Schneider Electric warns AI data centres face power squeeze

Schneider Electric has outlined its approach to power distribution for AI data centres, arguing that rising electricity demand is becoming a central constraint on AI deployment.

Conventional data centre electrical systems were built for lower rack densities and cannot readily support the sharp increase in power needs from AI training and inference workloads. Schneider Electric said rack densities are rising from about 10 kW in conventional facilities to more than 140 kW in high-density AI pods, while single AI training sites can approach 1 GW of demand.

That shift reflects a broader change in data centre design. Traditional enterprise and colocation sites typically run pods of up to 200 servers at around 10 kW per rack, with total site loads near 10 MW, while larger cloud facilities can exceed 50 MW. AI facilities differ because graphics processing units draw far more power per server than central processing units and place heavier demands on cooling systems.

Schneider Electric identified three converging pressures on operators: capacity, reliability and distribution. Utilities in a number of regions are already struggling to connect large new data centre loads within workable timescales, while the economics of AI workloads make power interruptions more costly.

According to Goldman Sachs Research, global data centre power demand could be 175% higher in 2030 than in 2023, with AI accounting for roughly 19% of total data centre demand by 2028. ITIC research cited by Schneider Electric found that 44% of enterprises estimate the cost of a single hour of downtime at USD $1 million to USD $5 million.

Cooling requirements are also reshaping electrical design. Schneider Electric said cooling load is projected to rise from 12-16 kW per square foot in current cloud environments to 80-120 kW per square foot in AI training centres, requiring far more power distribution close to the rack.

Architecture shift

Schneider Electric described three power architectures now in use across the market. Conventional systems step utility medium-voltage power down to 480 VAC and distribute it through switchboards, uninterruptible power supplies and busway to racks drawing less than 10 kW.

Large cloud providers using Open Compute Project designs move AC-to-DC conversion to the rack at 48 VDC, improving efficiency by two to three percentage points across very large sites, according to Schneider Electric. For AI workloads, the company highlighted a newer Sidecar design that moves AC-to-DC conversion and battery backup out of the rack and into an adjoining electrical panel, converting incoming AC to direct current at up to 800 volts.

According to Schneider Electric, this approach allows more IT equipment to be installed in each rack, reduces cabling constraints and separates power capacity management from compute capacity. It added that higher-voltage distribution closer to the rack is becoming necessary because conventional low-voltage architecture cannot support the projected density of AI training environments.

Prefabrication is another part of the shift. Factory-built skids and pods that combine uninterruptible power supplies, switchgear and batteries are increasingly being shipped to sites in transportable frames rather than assembled conventionally on location.

Schneider Electric said this can cut deployment timelines by 30% or more and reduce costs by more than 15% compared with traditional designs. It also pointed to a co-engineered GB300 NVL72 reference design with NVIDIA that supports rack densities of up to 142 kW and includes integrated power management and liquid cooling controls.

Labour squeeze

Schneider Electric said the role of electrical contractors is widening as AI infrastructure scales, with customers seeking expertise in high-density power distribution, system scalability and the interaction between electrical and cooling systems. That demand comes as the industry faces a shortage of skilled workers.

Data from the Uptime Institute's Global Data Centre Survey, cited by the company, found that 53% of operators report difficulty finding qualified employees, up from 38% in 2018. Schneider Electric said experienced specifiers and installers are leaving the sector faster than they are being replaced, pushing more projects towards modular and factory-integrated systems that require less on-site labour.

The company also positioned its product range as part of an integrated grid-to-chip model for AI power systems, covering medium-voltage switchgear, transfer switches, UPS units, switchboards, busway and modular pod systems. It added that its EcoStruxure Pod Data Centre for AI and Accelerated Compute can be assembled in a single day for up to 42 racks at 132 kW per rack.

A monitoring and service layer sits over that hardware, including PowerLogic metering and EcoCare services, which Schneider Electric said can reduce unplanned downtime by up to 75%.

Commenting on the shift in customer demand, Marta Asack, Senior Vice President, Power Products North America, Schneider Electric, said: "The AI era has fundamentally changed what our customers need from us. It is no longer enough to supply great products. Data center stakeholders need an energy technology partner who can deliver the full powertrain, grid-to-chip, chip-to-chiller, with the reference designs, prefabrication, and services that let them capture AI revenue faster."