From 400kW to 1MW per rack: Schneider Electric’s Steve Carlini on the next phase of AI infrastructure

AI racks are set to climb from around 400kW to at least 1MW as data centres move to 800V direct current (DC) power architectures, according to Steve Carlini, chief AI advocate at Schneider Electric.

Speaking to Capacity, Carlini said 800V DC will arrive alongside NVIDIA’s Rubin Ultra platform in 2027, with hyperscalers first in line before the technology filters down to colocation providers and neoclouds.

He also expects behind-the-meter power to dominate the largest US AI campuses, liquid cooling to sharply cut data centre water use, and faster utility grid queues to be a potential tipping point towards the end of 2027.

Carlini’s role is to track the business and technology trends reshaping the data centre market, from edge computing and sustainability to AI, and feed them back into how Schneider supports customers. He also spends a lot of time with NVIDIA, which Schneider works with closely. Schneider publishes reference designs for each new NVIDIA cluster around six months before it launches, giving operators time to build out facility-level power and cooling.

“With NVIDIA hitting the industry with a cattle prod, things are changing very, very fast,” he said.

Why 800V DC, and why now

The shift to 800V DC is a response to physical limits as traditional AC power distribution can no longer feed the densest GPU racks.

“We’ve reached the physical limits,” Carlini said. “We were running out of room for the size of the copper to be able to power each server.”

The fix is to take the power supplies out of the servers altogether and move them into a sidecar, which Schneider brands as its NetShelter power rack. That rack also holds batteries and capacitors to smooth out sharp swings in power draw from GPU clusters. The compute racks are left holding only server boards and GPUs, which can then be packed far more tightly.

“The way AI operates, you want to put the GPUs physically as close together as possible,” he said. “That’s what’s driving the densification.”

There will not be a single standard. NVIDIA’s design runs from 0V to 800V, while Mt Diablo, which Schneider is working on with Google and the Open Compute Project, uses plus and minus 400V. Battery and capacitor configurations also differ between versions.

“There’s not one 800V DC solution that’s going to fit everybody,” Carlini said.

Hyperscalers will be the first adopters, largely because they will hold the initial allocation of Rubin Ultra systems. Carlini expects colocation providers and neoclouds to follow, though probably not within the first year.

Bring your own power

On the largest US projects, grid power is increasingly the exception. Carlini said gigascale developments are sourcing most of their power on site or from adjacent generation, with behind-the-meter supply ranging from 50% to 100% of total load.

“They just can’t get the power from the grid,” he said.

Schneider is seeing growing demand for what it calls energy parks, built mainly around natural gas turbines. With 50MW turbines more readily available than larger units, developers are paralleling them into 200MW blocks. Schneider does not make the turbines, but supplies the switchgear and the software that controls power from these sites.

Carlini sees little sign of demand cooling. Schneider has supply chain agreements with the hyperscalers, neoclouds and large colocation providers, covering switchgear, UPS systems, busways and power distribution.

“Their next two to three years, they’ve secured capacity,” he said.

If overbuild does happen anywhere, he expects it to be the US, simply because it is building more than any other market. “We don’t see that happening,” he added.

He described the current period as a year of transition for AI. Training and scientific research will still take capacity, but more inference workloads are rolling out, which is where AI companies start to monetise. Enterprises are also moving models from pilot into production, and Carlini expects demand for compute to keep growing as fast as supply allows for the next two to three years.

The end of secrecy

Community opposition has become one of the most cited constraints on data centre buildouts, nowhere more so than in Texas, where residents have pushed back over water use and where projects are being sited.

Carlini said residents have legitimate concerns, and that both operators and the local authorities that granted permits share responsibility.

“There’s definitely been some mistakes made,” he said. Some sites were approved very close to homes, and operators “should have realised it as well”.

Many projects were developed under code names and shell companies. Carlini argued this was mostly about keeping plans from competitors rather than residents, but the effect was the same. Communities often learned a data centre was coming only once construction began.

