For the past few years, a lot conversations about AI data centre power have started in roughly the same place: where are we going to find all the electricity? It’s a fair question. AI infrastructure needs enormous amounts of power, and getting enough of it to a new data centre can be difficult.
But there’s another problem waiting on the other side of the grid connection. Once all that electricity reaches the building, it still has to get to the servers that actually need it. And those servers are becoming much more demanding. The International Energy Agency (IEA) says the power density of AI servers increased elevenfold between 2020 and 2025, and is expected to increase another fourfold by 2027.
To put that into perspective, the IEA estimates that a single advanced server rack could then have the same peak power demand as 65 households. That’s starting to change the way data centres themselves need to be built.
Instead of only asking whether enough electricity can reach the site, operators also have to ask whether the infrastructure inside the building can move that much power into a relatively small group of racks. 800 VDC is emerging as one possible answer. The name simply means 800-volt direct current, but the change behind it goes much further than swapping one voltage for another.
It changes how electricity travels through the data centre, where it gets converted and, potentially, how several other parts of the facility are designed around it.
AI Racks Are Outgrowing Conventional Power Distribution
To understand why 800 VDC is suddenly getting so much attention, it helps to start with something that isn’t always obvious from the outside. Electricity doesn’t enter a data centre and go straight to a GPU. It has quite a journey first. Electricity from the grid arrives as alternating current, or AC.
The processors inside a server ultimately need direct current, or DC. Along the way, the electricity may also need to be stepped down to different voltages, conditioned to keep the supply stable and passed through backup power equipment. Each of those jobs requires electrical equipment.
Uptime Institute says a conventional setup using double-conversion uninterruptible power supplies, better known as UPS systems, and standard IT power supplies can involve as many as five conversion stages before the electricity reaches the computing equipment. That architecture has supported data centres for years.
The difficulty is that AI is starting to ask it to do something very different. Think about the problem physically for a moment. If you want to deliver much more electrical power through the same kind of lower-voltage system, you have to increase the current. And the more current you need to carry, the more substantial the equipment carrying it has to become.
That can mean thicker cables, larger conductors and more copper. It also means more of the limited space around the rack is being used simply to get electricity into it. At ordinary rack densities, this isn’t the same problem. Even now, the wider data centre market is nowhere near running megawatt racks as standard.
But the highest-density AI systems are moving into very different territory. Schneider Electric says conventional AC and 48 VDC distribution start running into practical constraints as next-generation AI racks move beyond hundreds of kilowatts. This is the part that’s easy to miss when we talk about AI’s appetite for electricity.
The challenge isn’t only that AI needs more power overall. Increasingly, it needs a huge amount of that power concentrated into a very small physical space. Once that happens, changing how the electricity gets there starts to make sense.
Why 800 VDC Changes The Equation
There are two basic ways to deliver more electrical power. You can increase the current, or you can increase the voltage. We’ve already seen the problem with simply increasing the current. Eventually, all the copper, cabling and other equipment needed to carry it starts becoming impractical.
Increasing the voltage offers another option because the same amount of power can then be delivered with less current. That’s the basic thinking behind 800 VDC architecture. The Open Compute Project (OCP) says higher-voltage DC can move more power using less conductor and copper than lower-voltage alternatives.
For an AI data centre trying to fit more compute and more power into the same physical space, that starts removing some of the pressure from the electrical infrastructure around the rack. There’s another piece to this as well. Remember that journey electricity currently takes through the building? Every time it moves between AC and DC, another conversion has to happen. With 800 VDC, some of that conversion can move further away from the server rack.
The electricity is converted from AC to DC earlier, then travels through more of the facility as high-voltage DC before being reduced to the lower DC voltages the computing equipment actually uses. In practical terms, that can mean fewer conversion stages and less power equipment competing with GPUs for space inside the rack.
Schneider identifies both reduced congestion around electrical feeds and more room for compute as benefits of moving the main conversion outside the IT rack. Of course, none of this creates more electricity. An 800 VDC system can’t solve a shortage of grid capacity or make a 100 MW connection provide 120 MW.
What it can do is change how efficiently and practically the power that is available moves through the building. There’s also a reason we’re talking specifically about 800 VDC rather than a collection of different higher-voltage systems. Google, Microsoft and NVIDIA are working together through OCP to develop common requirements for the architecture.
The aim is to give operators and equipment manufacturers a shared set of interfaces and technical requirements rather than leaving every company to develop its own version. That work is still underway, which is important. 800 VDC isn’t established infrastructure across the data centre industry today.
It’s an emerging response to the much higher rack densities being planned for the next generation of AI systems. And because it’s emerging, operators don’t have to choose between keeping everything they already have or rebuilding the entire facility around DC.
