• • • DATA CENTRES • • •
CAN 800VDC MEET THE RISING POWER DEMANDS OF AI DATA CENTRES?
BY ALEXANDRA RANGEL-SANTOS,
SENIOR MANAGER IN PRODUCT DEVELOPMENT DIRECT CURRENT, EATON
applications for enterprises and consumers. A key challenge is how these new AI data
G
centres consume considerably more power than traditional data centres. To run AI tasks, much higher power densities are needed, with racks of AI chips requiring more than 120kW and, in some cases, over 1MW. Conventional racks, by comparison, typically use between 10kW and 50kW. Data centre operators and engineers must
rethink how power is converted, distributed, protected and managed within a facility. The electrical architecture can no longer be considered separately from computing and cooling. Instead, these systems need to be designed together from the beginning, with power, thermal management, controls and rack architecture working as part of one coordinated system.
The case for higher-voltage DC As AI rack densities rise, 800VDC distribution presents a way to reduce distribution losses, conductor size and some of the complexity associated with high-amperage systems. Within this approach, the conversion layer
between medium and low voltage is particularly important. Medium-voltage solid-state transformers (MVSSTs), combined with 800VDC distribution, can supply advanced AI racks through busway systems or direct DC remote power panels.
lobally, new data centre infrastructure is being built to accommodate the rapid growth in artificial intelligence
Design choices 800VDC is not simply a replacement for conventional AC distribution and introduces new design and protection requirements. For example, two-wire systems offer simplicity and use less material, while three-wire configurations may provide greater flexibility as requirements evolve. As standards continue to develop, maintaining flexibility rather than committing too early to one configuration is a practical approach. At the same time, engineers should assess fault
behaviour early in the development process, as DC systems require a different protection approach. This includes accounting for the initial high- current discharge from capacitors in the source and rack converters, followed by a current-limited phase that can continue for several seconds.
Protecting high-density systems Protection planning must, therefore, account for how these different fault conditions behave. For pole-to-pole faults, the full output of the DC
source can be imposed on the system. Protective devices must, therefore, be selected with response times matched to the system’s fault profile. Fast devices such as solid-state circuit breakers and hybrid circuit breakers can respond to high-speed fault events, while slower protection devices such as fuses and moulded-case circuit breakers address longer-duration events. The grounding approach needs to be considered
when managing pole-to-ground faults. In solidly grounded systems, fault currents can be similar to pole-to-pole faults and identified through overcurrent protection. By contrast, resistance- grounded systems are more challenging because the resistor can limit fault current to levels that may resemble normal load conditions. As a result, these systems require specialist ground-fault detection relays and residual or leakage current monitoring. Alongside this, protection coordination is essential. The aim is for the device closest to a
22 ELECTRICAL ENGINEERING • SEPTEMBER 2026
fault to disconnect, allowing unaffected racks and feeders to remain operational. However, achieving this familiar principle becomes more complex in DC systems with power electronics and high-density loads.
Managing failure and inrush In addition to selecting appropriate protection, designers need to consider how AI rack converters behave during energisation and faults. These converters can create significant inrush currents that could trip upstream protection, so pre-charge circuits or controlled voltage ramping can help manage this risk. Similarly, energy stored in source and rack
capacitors can be released into a fault within microseconds. If current rises faster than a solid-state breaker can withstand, the breaker itself may be damaged. In some situations, adding inductance can help slow the current rise. Designers must also consider component failures, as a failed feeder breaker could affect several racks, reducing redundancy and creating wider disruption. Limiting this potential ‘blast radius’ is, therefore, important in high-density AI environments.
Preparing for AI workloads An 800VDC architecture has the potential to support the higher rack power densities created by AI workloads. However, its adoption will be guided by how the industry addresses protection, fault behaviour and evolving standards. The industry is moving toward power systems
designed less like traditional data centres and more like integrated, high-performance energy platforms. By considering power conversion, distribution,
protection, controls and rack requirements as one system, operators can create architectures better suited to the scale and complexity of AI.
www.eaton.com
electricalengineeringmagazine.co.uk
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48
