A complete engineering overview of how power moves through an AI data center — from utility interconnection and onsite generation through medium-voltage distribution, UPS and battery backup, PDU/busway distribution to the rack, and the redundancy architecture that keeps GPU clusters running.
Why AI Data Centers Are an Electrical Design Problem First
A traditional enterprise data center is, electrically speaking, a fairly modest load — a few megawatts spread across racks of general-purpose servers that rarely all run at once, pulling maybe 5-10 kW per rack at peak. An AI training or inference facility is a completely different animal. A single rack of GPU servers can draw 40-130+ kW, an entire hall can demand 20-100+ MW, and — critically — that load doesn't ramp gently. A large training job starting or checkpointing can swing megawatts of demand in seconds, something conventional data center electrical design was never built around. That combination — enormous continuous load plus violent step changes — is why AI data center electrical design has become its own specialized discipline rather than a scaled-up version of traditional data center engineering.
This guide walks through the electrical system from the utility connection down to the rack, in the order power actually flows, and links out to deeper articles on each stage. If you're new to why data centers need so much power in the first place, start with Why IT Load Equals Heat Load in a Data Center — every watt of electrical power a GPU consumes becomes a watt of heat the cooling system has to remove, which is the single fact that ties this guide's electrical content to the studio's existing cooling-focused articles.
Utility Interconnection: Where the Power Actually Comes From
Everything starts at the point of interconnection with the local utility. For a facility drawing tens of megawatts, that almost always means a dedicated substation feed at transmission or sub-transmission voltage (commonly 115 kV, 138 kV, or 230 kV depending on the region and the utility's own grid), stepped down on-site by the facility's own main power transformers to a medium-voltage distribution level — typically 13.8 kV or 34.5 kV — before it's distributed across the campus. The utility interconnection is frequently the actual bottleneck on how fast a new AI data center can come online: securing a multi-hundred-megawatt grid connection can take years of utility studies and queue time, which is a large part of why onsite generation (covered later in this guide) has become such an active area of AI data center design rather than a fallback option.
Utility power alone is rarely treated as sufficient for a facility of this scale — nearly every large AI data center also has some form of onsite backup or supplemental generation, and increasingly, onsite generation is being designed to run in parallel with the grid connection rather than purely as an emergency backup.
Medium-Voltage Distribution: Getting Power Across the Site
Once power is stepped down to medium voltage at the main substation, it's distributed across the site via medium-voltage switchgear and underground or overhead feeders to a series of unit substations positioned near each data hall or pod. Each unit substation steps that medium voltage down again — typically to 480V three-phase in North America — which is the voltage level most UPS systems, PDUs, and ultimately the rack-level power distribution actually operate at. This two-stage step-down (utility voltage → medium voltage → 480V) exists because transmitting power at higher voltage means lower current for the same power, which means smaller, cheaper conductors and lower resistive losses across the long distances a hyperscale campus can span.
The segregated architecture used at this stage — typically labeled "A side" and "B side" — is worth understanding early, because it shapes every downstream design decision. Critical IT loads are fed from two electrically independent paths from the utility connection all the way to the rack, so that a fault or maintenance event on one path never interrupts power to a properly-configured dual-corded server. See N+1 and 2N Redundancy: What the Terms Actually Mean for how this A/B segregation concept extends into the specific redundancy levels (N+1, 2N, 2N+1) used to describe both electrical and cooling system design.
UPS and Battery Backup: Bridging the Gap
Utility power, even with onsite generation as backup, is not instantaneous — a generator takes seconds to start and reach stable output after a utility outage, and even automatic transfer switching takes a brief but real amount of time. The uninterruptible power supply (UPS) exists specifically to bridge that gap: it sits electrically between the medium-voltage distribution and the IT load, continuously online, so that a utility outage causes zero interruption to the GPUs while generators spin up. Modern AI data centers overwhelmingly use double-conversion (online) UPS topology — converting incoming AC to DC and back to AC continuously, not just standing by — because it also provides clean, regulated power that filters out utility voltage sags, surges, and the AI cluster's own load-swing noise from ever reaching sensitive electronics.
The battery portion of the UPS (increasingly lithium-ion rather than traditional VRLA/lead-acid, for its smaller footprint, longer cycle life, and faster response) is sized for a short ride-through window — often just 3-10 minutes for a data center, since the generators are expected to be online and carrying load well within that window, not sized to run the facility for hours. This is a fundamentally different sizing philosophy than a residential or small-commercial backup battery system, which is often sized to carry a full outage.
Onsite Generation: No Longer Just a Backup Plan
Backup diesel generators have been standard in data center design for decades — sized to N+1 or 2N redundancy, tested regularly, and rarely run except during an actual utility outage or scheduled test. What's changed for large AI facilities is that onsite generation is increasingly being designed as a genuine part of the primary power supply, not just an emergency system, because utility interconnection capacity and timeline have become the binding constraint on how much AI compute a facility can actually deploy. Natural gas reciprocating engines and gas turbines are both seeing wider adoption specifically because they can be brought online faster than a new utility transmission connection, and because they can run continuously in parallel with (or ahead of) the grid connection rather than sitting idle. Fuel cells, and even behind-the-meter small modular reactors (SMRs), are being explored by some operators for the same underlying reason: grid capacity is scarce, and onsite generation shortens the path to megawatts.
This shift toward primary-role onsite generation, combined with battery energy storage systems (BESS) used for demand-shaving and fast-response duty rather than just backup, is enough of a departure from traditional data center design that it's covered in its own dedicated articles in this cluster rather than as a subsection here.
Power Distribution to the Rack: PDUs and Busway
Inside the data hall itself, 480V power from the UPS output is stepped down one final time — typically to 208V three-phase — by power distribution units (PDUs), then routed to individual racks either via traditional whip cabling or, increasingly for high-density AI deployments, overhead busway. Busway's advantage at GPU rack densities is largely about current capacity and flexibility: a busway run can be tapped at any point along its length to add or relocate a rack power drop without pulling new conductors, which matters enormously when a facility's rack layout and power density assumptions are likely to change as GPU generations evolve. At the rack itself, redundant power distribution — typically dual power feeds from the A and B distribution paths described earlier — feeds redundant power supplies in each server, completing the fault-tolerant path from utility all the way to the GPU.
The Load Itself Is the Hard Part
Everything above describes a fairly conventional (if very large) power distribution architecture. What makes AI data center electrical design genuinely different from a traditional data center is the character of the load itself. A large GPU training cluster's power draw isn't steady — it fluctuates sharply based on what the training job is actually doing at any given moment, with the sharpest swings occurring during synchronized events like checkpointing (when thousands of GPUs briefly pause compute to write state to storage) or a job starting/stopping across a whole cluster simultaneously. These aren't small ripples; megawatt-scale swings within seconds are realistic at hyperscale, and they stress everything from UPS response time to generator governor control to the utility interconnection's own voltage stability. Electrical engineers designing for AI data centers now have to account for load volatility as a first-class design constraint, not an edge case — which is a major reason utilities and facility designers alike are paying close attention to grid-forming inverter technology and fast-responding BESS as tools for smoothing exactly this kind of swing.
Where to Go Deeper
This guide is the entry point into the AI Data Center Engineering cluster. From here:
For hands-on tools, the studio's Rack Power Calculator, PUE Calculator, and Cooling Load Calculator let you work through real numbers against the concepts in this guide.