer on supercapacitors sets out that they store energy by adsorbing ions, not by chemical reaction, so they are strong on power and life but orders of magnitude lower in energy density than batteries. In the time layers of Fig. 4 this means they suit swings longer than an electrolytic capacitor can cover and shorter than a battery is efficient at (in our reading, from about one second to about a minute) (Commentary; the band in Fig. 4 is a rough guide with no evidence of deployment behind it).

• What we confirmed: the NVIDIA GB300 power shelf uses electrolytic capacitors Sourced [Ref. 6]. They are not supercapacitors.
• What we could not confirm: we could not find, in public-agency or operator primary sources, the scale or track record of supercapacitor use in data centers. A supercapacitor maker listing AI data centers among its customer segments on its own website, as covered in the supercapacitor explainer, remains a company statement.

Which storage element goes in which time layer depends on cost, volume, life and maintenance together. We do not rank any approach.

10. Limits, and what we could not confirm

• We use no market sizes. We did not use market-size or market-share estimates for UPS batteries or lithium-ion, since they come from research firms. The Uptime Institute's "adopted" share has no stated definition, so we do not treat it as a market share.
• The Samsung SDI specification is a "first draft", and we could not confirm a final version. It gives no cycle life, warranty or cell chemistry.
• Operating results (Google in Belgium, Microsoft in Dublin) are plans or equipment descriptions as of the announcements; we could not confirm results after commissioning from primary sources.
• Individual accidents such as the Pangyo fire: we do not state causes because we could not confirm a published root cause. We also could not confirm official investigation results for the 2025 fire at a South Korean government-systems data center, so we do not cover it.
• The nearly 4 GW event in Virginia in July 2026 was still under analysis at the time of NERC's document (September 2026), so the details may change.
• Our calculations (Chapters 5 and 7) are structural explanations under stated assumptions. They do not show real UPS counts, real AI training loads or real battery life.
• UL 1973: apart from its appearing as a test standard in the Samsung SDI specification, we could not confirm its title and edition on the standards body's own page.

11. How to read evaluations and standards

Before comparing which number is larger, align the basis of comparison in the following order. This is an editorial reading checklist, not a set of test conditions or acceptance limits prescribed by a standard.

1. What was measured?

deliverable power, hold-up time, transfer time and total ownership cost. Check the units and whether the result describes a material, component or complete system.

2. Under which conditions?

load changes, redundancy, cooling, backup boundary and end-of-life capacity. Treat results under different conditions as separate results before comparing them.

3. The comparison trap

Do not confuse kW with kWh or nameplate capacity with runtime at the actual load.

4. What does the standard cover?

Check the identifier, title, edition and scope in the issuing body’s material. Distinguish test methods, safety, interoperability and management requirements. Naming a standard alone does not establish that a test, conformity assessment or certification has been completed.

Keep a comparison note containing the object, units, conditions, sample count and variation, comparator, standard edition and unresolved points. Do not invent numerical conditions or acceptance limits from a public abstract of a standard.

12. Glossary

13. References

1. NERC "Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions", 8 January 2025 (PDF) — https://www.nerc.com/globalassets/our-work/reports/event-reports/incident_review_large_load_loss.pdf
2. NERC "Reliability Insights: Large Computational Load Risks Due to Voltage Sensitivity", September 2026 (PDF) — https://www.nerc.com/globalassets/who-we-are/news/2026/09/reliability-insights_large-computational-load-risks-due-to-voltage-sensitivity.pdf
3. Microsoft "Microsoft datacenter batteries to support growth of renewables on the power grid", 7 July 2022 — https://news.microsoft.com/source/features/sustainability/ireland-wind-farm-datacenter-ups/
4. Centrica "Fluence and Centrica work with Data Centre in Belgium to Provide and Optimise its first Zero-Emission Backup Power System", 21 April 2022 — https://www.centrica.com/media-centre/news/2022/fluence-and-centrica-work-with-data-centre-in-belgium-to-provide-and-optimise-its-first-zero-emission-backup-power-system
5. Google "How we're making data centers more flexible to benefit power grids", 4 August 2025 — https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/how-were-making-data-centers-more-flexible-to-benefit-power-grids/
6. NVIDIA "How New GB300 NVL72 Features Provide Steady Power for AI", 28 July 2025 — https://developer.nvidia.com/blog/how-new-gb300-nvl72-features-provide-steady-power-for-ai/
7. Choukse et al. (arXiv) "Power Stabilization for AI Training Datacenters", arXiv:2508.14318 (August 2025) — https://arxiv.org/abs/2508.14318
8. NVIDIA "Designing Production-Ready Battery Energy Storage Systems for AI Factories", 10 June 2026 — https://developer.nvidia.com/blog/designing-production-ready-battery-energy-storage-systems-for-ai-factories/
9. Samsung SDI "Lithium Ion Battery System for UPS – U6A4, Product Specification for End Customers", first draft, 28 April 2025 (PDF, hosted on Vertiv's site) — https://www.vertiv.cn/49f6fc/globalassets/shared/lithium-ion-batteries-system-for-ups---u6a4-ul-250428-product-specification.pdf
10. Uptime Institute "Lithium Ion Batteries for the data center. Are they ready for production yet?", November 2019 — https://journal.uptimeinstitute.com/lithium-ion-batteries-in-the-data-center/
11. Uptime Institute "Major data center fire highlights criticality of IT services", January 2023 — https://journal.uptimeinstitute.com/major-data-center-fire-highlights-criticality-of-it-services/
12. IEC "IEC 62040-3:2021 Uninterruptible power systems (UPS) – Part 3: Method of specifying the performance and test requirements" — https://webstore.iec.ch/en/publication/60140
13. IEC "IEC 62619:2022 Secondary cells and batteries containing alkaline or other non-acid electrolytes – Safety requirements for secondary lithium cells and batteries, for use in industrial applications" — https://webstore.iec.ch/en/publication/64073
14. UL Solutions "UL 9540A Test Method for Battery Energy Storage Systems (BESS)" — https://www.ul.com/services/UL-9540A-test-method
15. NFPA "NFPA 855 Standard for the Installation of Stationary Energy Storage Systems" (current edition 2026) — https://www.nfpa.org/codes-and-standards/nfpa-855-standard-development/855
16. IEA "Energy and AI" (executive summary) — https://www.iea.org/reports/energy-and-ai/executive-summary
17. U.S. Department of Energy (DOE) "DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers", 20 December 2024 (introducing the LBNL report) — https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
18. Redwood Materials "Redwood Energy: Fast, low-cost storage to power the age of AI and a changing grid", 27 June 2025 — https://www.redwoodmaterials.com/news/redwood-energy-fast-low-cost-storage-to-power-the-age-of-ai-and-a-changing-grid/
19. Redwood Materials "Redwood and Crusoe expand compute to 7x scale after 99.2% microgrid uptime", 24 March 2026 — https://www.redwoodmaterials.com/news/redwood-and-crusoe-expand-compute-to-7x-scale/

14. Claim-to-source audit

Last updated 10 October 2026. Sources are limited to primary material (standards bodies, public agencies, company announcements and a preprint). Company figures are treated as company claims and kept apart from independently verified values. We use no market-size or share estimates. Our calculations (Chapters 5 and 7) explain structure under stated assumptions and do not show a real design, load or battery life. That the Samsung SDI specification is a first draft, operating-result figures, the root cause of the fire, and the scale of supercapacitor use could not be confirmed from primary sources and are not asserted. Figs. 1 and 6 are AI-generated images and Figs. 2 to 5 are vector drawings; none of them shows a real facility, product or measured data.