Moving To 800 VDC Doesn’t Mean Rebuilding Everything
At first glance, moving from conventional AC distribution towards 800 VDC sounds like a fairly dramatic retrofit. If a building was designed around AC, surely changing the way it distributes electricity means changing the whole electrical system? Not necessarily. The more useful question is where the electricity changes from AC to DC.
One of the earliest options is to put that conversion right beside the AI racks. Schneider calls this an 800 VDC power rack, although you’ll also see the term “sidecar”. Instead of squeezing all the necessary power-conversion equipment into the same rack as the GPUs, a separate rack handles the conversion and supplies the neighbouring compute equipment with 800 VDC.
That makes it possible to introduce 800 VDC without redesigning everything upstream. The facility can continue distributing AC power through much of the building, then convert it closer to the high-density equipment that needs something different. NVIDIA is following the same general route with an 800 VDC power rack designed to work inside existing AC facilities.
Its August 2026 roadmap describes this hybrid approach as the starting point, followed by more centralised designs where conversion moves progressively further away from the individual rack. For example, the next step could be a central power system serving an entire row of racks through an overhead 800 VDC busway.
A busway is essentially a large enclosed conductor that distributes electricity along the row, allowing equipment to connect to it at different points. Further down the road, NVIDIA is planning designs that would convert medium-voltage AC directly to 800 VDC much closer to where power enters the facility.
Those later designs are roadmaps, not evidence that data centres everywhere are already being built this way. But they show why thinking about the change as “AC versus DC” isn’t particularly useful. A data centre can use both.
For an existing enterprise facility adding a few high-density AI racks, moving the conversion point closer to those racks may be enough. A new AI-focused facility being designed from scratch has the option to move that conversion much further upstream. Where it happens then starts influencing some of the other infrastructure around it.
The Change Doesn’t Stop At The Server Rack
Power and cooling are usually easy to talk about as separate data centre problems. One supplies the equipment with electricity. The other takes away the heat that equipment produces. At very high rack densities, the two become harder to separate in the design. To be clear, 800 VDC isn’t the reason AI data centres need more advanced cooling.
That shift was already happening as increasingly powerful GPUs packed more heat into each rack. What changes is the possibility that the cooling equipment itself could eventually connect to the same DC power architecture. And examples of what that might look like already exist.
Trane Technologies, working with Eaton and Danfoss, demonstrated a chiller that could operate directly from an 800 VDC supply. A chiller is the large piece of cooling equipment responsible for removing heat from the water used by many data centre cooling systems. The laboratory demonstration delivered more than 3.5 MW of cooling capacity.
Trane says removing some of the usual power conversions could improve system efficiency by up to two per cent compared with an equivalent conventional AC setup. That figure comes from the companies behind the demonstration, so it shouldn’t be treated as independent evidence of what every data centre could achieve.
What the test does show is that the 800 VDC conversation is already moving beyond the server rack itself. Batteries are another example. Batteries naturally store DC electricity. If more of the data centre also distributes power as DC, infrastructure teams can start asking whether every conversion between the battery and the equipment it supports is still necessary.
OCP’s emerging architecture includes options for directly integrating battery energy storage into an 800 VDC backbone. Then there are the sudden changes in power demand created by AI workloads. The IEA notes that AI training and model use can cause large, rapid swings in electricity demand.
The electrical system has to cope not only with a lot of power, but with that demand changing quickly as the computing workload changes. Put those pieces together and 800 VDC starts looking less like a new way to power a rack and more like a wider infrastructure design question.
Power conversion, storage, cooling and protection increasingly have to be considered together because a decision made in one place can change what makes sense somewhere else. Of course, connecting more of the facility around a different electrical architecture also introduces some new problems of its own.
800 VDC Solves One Scaling Problem By Creating New Design Questions
On paper, the attraction is fairly straightforward. Higher voltage reduces the amount of current needed. Moving the conversion equipment can free space around the compute. Removing unnecessary conversion stages can simplify the route electricity takes through the facility.
But electrical systems also have to be safe when something goes wrong. And DC behaves differently from the AC systems most data centre teams have spent years working with.
Protection and safety work differently with DC
Here’s the simplest version. AC electricity constantly changes direction. As it does, the current naturally passes through zero. That brief zero point can help electrical protection equipment interrupt a fault. DC keeps travelling in the same direction, so there isn’t an equivalent natural zero point.
If an electrical arc forms, breaking that current can therefore require a different approach. There’s another complication in an 800 VDC system. Components called capacitors can store small amounts of electrical energy so they can release it very quickly when the system needs it. During a fault, that stored energy also has to go somewhere.
This is why questions such as where those capacitors sit, how faults are detected and how quickly the affected part of the system can be disconnected become important. Schneider modelled two different 800 VDC designs in a 2026 arc-flash study and found that the outcome changed depending on the architecture, including where capacitors were positioned and how quickly protection systems cleared the fault.
