The hidden power infrastructure behind AI

AI’s biggest bottleneck isn’t compute, it’s power. As rack densities climb from around 10 kW in conventional data centers to more than 140 kW in high-density AI pods, and single AI training facilities approach 1 GW of demand, a real constraint on AI deployment is the electrical infrastructure that delivers power from grid-to-chip. Conventional data center power architecture cannot keep up, and AI power infrastructure is evolving faster than anything the industry has built before.

From cloud data centers to AI factories

Traditional data centers were designed around general-purpose cloud workloads with lower rack-level power requirements. AI training and inference computing work differently; instead of central processing units, AI servers run graphics processing units, which draw far more power per server and require correspondingly higher cooling loads.

The scale shift is substantial. A conventional enterprise or colocation data center typically runs pods of up to 200 servers at roughly 10 kilowatts per rack, with total site loads around 10 MW. Cloud data centers built on Open Compute Project designs can exceed 50 MW. A single AI training facility could exceed 1 GW, more than the electricity demand of some small countries.

An AI factory runs compute at sustained utilization in a way a conventional data center cannot, which is exactly why the power system underneath it has to change.

Why AI power infrastructure is the new bottleneck

Three forces are converging to make AI power infrastructure the constraint that matters.

First, capacity. Goldman Sachs Research projects global data center power demand could be 175% higher in 2030 than in 2023, with AI accounting for roughly 19% of total data center demand by 2028. Utilities in many regions are already struggling to interconnect new data center loads along usable timelines.

Second, reliability. AI workloads are uniquely punishing when disrupted. Training runs that span days or weeks must restart from a prior checkpoint if power fails mid-iteration. ITIC’s Hourly Cost of Downtime research found that 44% of enterprises estimate a single hour of downtime at $1 million to $5 million, a figure that only climbs as facility scale grows.

Third, distribution. Cooling load alone is projected to rise from 12–16 kW per square foot in today’s cloud environments to 80–120 kW per square foot in AI training centers, a 5- to 10-fold increase. Delivering that volume of power with conventional low-voltage architecture is structurally impossible. Higher-voltage distribution, moved closer to the rack, is the only path forward.

At GTC 2026, NVIDIA CEO Jensen Huang reframed the economics of this shift with a formula the electrical industry should pay close attention to: Revenue = Tokens per Watt × Available Gigawatts. Every inefficiency in the power system reduces tokens per watt, which means power infrastructure is no longer a support function for AI, it is one of the primary levers on AI revenue.

Rethinking power architecture for AI workloads

Three architectures now coexist in the data center market, and each one solves a different part of the problem.

Conventional architectures, used in enterprise and colocation sites, step utility medium-voltage power down to 480 VAC and distribute it through switchboards, UPS, and busway to racks drawing under 10 kW. Open Compute Project architectures, adopted by large cloud providers, move AC-to-DC conversion to the rack at 48 VDC, gaining two to three percentage points of efficiency across very large footprints.

For AI, a newer approach called Sidecar architecture moves the AC-to-DC conversion and battery backup out of the rack entirely, converting incoming AC to direct current at up to 800 volts in an adjoining electrical panel. The result is more IT equipment per rack, fewer cabling constraints, simpler maintenance, and power capacity that can be managed independently of compute capacity.

The architectural shift does not stop there. Prefabrication is becoming central. Factory-built skids and pods containing compatible UPS, switchgear, and battery equipment are moving to site in transportable frames, cutting deployment timelines by 30% or more and costs by over 15% compared with conventional designs. Schneider Electric and NVIDIA’s co-engineered GB300 NVL72 reference design, supporting rack densities up to 142 kW with integrated power management and liquid cooling controls, is one example of how reference designs open specifications for AI factories, helping the industry move faster.

The expanding role of electrical contractors and specifiers

As AI infrastructure scales up, the role of electrical contractors is expanding from installation into strategy. Data center clients look for partners who understand high-density power distribution, system scalability, integration between electrical and cooling systems, and the trade-offs between efficiency, resilience, and cost.

There is also a workforce dimension. The Uptime Institute’s Global Data Center Survey found that 53% of operators report difficulty finding qualified employees, up from 38% in 2018, and staffing has consistently ranked as operators’ leading operational requirement. Experienced specifiers and installers are leaving the industry faster than they are being replaced, even as the pace of AI deployment accelerates. That combination is pushing demand toward prefabrication, modular deployment, and factory-integrated systems that reduce on-site labor without sacrificing quality.

Contractors and specifiers who build this expertise now will be positioned to win the most complex, highest-value AI infrastructure work for the next decade.

Powering AI with an energy technology partner

This is why Schneider Electric is focused on the full grid-to-chip powertrain, not one product at a time, but as an integrated system designed specifically for the economics of AI.

Our response covers every layer of AI power infrastructure: SureSeT and GHA medium-voltage switchgear at the grid edge; ASCO 7000 Series transfer switches backed by 100 years of ASCO transfer switch leadership; Galaxy VX and Galaxy VL UPS with eConversion mode running at up to 99% efficiency; QED-2 switchboards distributing power from the grid edge to server pods; I-Line Track busway rated up to 1,000 A; the EcoStruxure Pod Data Center for AI and Accelerated Compute, assembled in a single day for up to 42 racks at 132 kW per rack; and the EcoStruxure platform tying it all together with AI-enhanced monitoring, PowerLogic metering, and EcoCare services that can reduce unplanned downtime by up to 75%.

“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.” —Marta Asack, Senior Vice President, Power Products North America, Schneider Electric

The infrastructure that powers AI may be hidden, but it determines whether AI can scale. For a deeper technical walkthrough of Schneider Electric’s end-to-end approach, including the NVIDIA GB300 reference design, Sidecar architecture, and the full grid-to-chip portfolio, read the data bulletin: From Grid to Chip: Scalable Power Distribution for AI Data Centers.