Grid Integration of AI Data Centers: A Critical Review of Energy Storage Solutions (arXiv preprint)

Overview

AI workloads impose highly dynamic, hard-to-forecast power profiles on the grid, creating reliability and stability challenges that conventional data-centre architectures are not designed to address. The review compares storage layer by layer on response timescale, power and energy rating and role, with fuel cells, thermal storage and second-life batteries as supplements. It argues that sub-second training-load variability makes conventional ESS dispatch insufficient and that coordinated multi-layer storage is needed, and lists gaps in simulation tools, degradation modelling, sub-second load forecasting and multi-layer sizing.

Layer taxonomy: chip-level (on-chip capacitors, HBM buffers) sub-ms to ms at W-kW; rack/server BBUs and supercapacitors ms to seconds at kW-tens of kW; traditional UPS under 20 ms at 100s kW-MW; grid-interactive UPS 0.5-10 s FFR; BTM BESS minutes-hours. UPS battery energy sits idle for more than 99% of its lifetime, which the authors present as a significant opportunity cost. In grid-interactive UPS recharge mode, maximum recharge power is typically limited to 20-25% of nominal UPS capacity.

GiUPS units typically provide reactive power of 20-40% of rated kVA; a 1 MW/1.1 MVA unit can supply about ±250-300 kVAr, helping raise PCC power factor from 0.92-0.95 to above 0.99. Typical LVRT profiles cited for GiUPS: remain connected down to 0.2 pu for 100-150 ms and down to 0.5 pu for up to 500 ms, delivering 2-3 pu reactive current within 2-5 ms. Rapid AI load variations are described as often inducing 5-10% voltage swings within milliseconds.

Limits: as a review it has no new measurements; the authors found no software specifically developed for simulating energy management systems of AI data centres. This is the preprint version of the same work as the published article.

Key points

1. Layer taxonomy: chip-level sub-ms to ms (W-kW); rack/server BBUs and supercapacitors ms to seconds (kW-tens of kW); traditional UPS under 20 ms (100s kW-MW); grid-interactive UPS 0.5-10 s FFR; BTM BESS minutes-hours.
2. UPS battery energy sits idle for more than 99% of its lifetime, which the authors present as a significant opportunity cost.
3. In grid-interactive UPS recharge mode, maximum recharge power is typically limited to 20-25% of nominal UPS capacity.
4. GiUPS units typically provide reactive power of 20-40% of rated kVA; a 1 MW/1.1 MVA unit can supply about ±250-300 kVAr, raising PCC power factor from 0.92-0.95 to above 0.99.
5. Typical LVRT profiles for GiUPS: connected down to 0.2 pu for 100-150 ms and down to 0.5 pu for up to 500 ms, delivering 2-3 pu reactive current within 2-5 ms.
6. Rapid AI load variations often induce 5-10% voltage swings within milliseconds.
7. The authors found no software specifically developed for simulating energy management systems of AI data centres.

Using it in design

• Layer storage by timescale (chip/server sub-ms to ms, rack ms to seconds, UPS and GiUPS sub-second to 10 s, BESS minutes to hours) and show each layer's duty in the coordination scheme.
• For grid-interactive UPS, check the recharge power limit (20-25% of nominal capacity) and LVRT profile (0.2 pu for 100-150 ms, 0.5 pu for 500 ms) against grid-connection requirements.
• Assess GiUPS reactive-power support (20-40% of rated kVA) together with the point-of-connection power factor target (above 0.99).

Who should read it

Engineers designing UPS, rack BBU and BESS architecture and grid services for AI data centres.

Document contents

1. II. AI DC Power Architecture (IT equipment, power delivery systems, cooling systems, miscellaneous units)
2. III. UPS Systems in AI DCs (traditional UPS operation and limitations, GiUPS systems, BESS sizing in GiUPS systems)
3. IV. BESS in AI DCs (control modes, grid support roles, cost and sizing)
4. V. Fuel Cells
5. VI. Thermal Energy Storage
6. VII. Rack- and Server-Level Storage (BBUs, server- and GPU-level storage, sizing, coordination across layers)
7. VIII. Second-Life Battery Energy Storage
8. IX. Challenges and Future Directions

Related standards

• IEC 62040-3
• IEEE 2800-2022
• OCP Open Rack V3

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

arXiv preprint (University of Michigan)

Year

2026

Document number

arXiv:2603.00415

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.48550/arXiv.2603.00415

Licence

CC BY 4.0 — share with attribution

Licence evidence

arxiv.org

Citation