He believes that approach is finished. Operators will need to disclose noise levels, power use and sourcing, and water use and sourcing up front, and win local approval before permits are granted.

“It’s not acceptable to tap into the town’s water supplies,” he said. “It’s not acceptable for the data centres to drive up utility prices for local residents.”

Traffic and noise are the two biggest complaints, he said. Rising US electricity prices add to the pressure. Carlini said much of that increase is not being driven by data centres, but they are an easy target for residents struggling with bills.

Behind-the-meter generation could help change that narrative. Carlini said operators are increasingly planning to generate surplus power that can eventually be shared with the grid and local communities.

Hotter chips, less water

Carlini also pointed to a major shift in cooling. NVIDIA’s Rubin systems are designed to run liquid cooling at 45°C (113°F), higher than earlier generations.

That means many new liquid-cooled facilities can rely on dry coolers almost all year round, depending on climate, instead of chillers. Evaporative cooling and water-spray assists can be removed, and mechanical compression kept to a minimum.

“You’re going to see a dramatic shift in water use in data centres going forward with liquid cooling,” he said.

Operators are also moving away from drinking water supplies, recycling wastewater and treating it to the standard needed for liquid cooling loops.

“It’s a whole changing paradigm in the cooling architecture for data centres,” he said.

The UK: a long queue, but room to move

The UK is often described as lagging in AI infrastructure. Carlini said the bigger story is the size of its grid connection backlog, though much of that queue may not be real.

He pointed to measures that would require developers to pay to keep their place in the queue, weeding out speculative projects. Combined with areas of the country that have spare power and have been designated for AI development, he believes the UK is better placed than the rest of Europe.

“The UK is in a position that they can move faster,” he said.

Across wider Europe, he said there is a clear desire to move quickly, but regulation remains a hurdle.

Prefab underhyped, Omniverse ahead of itself

Asked which technology will be mainstream by the end of 2027, Carlini pointed to prefabrication. Schneider has been in the prefab business for 20 years, and he said NVIDIA CEO Jensen Huang’s recent keynote has put it firmly in the spotlight.

The case is speed, predictability and quality. Rack densities are changing so quickly that operators struggle to keep up with on-site builds. Factory-built power and cooling systems can simply be dropped into place, while facility work runs in parallel.

Schneider also offers prefabricated IT pods pre-designed and tested for specific NVIDIA and AMD Helios clusters. The pods ship flat-packed and can be assembled two storeys high.

“We’ve seen an exponential increase in our business with prefabrication,” he said.

At the other end of the scale is NVIDIA Omniverse, the digital twin platform for designing and simulating data centres. Schneider-owned software runs its simulation engines, with AVEVA handling cooling and ETAP handling power.

The limitation, Carlini said, is the shortage of SimReady assets, digital models that match real, purchasable products. Once enough are available, operators will be able to design a full facility in Omniverse and generate a bill of materials from it.

“That’s a little bit overhyped now,” he said, “but I think in the future it’s going to be the way the industry goes.”

The tipping point: faster grid queues

Asked about the biggest tipping point of the next 12 months, Carlini pointed not to a chip or a cooling system, but to easing regulation and speeding up utility connection queues in the US and Europe.

Utilities must run impact studies on every large load application, simulating how the grid would cope if part of it failed. Each study can take years. If utilities can speed those studies up while clearing speculative projects out of the queue, he expects waiting times to start shortening. That could become a tipping point, probably towards the end of 2027.

Right-sized AI, measured in tokens

Carlini also expects enterprises to change how they run AI. Many early deployments used frontier models for simple tasks and cost far more than expected. He predicts a move towards right-sized IT stacks and dedicated models built for specific applications.

That shift will change how capacity itself is measured. Carlini expects tokens, the units AI models break text, images and audio into for processing, to join megawatts as a way of sizing data centres.

“It could be a one gigawatt data centre,” he said, “but how many tokens is it good for?”