WHAT GOES IN TO POWERING AI DATA CENTRES WITH DC, AC AND HIGH-VOLTAGE INFRASTRUCTURE

Article contents

Key Insights:

• AI is pushing data centre power consumption higher, with rack loads moving from tens of kilowatts into the hundreds and eventually towards megawatt-scale systems.
• Higher-voltage DC can reduce current and electrical losses, helping manage the power demands of increasingly dense AI racks.
• AC, DC and hybrid architectures each have a role, with the right choice depending on the facility, existing infrastructure, equipment and maintenance requirements.
• PUE doesn’t tell the whole story of data centre energy efficiency. Moving conversion losses to another part of the system can change how energy is classified and affect the headline PUE figure.
• Power infrastructure needs to be designed with future rack densities in mind, rather than leaving the facility to retrofit around higher loads later.

What does it take to power an AI data centre?

Walk into an AI data centre today and the first thing that stands out is the density. Rack after rack is being filled with increasingly powerful compute, driving data centre power consumption far beyond the loads many facilities were originally designed around.

But all that power has to get there first. Before a GPU ever starts crunching numbers, electricity has travelled through the grid, transformers, switchgear, UPS systems and distribution equipment. 

At every stage, the infrastructure has to handle a larger load without introducing unnecessary losses or compromising resilience. That puts more pressure on the way power is distributed through the facility, from the grid supply to the equipment at rack level.

So, how will AI data centres be powered as density, and the demands placed on electrical infrastructure continue to climb? We’re going to follow the power from the grid connection to the rack, looking at the role of AC, DC and hybrid architectures along the way - and what each approach means for infrastructure design.

How does power actually reach an AI data centre?

Before comparing AC and DC, it helps to look at what actually happens to the power on its way to the rack. Electricity doesn’t arrive at a server in one neat step, it passes through a chain of infrastructure - each stage changing its voltage, controlling its flow or conditioning the supply:

Every link in that chain has a job, but it all comes with a footprint. As rack densities rise, there’s less room to absorb the infrastructure needed to get power where it needs to go.

Bringing high-voltage power onto site

Larger data centres increasingly need high-voltage utility connections simply to bring enough capacity onto site. That power then passes through substations and transformers, which step voltage down to levels the facility can use.

The important bit is that this infrastructure cannot be designed in isolation. The incoming supply, substation and transformers all need to account for the loads the facility is ultimately expected to support. 

Get that capacity or architecture wrong, and the facility is constrained before a single rack is even installed.

Protecting and conditioning the power supply

Once the power is on site, it needs to be protected and conditioned. Switchgear allows circuits to be controlled and isolated, while UPS systems and energy storage can maintain the supply when the grid drops out or fluctuates.

That resilience is becoming increasingly important for AI infrastructure. Higher-density compute means more power concentrated in fewer racks - so an interruption can affect a substantial amount of computing capacity all at once.

Getting power from the electrical room to the rack

From there, power has to make its way through the facility to the rack. Distribution boards, PDUs, busways and cabling move it closer to the IT load, with further voltage changes and AC/DC conversions occurring along the way.

Those conversion stages aren’t free. Each one introduces another point of electrical loss and another piece of infrastructure into the power path - all of which take up valuable space. In a high-density AI factory, that creates a direct trade-off between the infrastructure needed to deliver power and the space available for compute.

AI is putting traditional power architecture under pressure

For years, the basic power path in a data centre has been fairly well established: AC comes in from the grid, gets transformed and distributed through the facility, then converted again as it gets closer to the IT equipment. 

It works - but AI is asking that same architecture to deal with a very different kind of load:

1. More power is being consumed across the facility

AI is pushing data centre power consumption higher at facility level. More compute means a bigger overall electrical load, putting greater demands on everything from the utility connection to the distribution infrastructure.

2. More power is being packed into each rack

Then there’s the rack itself. AI racks are moving from tens of kilowatts into the hundreds, with the industry already looking towards racks measured in megawatts. The footprint might not have changed much, but the amount of power being pushed through it certainly has.

And that has a knock-on effect throughout the electrical system. More current means bigger conductors and more substantial electrical equipment. There’s also only so much room in a facility, so giving more of it over to power infrastructure leaves less space for the thing the whole building exists to support: compute.

There’s another issue too… 

Every time power is converted, some energy gets lost along the way. That might not look like much at lower loads, but start pushing hundreds of kilowatts or more through the same power path and it soon adds up - affecting both efficiency and the amount of heat the facility has to deal with. 

