Data center power is the electrical infrastructure that receives, conditions, backs up, and distributes electricity to every server, switch, and cooling system inside a facility. It includes the utility feed, transformers, uninterruptible power supplies, generators, power distribution units, and all the switchgear and cabling that ties them together. Getting this chain right determines whether your compute stays online or goes dark.
Worldwide data center electricity consumption is estimated to rise from 448 terawatt-hours (TWh) in 2025 to 980 TWh by 2030 (Source: International Energy Agency, 2025). That demand grew 17% year-over-year in 2025 alone, dwarfing the 3% growth in global electricity demand during the same period. AI-optimized servers already account for roughly 21% of total data center power usage in 2025, and that share is projected to hit 44% by 2030 (Source: Gartner, 2025).
If you are planning a modular edge deployment, understanding power architecture is not optional. Our research concept paper on modular edge data centers covers the full picture of cooling, power, and compliance. This article goes deeper on the electrical side.
What Are the Core Components of a Data Center Power System?
A data center power system consists of six primary subsystems that work in series to deliver clean, uninterrupted electricity from the utility grid to the IT load. Failure at any single link can cascade into a full outage. Here are the essential components:
- Utility feed and main switchgear – The incoming medium-voltage service (typically 12.47 kV or 13.8 kV) with automatic transfer switches (ATS) that route power between utility and backup sources.
- Step-down transformers – Convert medium voltage to usable facility voltage, commonly 480VAC three-phase in modern builds.
- Uninterruptible power supplies (UPS) – Condition incoming power, regulate voltage, and provide battery-backed ride-through during transfer events.
- Backup generators – Provide sustained on-site power during extended utility outages, usually fueled by diesel, natural gas, or increasingly, hydrotreated vegetable oil (HVO) blends.
- Power distribution units (PDUs) – Step down and distribute power from the UPS output to individual racks and circuits, with metering and branch-circuit protection.
- Busways and whips – The physical conductors, bus ducts, and flexible connections that route power from PDUs to rack-mounted equipment.
- Monitoring and management software – Platforms like Schneider Electric’s EcoStruxure or Vertiv’s Trellis that provide real-time visibility into load, capacity, and efficiency across the entire power chain.
Every watt your IT load consumes must pass through each of these layers. Losses compound at every stage, which is why power usage effectiveness (PUE) matters so much. The industry average PUE was 1.56 in 2024 and improved slightly to 1.54 in 2025 (Source: Uptime Institute, 2025). That means for every 1.0 W of IT load, facilities consumed an additional 0.54 W on overhead. Google’s global fleet achieved a PUE of 1.09 in 2024, demonstrating what aggressive optimization can accomplish.
How Much Power Does a Data Center Use?
A single data center can consume anywhere from tens of kilowatts for a small edge site to hundreds of megawatts for a hyperscale campus. The answer depends entirely on rack count, rack density, and the ratio of AI to general-purpose compute. There is no single number that fits all deployments.
Here is where rack density becomes the critical variable. Common rack densities in 2025 range from 16 to 30 kW per rack for general-purpose workloads. AI-specific racks tell a very different story.
General-Purpose vs. AI Rack Power Demands
| Parameter | General-Purpose Rack | AI / GPU Rack (2025) | Next-Gen AI Rack (2026-2027) |
|---|---|---|---|
| Typical Power per Rack | 10 – 30 kW | 100 – 240 kW | 250 – 900 kW |
| Approximate Cost per Rack | ~$500,000 | ~$3.9 million | TBD |
| Infrastructure Build Cost | $20,000 – $50,000/rack | $200,000 – $300,000/rack | Higher |
| PDU Amperage Needed | 30A – 60A | 100A+ | Significantly higher |
| Cooling Approach | Air-cooled, raised floor | Direct liquid cooling, rear-door heat exchangers | Full immersion or direct-to-chip liquid |
NVIDIA’s current Blackwell-based GPU servers require approximately 132 kW per rack. The next generation, including Blackwell Ultra and Rubin platforms, is projected to need between 250 and 900 kW with up to 576 GPUs per rack. Google’s Project Deschutes has introduced a 1 MW rack design.
The misconception that all racks have similar power densities is dangerous when planning infrastructure. A facility designed around 10 kW racks cannot simply accept a 132 kW AI rack without a complete redesign of its power and cooling pathways. For edge deployments, this means sizing your electrical infrastructure for the workload you intend to run, not the workload you are running today.
What Is the Role of an Uninterruptible Power Supply in a Data Center?
