Powering the AI mega-rack: how DC-DC conversion is evolving for AI data centres

Advanced Energy Director Harry Soin tells Power Electronics how the shift to 800 V DC architectures is reshaping power conversion in AI data centres.

With three generations of AI computing platforms in as many years, NVIDIA has rapidly increased compute density while driving a transition toward rack-scale 800 V DC power architectures.

At the same time, hyperscalers including Google, Meta, and Microsoft are collaborating on Open Compute Project (OCP) initiatives, such as Mount Diablo, which re-architect data centre power distribution to provide a pathway to megawatt-scale AI racks at higher voltages. The impact on power conversion has been profound with industry players developing systems set to provide higher-voltage DC-DC conversion and greater efficiencies at megawatt-scale rack power levels.

A case in point is the ADH series of DC-DC converters from US-based Advanced Energy, introduced in June this year. Targeting 800 V DC AI data centre power architectures, these DC-DC converters can convert that 800 V DC input into 50 V DC output and operate at multi-megahertz switching frequencies. Devices are said to achieve a peak power efficiency of 98.2%, helping to reduce power losses in high performance computing applications. Each converter can deliver up to 8 kW of peak power and 6 kW of continuous full-load power within a standard half-brick package.

Keeping cool: a Google data centre aisle, with coolant distribution units, at New Albany, Ohio. [Google]

“From the Grace Hopper to Vera Rubin Ultra NVIDIA platforms, power is essentially doubling and quadrupling, so currents are increasing by huge amounts and power dissipation losses are significantly higher,” highlights Harry Soin, Senior Director, Hyperscale, at Advanced Energy. “To keep up with this, we transitioned from 12 V to 48 V outputs... and with industry now moving towards 800 V architectures, we have introduced our latest converters.”

Megawatt power racks
The transition to 800 V DC has been rapid. In 2022 standard power shelves from Advanced Energy handled 18 kW but within that same year, capacity swiftly rose to 33 kW. This year, the company will be releasing a 72 kW power shelf, which can be either combined or used in an external sidecar configuration to reach the 800 kW to full megawatt power racks data centres are now demanding.

But once megawatt scale is well and truly reached, traditional power shelves hit a bottleneck; distributing this much power through standard low-voltage copper busbars requires them to become prohibitively thick to avoid over heating. Given this, next-generation power rack architectures, such as NVIDIA’s Kyber platform, designed to scale to the massive NVL576 architecture with 576 connected Rubin Ultra GPUs, removes the power shelf from the main cabinet and instead runs a 800V DC power line directly into the server.

“GPU suppliers are working on a solution where essentially, the high voltage goes directly to the server - this is a revolutionary move, right?” says Soin. “This means the server itself needs power conversion.”

The NVIDIA Rubin platform: The AI supercomputer architecture is designed to power massive agentic AI. [NVIDIA]

According to Soin, Advanced Energy’s DC-DC team is currently working on Power Distribution Boards (PDBs) to do just this. These will house 4 kW to 6 kW DC-DC converters in a standard brick package – and four of these bricks can be placed onto the server’s PDB. This will then step down the incoming 800 V to low voltages at the board level, to power the GPUs, such as Rubin Ultra.

As the demand for high power density AI infrastructure - that generates hugely concentrated thermal loads - continues to grow, the move from air cooling to liquid cooling will be critical for AI workloads. As Soin points out, he and colleagues have been focused on developing DC-DC converters for the various GPU architectures that link tens and tens of GPUs together.

However, the architectures that accompany the Vera Rubin and upcoming Kyber platforms double the computing density of existing systems. This extreme density is necessary to power the massive agentic AI models that process gargantuan levels of data.

“We’re now working with less space for the power and we need for these high-density platforms to be more efficient... If you can’t get the heat out, you cannot provide the power, so managing heat with direct liquid cooling has become key, especially for architectures like the Kyber platform,” says Soin.

With this in mind, Advanced Energy has designed its board-mounted DC-DC converters so that liquid cooling can be directly integrated via a baseplate. “By tapping into the liquid plumbing used by all the hyperscalers and chip suppliers, including AMD and NVIDIA, we can directly remove the heat,” explains Soin.

“We used to be working with 1 to 2 kW in a single brick board mount power module but are now working with 6 to 8 kW modules and are considering 12 kW modules – this is only possible with direct liquid cooling,” he adds.

Cooling aside, Soin and colleagues at Advanced Energy are working with both silicon and wide bandgap materials to take the peak power efficiency of their devices beyond 98.2%. Soin says they have embraced silicon carbide semiconductors, thanks to solid reliability reports, and are now using gallium nitride as more and more reliability data becomes available.

“We are looking at incremental [increases] in efficiency … and the only way to move forward and reach [these] now is to move beyond silicon,” he says. “We’re removing the parasitics and the switching losses. After that, it’s going to be a race for the conduction losses and this can only be achieved with the lower on-resistance found in wide bandgap devices.”

Market moves
The Vera Rubin Ultra platform with its multi-chip GPU design is, according to NVIDIA, expected to launch in the second half of 2027, at which point Advanced Energy anticipates rising demand for its ADH series of DC-DC converters. As Soin puts it: “We’re developing solutions for hyperscalers as well as the GPU designers -– these guys are leading this revolution, right?”

“We are designing solutions and reference designs that could be part of their ecosystem and will essentially be deployed,” he adds. “Public information indicates that Vera Rubin Ultra will be launched in huge quantities in the middle of next year, and that’s exactly when you will see our [devices] verified and implemented – they will be needed to address the power needs of those upcoming architectures.”

As today’s data centres transition to 800 V DC via additional sidecars to meet rising AI energy demands, Soin reckons native 800 V infrastructure will come online around mid-2029. “When that 800 V goes directly to the server and [delivers] higher efficiencies – that really will shape the future.”

Soin is also clear that industry will not stop at 800 V. “If power keeps increasing as it is today, then we could approach 1500 V DC in the future,” he says. “This is a really exciting time.”