It also found that existing methods for calculating arc-flash risk don’t fully capture some of the behaviour found in these newer converter-fed DC systems. That doesn’t give us a simple conclusion that DC is either safer or more dangerous than AC. It tells us something more useful: the risk depends on how the system is designed.
For operators, that brings circuit protection, grounding, fault isolation and maintenance procedures into the conversation much earlier. You can’t simply change the voltage and assume all the existing safety assumptions still apply.
Standards and operations are still catching up
Some of those questions are still being worked out because the architecture itself is still developing. OCP is working with organisations including UL Solutions, the National Fire Protection Association, IEEE and the International Electrotechnical Commission on the safety and regulatory frameworks needed for wider 800 VDC deployment.
Google, Microsoft, NVIDIA and the wider OCP community are also developing common requirements around areas such as power quality and the interfaces connecting different parts of the system. That interoperability work is particularly important. A data centre is built from equipment supplied by many different companies.
An open 800 VDC architecture only becomes genuinely useful if operators can combine compatible power, cooling, protection and computing equipment without becoming trapped inside one supplier’s design. Then there’s the human side. Operating a DC environment introduces different protection and maintenance requirements, so teams need people who understand them.
That skills requirement arrives when data centre recruitment is already difficult. More than half of the owners and operators responding to Uptime Institute’s 2026 Global Data Center Survey said they were struggling to find qualified candidates for open roles. So 800 VDC can make one part of the scaling problem easier while moving some of the complexity elsewhere.
Less current and fewer conversions don’t remove the need for careful infrastructure design. They change where that work needs to happen.
What Infrastructure Leaders Need To Decide Now
With all the activity around 800 VDC, it would be easy to come away with the impression that the entire data centre industry is preparing to move to megawatt racks. It isn’t. Uptime Institute’s 2026 Global Data Center Survey collected responses from more than 800 data centre owners and operators.
It found that average rack densities are still rising slowly, although a growing number of operators are now reporting peak densities of 30 kW or more. That’s an important reality check. The 800 VDC discussion is being driven by the extreme end of AI infrastructure, where rack densities are moving towards hundreds of kilowatts and potentially beyond.
It isn’t evidence that every enterprise data centre suddenly needs a completely new power architecture. For most infrastructure leaders, the useful starting point is much closer to home: what rack density will this facility actually need to support during its working life?
An existing data centre adding a handful of higher-density AI systems has a very different problem from a new facility being designed around future generations of accelerated computing. The first may eventually need a separate power rack beside its AI equipment. The second has the freedom to consider whether converting to DC much earlier in the electrical path makes more sense.
There’s also a timing question. NVIDIA said in August 2026 that more than 80 equipment manufacturers and infrastructure companies were already building products around the emerging OCP specifications. At the same time, OCP is still developing standards, safety frameworks and common technical requirements.
That leaves infrastructure teams with a familiar balancing act. Designing only for what today’s racks need could make future AI upgrades harder. Building too aggressively around an emerging architecture could leave the facility committed to choices that are still changing.
Which is why the immediate decision isn’t really, “Should we switch to 800 VDC?” It’s whether the AI systems an organisation expects to run over the next few years are starting to change the assumptions its current data centre electrical infrastructure was built around.
Once the answer to that becomes yes, where power gets converted, how it reaches the rack and what needs to change around it become much more useful questions.
Final Thoughts: AI Is Changing How Power Moves Through The Data Centre
AI’s power problem has always sounded like a question of quantity. More compute needs more electricity, so data centres need access to more power. That’s still true. It’s just no longer the whole story. When enough computing power is packed into one rack, simply getting electricity to the building doesn’t solve the problem.
The facility also needs a practical way to move an extraordinary amount of that electricity through a relatively small space and into the hardware waiting for it. 800 VDC is emerging because higher voltage offers one way to do that without simply pushing more and more current through increasingly bulky lower-voltage infrastructure.
Moving towards DC can also change where power is converted and how much electrical equipment needs to sit close to the compute. Whether that eventually means a power sidecar beside a few AI racks, an 800 VDC busway running across a data hall or a much more DC-native facility will depend on the building, the workloads and how the technology develops.
There isn’t one migration path, and there doesn’t need to be. The more interesting point is what 800 VDC tells us about AI infrastructure itself. Rack density is becoming high enough to change decisions that used to sit much further down the data centre design process.
Power, cooling, storage, protection and compute are becoming harder to plan separately because each increasingly affects the choices available to the others. That shift is still unfolding. And as AI continues changing the physical infrastructure behind enterprise computing, EM360Tech will keep following where those changes lead.
Comments ( 0 )