An uninterruptible power supply in a data center provides continuous, conditioned power to the IT load during the brief window between a utility failure and generator start-up, typically 10 to 30 seconds. Beyond that bridging function, a modern UPS also filters harmonics, regulates voltage sags and swells, and protects sensitive electronics from power quality events that would otherwise cause crashes or data corruption.
The average cost of data center downtime can reach $9,000 per minute (Source: Uptime Institute, 2024). A properly sized UPS is your first and fastest line of defense against that cost.
UPS Topologies for Edge and Modular Deployments
Three common UPS topologies exist in data center environments:
- Standby (offline): The load runs directly on utility power; the UPS switches to battery only when it detects a failure. Transfer time is typically 5-12 milliseconds. Suitable only for non-critical loads.
- Line-interactive: Adds voltage regulation via an autotransformer while the load runs on utility power. Transfer time is shorter (2-4 ms). Common in small edge closets and network rooms.
- Double-conversion (online): The load always runs on inverter-supplied power, with the battery constantly in the circuit. Zero transfer time. This is the standard for any deployment where uptime is non-negotiable.
Modern double-conversion UPS systems have reached impressive efficiency. The Vertiv PowerUPS 9000, for example, offers 97.5% efficiency in double-conversion mode. That may sound marginal until you consider scale: even a 1% to 1.5% improvement in UPS efficiency translates to significant annual energy savings across a facility running megawatts of IT load.
The global data center UPS market is projected to grow from $8.76 billion in 2025 to $12.47 billion by 2030, reflecting a 7.3% compound annual growth rate (Source: Mordor Intelligence, 2025). That growth is driven largely by AI deployments demanding higher-capacity, higher-efficiency units.
Battery Technology Choices
UPS systems are evolving beyond simple backup appliances. Lithium-ion batteries have largely replaced traditional valve-regulated lead-acid (VRLA) in new builds, offering smaller footprints, longer service life (8-10 years vs. 3-5 years), and the ability to support grid-interactive functions like demand response and peak shaving. For edge sites where floor space is at a premium, lithium-ion’s higher energy density is often the deciding factor.
Why Are Data Centers Adopting 480V Three-Phase Power?
480VAC three-phase distribution reduces resistive (I-squared-R) losses by roughly half compared to traditional single-phase 208V distribution, because higher voltage means lower current for the same power delivery. This allows smaller conductor sizes, fewer parallel runs, and less copper throughout the facility.
The transition from 208V single-phase to higher voltages is one of the most impactful efficiency moves a data center operator can make. Here is how the voltage options compare:
Voltage Distribution Options
| Distribution Voltage | Configuration | Relative I-squared-R Loss | Conductor Size | Adoption Trend |
|---|---|---|---|---|
| 208V | Single-phase | Baseline (highest) | Largest | Legacy, declining |
| 277V | Single-phase | ~44% less than 208V | Smaller | Growing in North America |
| 415V | Three-phase (EU standard) | ~75% less than 208V | Significantly smaller | Standard in Europe |
| 480V | Three-phase (US standard) | ~81% less than 208V | Significantly smaller | Rapidly growing in US |
| 600-800V DC (HVDC) | Direct current | Lowest losses | Smallest | Emerging, especially in hyperscale |
For modular edge deployments, 480V three-phase is becoming the default for anything above a single-rack closet. It simplifies the electrical distribution, supports higher-density racks without oversized cabling, and aligns with the input requirements of modern high-capacity UPS systems and PDUs.
High Voltage Direct Current (HVDC) distribution at 600 to 800V is gaining traction in hyperscale environments for even further efficiency gains, eliminating AC-to-DC conversion steps inside the server power supply. Edge deployments may eventually follow, but 480VAC three-phase remains the practical choice for most sub-megawatt facilities in 2025.
How Do Generators Provide On-Site Power for Data Centers?
On-site power for data centers comes primarily from diesel generators that start automatically when the utility feed fails, typically reaching full load within 10 to 15 seconds. During that transition window, the UPS carries the IT load on battery. Generators provide sustained backup for hours or days, limited only by fuel supply.
The global data center generator market is projected at $7.57 billion in 2025 (Source: Mordor Intelligence, 2025). Sub-1 megawatt generators accounted for 45.79% of the market in 2025, with units above 2 MW forecast for significant growth as hyperscale and AI facilities demand more backup capacity. Diesel generators sized around 750 kW remain popular for data center backup at the edge and mid-market tier.
Fuel Flexibility and Sustainability
The assumption that diesel generators are the only viable backup option, and that they are inherently at odds with sustainability goals, does not hold up in 2025. Operators are increasingly adopting fuel-flexible strategies:
- Natural gas generators offer lower Scope 1 emissions and may have more reliable fuel supply via pipeline.
