As AI clusters become denser and more dynamic, Siemens is turning its attention to the electrical backbone supporting the compute
AI is changing more than how much power data centres consume – it is changing how that power needs to move through the facility and Siemens is betting the electrical architecture itself will become a critical part of AI infrastructure.
For years, data-centre power infrastructure was largely about delivering reliable electricity from the grid to IT equipment, with sufficient redundancy and headroom for growth, AI is making that model harder to maintain. As GPU clusters become denser and workloads become more dynamic, operators are dealing with higher power requirements, tighter efficiency targets and potentially faster changes in electrical demand. The result is that power conversion and distribution are moving closer to the centre of the data-centre design conversation – Siemens is positioning itself around that shift.
In August, the company and transformer specialist Reinhausen announced a strategic partnership to develop and industrialise a solid-state transformer (SST) for 800VDC AI data centres. The modular system is designed to connect to grid voltages of up to 36kV and deliver 800VDC directly into downstream data-centre infrastructure.
The case for rethinking the electrical architecture
The significance of the project is bigger than the transformer itself – conventional data centre electrical architectures can involve multiple stages of conversion before electricity reaches the IT load. Siemens and Reinhausen are exploring an alternative: converting medium-voltage AC into 800VDC as part of a more direct power architecture. For AI operators, that raises an important question: does the traditional electrical design still make sense when the workload is fundamentally different?
Higher rack densities increase the amount of power that has to be delivered into a relatively small physical footprint. At the same time, AI training workloads can create rapid changes in power demand. Siemens says it’s SST concept is designed to respond to those changes within milliseconds – this is a company stated capability and, because the technology is still at the development and industrialisation stage, should not be treated as an independently demonstrated performance benchmark, yet, but the direction is revealing.
The industry is starting to look at the electrical system not simply as infrastructure sitting underneath the compute, but as something that needs to be designed around the behaviour of the compute itself.
AI makes power an architectural problem
This is where Siemens’ broader data-centre strategy becomes interesting – they already operate across power distribution, automation, power-quality monitoring, cooling, microgrids and infrastructure management, their positioning around AI data centres increasingly connects those areas rather than treating them as separate pieces of equipment which makes sense.
Once AI clusters become sufficiently dense, decisions about electrical distribution affect more than resilience, they can influence efficiency, available capacity, cooling requirements, expansion plans and how effectively a facility can respond to changes in workload – in other words, power is becoming part of the AI architecture.
This is an important shift for data centre operators – the question is no longer simply whether a site has enough megawatts, It is whether the electrical infrastructure can deliver those megawatts efficiently, reliably and in a form that matches the requirements of increasingly dense AI systems.
The hardware supply chain matters
Siemens is also expanding the physical manufacturing capacity behind its data centre strategy – they recently announced more than $200 million of investment in two new US manufacturing facilities in Georgia and Texas, which they say will create more than 1,500 jobs and expand production of electrical equipment supporting AI data centres – they also reported approximately €6 billion in orders during the first nine months of its fiscal 2026, which highlights the scale of the market opportunity they are targeting.
The AI infrastructure bottleneck is not necessarily limited to GPUs or electricity generation – transformers, switchgear, distribution equipment, power-management systems and other electrical infrastructure all have to be manufactured, delivered and integrated before a new AI facility can become operational, and that matters
The electrical backbone is now strategic
For operators, the bigger lesson from Siemens’ strategy is that electrical infrastructure needs to be considered much earlier in planning. Power density, conversion efficiency, grid connections, protection, load variability and future expansion are increasingly interconnected decisions – whilst Siemens’ 800VDC SST project represents one possible response to those challenges rather than a finished industry standard, it does reflect a wider direction: AI is forcing data centre power architecture to evolve alongside the compute.
The GPUs remain the most visible part of an AI facility, however, the infrastructure that delivers power to those GPUs could determine how efficiently and how quickly the next generation of AI capacity can actually be built.
