Grid Integration of AI Data Centers: A Critical Review of Energy Storage Solutions

Overview

The review starts from the premise that AI workloads impose fast, hard-to-forecast power profiles on the grid, creating reliability and stability problems that conventional data-centre architectures were not designed for, and asks how energy storage responds. The method is a critical literature review: storage is analysed by response timescale, power and energy ratings, operational role, integration challenges and coordination across a four-layer taxonomy, supplemented by fuel cells, thermal storage and second-life batteries.

Traditional UPS units act only as passive protective devices that interact with the grid through battery charging and synchronisation; the review notes UPS battery energy sits idle for more than 99% of its lifetime. For grid-interactive UPS (GiUPS), battery sizing must cover backup autonomy as the baseline plus committed grid-service power and cycling; moderate oversizing can lower effective DoD and extend life, while excessive oversizing adds cost, footprint and thermal burden. Rack BBU shelves must carry full rack power for the target ride-through with one module failed; in ORV3 designs this is 5+1 redundancy, and OCP reference designs target minutes of ride-through.

Numerical context: BESS cost fell from about $7,500/kWh in 1991 to $133/kWh in 2024, with about $80/kWh projected by 2030. A cited study shows a 560 kWh Li-ion BESS with a 960 kWh lead-acid BESS can provide fast frequency response and backup for a traditional data centre. AI racks can exceed 100 kW and reach MW scale in dense configurations, which greatly increases BBU shelf energy and raises Li-ion thermal risk when cells sit next to IT equipment. Server-level supercapacitors are sized by the energy of GPU power swings rather than by cycle life.

Conclusions: AI load profiles are fundamentally different from traditional loads, with sub-second variability, high peak-to-average ratios and unpredictable workload transitions; GiUPS is the most significant architectural evolution, turning passive UPS batteries into active grid assets capable of FFR, VRT, peak shaving and synthetic inertia; and fuel cells and thermal storage are underused in the literature relative to their potential. Limits: as a review it has no new measurements, and the authors list gaps in simulation tools, degradation models, load forecasting and multi-layer sizing as research needs.

Key points

1. Traditional UPS units act only as passive protective devices; the review notes UPS battery energy sits idle for more than 99% of its lifetime.
2. GiUPS battery sizing must cover backup autonomy as the baseline plus committed grid-service power and cycling; moderate oversizing can lower effective DoD and extend life, excessive oversizing adds cost, footprint and thermal burden.
3. BESS cost fell from about $7,500/kWh in 1991 to $133/kWh in 2024, with about $80/kWh projected by 2030.
4. A cited study shows a 560 kWh Li-ion BESS with a 960 kWh lead-acid BESS can provide fast frequency response and backup for a traditional data centre.
5. Rack BBU shelves must carry full rack power for the target ride-through with one module failed; ORV3 designs use 5+1 redundancy and OCP reference designs target minutes of ride-through.
6. AI racks can exceed 100 kW and reach MW scale in dense configurations, raising BBU shelf energy and Li-ion thermal risk when cells sit next to IT equipment.
7. Server-level supercapacitors are sized by the energy of GPU power swings rather than cycle life, because they tolerate very high cycle counts.

Using it in design

• Layer storage by timescale: GPU/server supercapacitors for sub-second swings, rack BBUs for minutes, facility GiUPS and BESS for grid services and longer durations; show each layer's duty and coordination separately in the architecture.
• Size rack BBU shelves to carry full rack power for the target ride-through with one module failed (for example 5+1), and link the Li-ion thermal risk of cells next to IT equipment to fire and ventilation design.
• Size the GiUPS battery by counting backup autonomy and the grid-service commitment separately, and show the trade-off of oversizing between DoD and life and cost, footprint and thermal burden.
• Include fuel cells and thermal storage in the comparison of options for diesel backup and cooling.

Who should read it

Designers and energy engineers planning UPS, rack BBU and BESS architecture and grid services for AI data centres.

Document contents

1. I. Introduction
2. II. AI DC Power Architecture (A. IT equipment, B. Power Delivery Systems, C. Cooling Systems, D. Miscellaneous Units)
3. III. UPS Systems in AI DCs (A. Traditional UPS Operation and Limitations, B. GiUPS systems in AI DCs, C. BESS Sizing in GiUPS Systems)
4. IV. BESS in AI DCs (A. BESS Control Modes, B. Grid Support Roles of BESS, C. BESS Cost and Sizing Considerations)
5. V. Fuel Cells (FCs) in AI DCs
6. VI. Thermal Energy Storage (TES) in AI DCs
7. VII. Rack- and Server-Level ESSs Roles in AI DCs (BBUs as distributed UPS, server- and GPU-level storage, sizing, coordination across rack and server layers)
8. VIII. Second-Life Battery Energy Storage
9. IX. Challenges and Future Directions
10. X. Conclusions

Related standards

• IEC 62040-3
• IEEE 1547
• OCP ORV3
• NFPA 855

Read the document

Publication details

Authors

Sina Mohammadi; Wayne Wang; Marcus Chen I Wada; Rouzbeh Haghighi; Ali Hassan; Hualong Liu; Archit Bhatnagar; Ang Chen; Wencong Su

Publisher

Advances in Applied Energy (Elsevier BV)

Year

2026

Document number

arXiv:2603.00415

Volume

23

Pages

100291

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.1016/j.adapen.2026.100291

Licence

CC BY 4.0 — share with attribution

Licence evidence

arxiv.org

Citation