- Hydrotreated vegetable oil (HVO) can serve as a drop-in replacement for diesel in many generator models, reducing lifecycle carbon intensity.
- Hydrogen-ready generators are entering pilot deployments, though hydrogen supply chains remain immature for most edge locations.
- Hybrid configurations pair smaller generators with lithium-ion battery banks, allowing the generator to operate at optimal load rather than idling at partial load during minor events.
For edge sites subject to local emissions regulations or noise ordinances, generator selection must account for more than just kilowatt output. Compliance with fire safety standards is equally critical. Our guide to fire suppression and safety codes for modular data centers covers NFPA 75 requirements that directly affect generator room design, fuel storage, and ventilation.
Generator Sizing for Edge
Sizing a generator for an edge deployment requires accounting for the full facility load, not just IT equipment. You need headroom for cooling systems, lighting, fire suppression, and the inrush current that occurs when HVAC compressors and UPS battery chargers restart simultaneously after a transfer event. A common rule of thumb is to size the generator at 125% to 150% of the expected maximum facility load.
What Is a PDU and Why Is It Critical for High-Density Racks?
A power distribution unit (PDU) in a data center is the final link in the power chain, receiving conditioned power from the UPS and distributing it to individual servers and network equipment within a rack or row. PDUs provide branch-circuit protection, outlet-level or circuit-level metering, and often remote switching capabilities for capacity management.
The data center rack PDU market is expected to grow from $2.78 billion in 2025 to $3.01 billion in 2026 (Source: Mordor Intelligence, 2025), driven by the need for higher-amperage units that can support dense AI racks.
PDU Types
PDUs come in several configurations, each suited to different monitoring and management requirements:
- Basic PDUs: Simple power distribution with no monitoring. Acceptable only for non-critical edge closets.
- Metered PDUs: Provide real-time power consumption data at the unit or inlet level. Essential for capacity planning.
- Monitored PDUs: Add outlet-level metering and environmental sensors (temperature, humidity). Useful for tracking data center cooling efficiency at the rack level.
- Switched PDUs: Include remote on/off control per outlet, enabling remote power cycling of unresponsive equipment without a site visit, which is particularly valuable for unmanned edge locations.
- Automatic Transfer Switch (ATS) PDUs: Accept dual power feeds and automatically switch between them, providing rack-level redundancy.
For modern GPU servers and high-density AI racks, 100A+ PDUs are becoming mandatory. A standard 30A PDU maxes out at roughly 5 kW per phase. When a single rack pulls 100 kW or more, you need multiple high-amperage feeds and PDUs rated accordingly.
PDU Placement: Rack-Mount vs. Floor-Mount
Rack-mount (vertical) PDUs install inside the cabinet and connect directly to equipment via short patch cords. Floor-mount (standalone) PDUs sit in the row and feed multiple racks, often via overhead busway. Edge deployments with limited floor space typically favor rack-mount PDUs for simplicity, while larger modular builds may benefit from floor-mount units that consolidate monitoring and reduce per-rack cabling.
How Is AI Changing Data Center Power Infrastructure?
AI workloads are fundamentally reshaping data center power infrastructure by compressing what was traditionally a megawatt-scale problem into a per-rack problem. A single AI rack at 132 kW demands more power than an entire row of general-purpose servers from just five years ago.
This density shift forces changes at every layer of the power chain. Transformers must handle higher per-circuit loads. UPS systems need to scale without proportional footprint growth. PDUs require amperage ratings that did not exist in standard product lines a few years ago. And generators must cover not only higher total loads but also the significantly higher inrush currents that dense GPU clusters generate on startup.
The average AI rack is projected to cost $3.9 million in 2025, roughly seven times the $500,000 cost of a traditional server rack (Source: Gartner, 2025). Building out the supporting 100 kW electrical and cooling infrastructure runs $200,000 to $300,000 per rack on top of that equipment cost.
For edge AI deployments, these economics create a paradox: the workload demands enterprise-grade power infrastructure, but the deployment model calls for compact, rapidly deployable, and often unmanned sites. Solving this tension requires tight integration between power and cooling systems from the design phase, not as an afterthought. This is precisely the argument laid out in our modular edge data center concept paper, where power and thermal design are treated as a single problem.
Data center cooling, which typically represents the largest non-IT power draw, must also evolve alongside these electrical changes. Facilities targeting AI workloads are moving from traditional air cooling to direct liquid cooling, which shifts power consumption from large air handlers and fans to pumps and coolant distribution units. The net effect can be a meaningful PUE improvement, but only if the electrical infrastructure is designed to support the new cooling topology from the start.
