Schneider Electric has published a study assessing arc‑flash hazards in 800 VDC power architectures being adopted for AI‑focused data centres. The analysis compared two representative 800 VDC configurations based on design patterns from hyperscale operators and concluded that arc‑flash consequences depend strongly on system topology, capacitor placement and protection practices.
Key findings
- Even under conservative scenarios, arc‑flash risk in 800 VDC installations can be managed and in many cases reduced to levels comparable with conventional alternating current (AC) systems.
- Transient behaviour matters: capacitor discharge dominates the first milliseconds of a DC arc‑flash event, so time‑dependent modelling and millisecond‑level protective response are critical.
- Existing arc‑flash assessment frameworks developed for AC systems can be applied to 800 VDC if they account for topology and time‑dependent effects.
- Advanced simulation tools and digital twins, such as ETAP, improve accuracy of risk estimates and support more effective protection strategies.
Configurations evaluated
Schneider Electric produced the first practical arc‑flash analysis for both rack‑level and facility‑level 800 VDC architectures, covering common implementation approaches in the field:
-
Rack‑level 800 VDC architectures: under conservative assumptions the calculated incident energy (arc‑flash energy) remained well below the 1.2 cal/cm² reference used for personal protective equipment (PPE), even in models that did not include dedicated protective devices.
-
Centralised (facility‑level) 800 VDC architectures: a conservative, less realistic topology without overcurrent protection produced higher incident energy than the rack‑level cases. The study assessed how topology affects back‑feed, peak currents and arc‑flash outcomes — for example, faults occurring before or after reverse‑blocking diodes change the way energy feeds into the fault.
When faults are time‑limited by standard protective devices, arc‑flash energy falls to levels appropriate for work environments and generally aligns with the risk levels of typical AC architectures.
Design implications and recommendations
The report emphasises that reducing arc‑flash consequences depends mainly on design decisions rather than simply on the fact the system uses DC. Important factors include:
- placement and sizing of capacitors;
- use and location of back‑feed prevention devices such as reverse‑blocking diodes;
- millisecond‑level protection coordination and response;
- careful evaluation of system topology and switching logic;
- use of advanced simulations and digital twins to model real‑world behaviour.
Manish Kumar, Schneider Electric’s executive vice president responsible for "Secure Power & Data Centers", said the move to 800 VDC represents a significant shift in data‑centre design and that the company aims to provide a practical framework for engineers and safety professionals to assess arc‑flash risk and develop safe working practices.
Tanuj Khandelwal, CEO of ETAP, noted that traditional standards can be overly conservative because they do not fully reflect the behavior of complex DC systems, and that operators benefit from modelling and validating 800 VDC systems in operation using physics‑based models and digital twins.
Why this matters now
800 VDC distribution is being adopted more widely in AI data centres because it supports higher power densities and enables IT racks of 400 kW and above — an evolution supported by industry players including NVIDIA and various energy‑technology partners. Higher DC voltages necessitate a deeper understanding of potential failure modes, protective measures and safe working procedures.
Availability
The study’s full results are available in the white paper titled "DC Arc Flash Analysis: A Practical Study on 800 VDC in Data Centers", which documents the assessment methods, simulation role and recommended design practices.
Note: the term “AI factory” refers to a specialised computing infrastructure designed to manage the full AI lifecycle and extract value from data.
