IN Brief:
- Siemens, NVIDIA, Fluence, and nVent have developed a reference architecture for AI data centres.
- The design targets a 136MW facility with 100MW of IT load.
- Electrical, power, controls, and storage integration are becoming central to AI infrastructure deployment.
Siemens has developed a reference architecture for NVIDIA AI data centres, working with NVIDIA and Fluence, and incorporating design considerations aligned with nVent.
Aligned with NVIDIA DSX Vera Rubin NVL72 systems, the architecture is intended to translate high-density AI infrastructure into a deployable electrical, power, and controls design for hyperscalers, colocation providers, and specialist cloud infrastructure operators.
The reference design is sized for a total facility capacity of 136MW, including 100MW of IT load. It covers the complete electrical path from a nominal 34.5kV utility connection through medium-voltage distribution, modular low-voltage power blocks, and the rack interface.
Built around Tier III concurrent maintainability, the design allows a single component to be removed from service without affecting IT operations. Siemens uses repeatable and scalable electrical building blocks aligned with NVIDIA DSX Vera Rubin deployment units, supporting initial deployments at tens of megawatts and later expansion towards hundreds of megawatts without a fundamental redesign of the electrical architecture.
Fluence contributes battery energy storage capability, while nVent-aligned design parameters support compatibility with NVIDIA workloads and system architectures. A further supplement is planned for advanced thermal management, reflecting the tightening relationship between electrical engineering, cooling, controls, and system availability in high-density AI facilities.
Power infrastructure has become one of the most difficult constraints in AI deployment. Processor availability and accelerator supply remain important, but grid connection, power conversion, thermal rejection, redundancy, and commissioning speed now shape how quickly compute capacity can be brought online.
That pressure is visible across the wider electronics stack. Microchip’s 3.3kV silicon carbide modules for medium-voltage power conversion address higher-voltage conversion, while its PCIe 6.0 and CXL 3.1 retimers for AI fabrics target the data-movement side of accelerated infrastructure. Both areas now sit within the same design conversation: AI systems need more power, tighter signal integrity, and lower deployment risk at the same time.
The Siemens architecture also points towards more industrialised data-centre construction. High-density AI facilities increasingly resemble process plants in their dependence on prefabricated assemblies, repeatable electrical blocks, protection coordination, digital modelling, and staged expansion. Modular medium-voltage and low-voltage skids can reduce on-site integration risk, provided the design remains compatible with the density and load behaviour of successive IT generations.
Battery energy storage is also moving deeper into the data-centre design envelope. For AI facilities, storage can support black start, grid services, voltage and frequency stabilisation, and smoothing of demand in constrained grid areas. That places power electronics, controls, and storage management closer to the centre of compute infrastructure engineering.
As rack densities rise, the path from grid connection to processor package is becoming a core design domain. Siemens’ reference architecture puts that power path into a repeatable framework, where electrical infrastructure is planned as part of the AI system rather than as a downstream facilities package.