What Does the Future of Data Center Power Look Like?
The future of data center power is defined by higher voltages, tighter integration with renewable energy, and an increasing blur between the utility grid and the data center itself. Facilities will function less as passive consumers of grid power and more as active grid participants, using battery storage to absorb and dispatch energy in response to price signals and grid conditions.
Several trends are already in motion:
- On-site generation beyond backup: Solar arrays, fuel cells, and small modular nuclear reactors are being evaluated as primary power sources, not just backup, for large data center campuses.
- HVDC distribution at scale: 600 to 800V DC distribution eliminates multiple AC-DC conversion stages, reducing losses and component count. Adoption will accelerate as server power supplies standardize on DC input.
- Grid-interactive UPS systems: Lithium-ion UPS batteries can participate in demand response programs, providing revenue and grid stability services while still maintaining their primary backup function.
- Standardized modular power blocks: Pre-engineered, factory-tested power modules (transformer, UPS, switchgear, PDU) that ship as complete units and interconnect on-site. This approach reduces deployment time from months to weeks.
Regulatory pressure is also tightening. The Climate Neutral Data Centre Pact requires new data centers in cool climates to meet a PUE of 1.3 by January 1, 2025, and existing facilities to hit that target by 2030. While this pact applies primarily in Europe, similar standards are likely to emerge globally. Operators referencing the Uptime Institute’s annual PUE benchmarks will want to track these evolving requirements.
The interplay between power infrastructure and environmental regulations extends to cooling as well. The AIM Act’s phase-down of HFC refrigerants is driving adoption of lower-GWP alternatives that may require different compressor technologies and electrical characteristics. And ASHRAE TC 9.9 thermal guidelines continue to shape allowable temperature and humidity ranges, which directly affect how hard cooling systems, and their associated power draw, must work.
Fire protection standards from NFPA 75 also influence power infrastructure layout, dictating separation distances between generators and IT equipment, ventilation requirements for battery rooms, and emergency power-off (EPO) system design. Getting the electrical design right from the beginning avoids costly retrofits when the fire marshal shows up.
Frequently Asked Questions
How much electricity does a data center use annually?
Worldwide data center electricity consumption was estimated at 448 TWh in 2025 and is projected to reach 980 TWh by 2030. Individual facility consumption ranges from tens of kilowatts for small edge sites to hundreds of megawatts for hyperscale campuses, depending on rack count, density, and workload type.
What is the average PUE for data centers in 2025?
The industry average PUE was 1.54 in 2025, according to the Uptime Institute’s annual survey. This represents a modest improvement from 1.56 in 2024. Leading operators like Google achieve significantly lower PUE values, with Google reporting a fleet-wide average of 1.09 in 2024.
How much does data center downtime cost per minute?
Data center downtime can cost as much as $9,000 per minute, according to Uptime Institute research. Actual costs vary based on the business served, with financial services and e-commerce operations typically experiencing higher per-minute losses than other sectors due to direct revenue impact.
What are the power requirements for AI data center racks?
AI racks commonly require 100 to 240 kW per rack in 2025, with next-generation platforms projected to reach 250 to 900 kW. This is roughly 5 to 25 times the density of traditional 10 to 30 kW general-purpose server racks, requiring 100A+ PDUs and dedicated cooling infrastructure.
Why are data centers adopting 480V three-phase power?
Data centers adopt 480V three-phase distribution because it reduces resistive losses by roughly 81% compared to 208V single-phase, allows smaller conductor sizes, and supports higher-density racks with less copper. It is becoming the US standard for any new facility above a single-rack deployment.
What is a PDU in a data center and why is it important?
A PDU, or power distribution unit, is the final device in the data center power chain that distributes electricity from the UPS to individual servers and network equipment. PDUs provide branch-circuit protection, power monitoring, and often remote outlet switching, making them essential for capacity management and uptime.
What is the role of UPS systems in data centers?
UPS systems provide continuous, conditioned power during the seconds-long gap between a utility failure and generator start-up. Modern units also filter harmonics, regulate voltage, and support grid-interactive capabilities. Double-conversion UPS systems ensure zero transfer time, making them the standard for critical infrastructure.
How is edge computing impacting data center power demand?
Edge computing distributes power demand across many smaller sites instead of concentrating it in a few large facilities. Each edge site may consume only 50 to 500 kW, but these sites collectively add significant grid load. Edge power infrastructure must be self-contained, often unmanned, and designed for the same reliability as centralized facilities.