That’s where DC comes in: if power can be converted fewer times and delivered more directly to the load, there’s potential to cut some of those losses and simplify parts of the power path. 

We’re not talking about ripping AC out of the facility - just that there may be a more direct way to get power to increasingly hungry AI hardware.

What changes when more of the power system moves to DC?

A move to DC changes where the conversion happens in the power path. Instead of converting power at several points on its way to the rack, some of those conversion stages can be consolidated or removed altogether.

That means fewer opportunities to lose energy as heat, while potentially reducing some of the infrastructure needed to get power to the IT load.

There’s also a natural fit with battery storage and many forms of low-carbon generation, which produce DC power. Keeping more of the system in DC can mean fewer conversions between generation, storage and the IT load.

But DC comes with its own engineering headaches. Equipment and standards aren’t as mature as they are for AC, and maintenance teams need to be comfortable working with a different type of electrical system.

So DC has a place in the mix, but it’s not a case of switching everything over and calling it done. The right architecture will depend on the facility, the load and how the power system needs to operate.

Why higher-voltage DC is attracting attention for AI

This is where 800 VDC and other high-voltage DC architectures come into play. The basic relationship is straightforward: push the voltage higher and you need less current to deliver the same amount of power.

That becomes a much bigger deal when racks are pulling hundreds of kilowatts - and with platforms such as NVIDIA Rubin set to push rack power higher still, those demands are only going up. At lower voltages, the current gets very high very quickly, which means larger conductors and connectors are needed to carry it. At 800 VDC, the same amount of power can be delivered at a much lower current, reducing the need to keep increasing conductor sizes as rack loads rise.

There’s also the question of where all the power equipment sits. Sidecars and standalone power units can take some of that hardware out of the compute rack, freeing up space for IT equipment. And as rack power continues to climb, higher DC voltages could become more common. 

Of course, higher-voltage DC brings its own engineering requirements around protection, grounding, isolation and maintenance. The complexity is still present, just in different parts of the system - but for the highest-density AI facilities that can be a trade-off worth making.

AC, DC or hybrid?

Efficiency is only part of the equation when choosing between AC and DC.

The choice also depends on the facility around the power system. A new-build data centre has more freedom to design around DC from the outset, while an existing facility may already have substantial AC infrastructure in place. That’s where hybrid approaches can make sense, using DC closer to high-density compute while AC continues to serve other parts of the facility.

There’s also the practical side of running the system. Equipment availability and supply-chain maturity need to be weighed alongside the efficiency gains. There’s little point specifying an architecture that looks great on paper if the equipment is hard to source or the maintenance team can’t support it.

Ultimately, the architecture needs to suit the loads it has to handle and the people who’ll be operating and maintaining it day to day.

If DC uses less energy, will it actually improve PUE?

PUE data centre measurement is usually expressed simply: total facility energy divided by IT equipment energy. So if DC cuts conversion stages and reduces electrical losses, it’s easy to assume PUE should improve too. But unfortunately, it doesn’t work that way.

The reason comes down to where those losses are counted:

• In a traditional AC setup, some power conversion happens inside the server or rack, and that energy is usually counted as part of the IT load.
• Move that conversion somewhere else, such as a sidecar or separate power unit, and the same losses can be counted as facility energy instead.

So the facility could use less energy overall, while its PUE figure actually gets worse. The energy hasn’t necessarily gone up - it’s just being counted in a different part of the calculation.

That’s why data centre energy efficiency can’t be judged on PUE alone. You also need to look at total energy use and make sure the same metering boundaries are being used when comparing different power architectures. 

A lower PUE doesn’t automatically mean a more efficient facility, and a higher one doesn’t automatically mean a less efficient design.

Power architecture has to be designed around the AI factory

We opened by asking how will AI data centres be powered as rack densities keep climbing. After following the power from the grid to the rack, the answer is pretty clear: there’s no single architecture that works for every AI facility.

What works will depend on the available grid connection, cooling system, resilience requirements, existing electrical infrastructure and, crucially, where rack densities are heading next.

That’s why AC, DC and hybrid architectures all have a place. The engineering comes down to working out where each one fits, how the different systems interact, and what the facility needs to support over its lifetime.

That’s what we do at RED. We look at the power, cooling and infrastructure as one system, then work through the practical trade-offs to find an architecture that works for the facility and the hardware it needs to run.

Because, as we’ve explored here, and in our powering the AI factory piece, there’s not much point designing a power system for today’s racks if tomorrow’s racks have already outgrown it.

Talk to RED about designing power infrastructure for high-density AI environments.