Getting ready for 800V DC data centres: what engineers need to know
AI compute is forcing a fundamental rethink of data centre power architecture. 800V DC distribution isn’t a distant prospect – it’s being specified now. Here’s a straightforward account of what is driving the change, what it means in practice, and what engineering teams should be planning for.
The data centre industry has spent decades building around alternating current. The engineering conventions, the components, the safety standards and the skills base all reflect that history.
The move toward 800V DC power distribution for high-density AI compute is not just an incremental change to that architecture – in several ways, it’s a fundamental departure from it. That’s not a reason to be alarmed. It’s a reason to engage with the technical implications now, rather than when the first 800V DC project lands on your desk.
Why AI compute is driving the change
The proximate cause is rack power density.
A standard enterprise server rack has historically operated at somewhere between 5 and 15 kW. Modern AI training racks – built around dense GPU clusters – are already reaching 100 to 200 kW, and current development trajectories point toward 600 kW per rack within the next few years. Proposals associated with next-generation compute platforms suggest individual racks eventually approaching one megawatt.
At those power levels, delivering electricity at conventional voltages creates a conductor problem. The current required to feed a 600 kW rack at 50V – the standard server-level DC voltage used in most current deployments – would require cables and busbars of impractical size. The solution is to increase the distribution voltage.
“The move to 800V DC is not going to happen uniformly or immediately across the industry. But the direction of travel is clear, and the engineering teams who have worked through the technical implications in advance will be substantially better placed when the first serious project requirement arrives.“
Tom MacPartlin, Business Development Manager, Capital
At 800V DC, the same power can be delivered with conductor cross-sections that are physically manageable.
At 50V DC, feeding a 600kW rack would require cables and busbars of impractical size. Increasing distribution voltage to 800V DC delivers the same power with conductor cross-sections that are physically workable – and frees up space for more compute.
What 800V DC distribution looks like in practice
The architecture currently under active development works like this:
- Existing 400V AC power trains are converted to plus-or-minus 400V DC – effectively 800V DC across the rails – using converters mounted close to the IT load.
- This converted DC supply feeds the AI compute racks directly, replacing the multiple AC-to-DC conversion stages that are a feature of conventional UPS-based distribution.
The longer-term vision goes further: direct conversion from medium-voltage AC supply to 800V DC using solid-state transformers. Or alternatively rectification of secondary output of MV/LV transformer.
The implications for project programmes and supply chain management would be significant.
Three 800V DC engineering advantages that matter
Beyond the conductor sizing argument, 800V DC offers three further advantages:
- Energy efficiency: Every AC-to-DC conversion in a conventional power train carries losses. A UPS-based system with multiple conversion stages can lose several percent of the input energy to heat. By distributing at DC and eliminating conversion stages, 800V DC systems improve PUE directly. For an operator running hundreds of megawatts of IT load, this is a material operational cost reduction.
- Renewable integration: Connecting variable renewable generation sources – solar arrays, wind, local storage – to an AC distribution system requires inverters and synchronisation equipment. The same sources connect far more naturally to a DC bus, with lower losses and simpler control architecture. As data centre operators face increasing pressure to demonstrate renewable credentials, this matters.
- Power quality management: This is arguably the most significant for AI-specific applications. AI training and inference loads exhibit very high-frequency, high-amplitude power fluctuations – the result of GPUs cycling rapidly between computation states. Managing these fluctuations on an AC system requires synchronisation with the voltage waveform, which constrains the speed and scale of the response. A DC bus with local storage integration can buffer these spikes far more effectively, reducing stress on the upstream network and improving the stability of supply to sensitive compute hardware.
DC distribution connects naturally to renewable generation sources and local storage – without the inverters and synchronisation equipment that AC systems require. For operators under pressure to demonstrate renewable credentials, this is a meaningful advantage.
What the supply chain can and cannot offer today
It’s important to be clear-eyed about where the technology stands. 800V DC distribution is not yet a mature, off-the-shelf product category.
Key components – DC circuit breakers with appropriate interrupt ratings, solid-state protection devices, and the converters themselves – are in active development, and different manufacturers are taking meaningfully different approaches to the design challenges. Standards bodies are still working through the certification frameworks.
Engineers who are beginning to specify 800V DC projects today are, to some degree, working ahead of the supply chain. That’s uncomfortable but manageable, as long as expectations are realistic.
The key is maintaining close contact with component manufacturers to understand their development timelines, and designing systems with sufficient flexibility to accommodate component specifications that may change before the project reaches installation.
The protection engineering challenges are also genuinely novel for engineers whose careers have been built on AC systems.
Selective protection coordination with solid-state DC circuit breakers behaves differently from the fuse and MCCB coordination familiar from AC distribution. Arc flash behaviour at DC is different too. These are solvable problems – they’re being solved – but they require engineers to revisit some foundational assumptions.
Want to move from AC to 800V DC? What to do next.
For engineering teams whose projects currently run on conventional AC distribution, the practical next steps are:
- Start tracking the development of key components and the progress of relevant IEC and IEEE standards.
- Engage with switchgear manufacturers who are actively developing 800V DC products and who can offer early access to prototype testing.
- Consider how existing facilities might be designed or retrofitted to accommodate a transition – particularly with respect to MV infrastructure, which will still be needed in the near term even as LV architecture evolves.
The move to 800V DC is not going to happen uniformly or immediately across the industry.
But the direction of travel is clear, and the engineering teams who have worked through the technical implications in advance will be substantially better placed when the first serious project requirement arrives.
Designing for high-density AI compute? We’re actively developing our 800V DC product range and welcome conversations with engineering teams planning for this transition. Get in touch
Insights
Most data centre procurement teams are measured on capital cost. That's understandable – but it's the wrong metric for LV switchgear. Over a 10-to-15-year asset life, the decisions made at procurement stage can cost or save millions. Here's how to think about it properly.
Switchgear lead times are one of the most stubborn constraints in data centre construction. But schedule pressure is rarely the supplier's fault alone. Here's what experienced project teams do differently – and how you can shave weeks off your next deployment.