Power — Trivia & Interesting Facts¶
Surprising, historical, and little-known facts about power systems in datacenter and infrastructure contexts.
A hyperscale datacenter campus can draw more power than a small city¶
Microsoft's Quincy, Washington campus draws over 200 MW. Meta's Altoona, Iowa facility exceeds 200 MW. For context, 200 MW is enough to power approximately 160,000 homes. The single largest constraint on new datacenter construction is not land or cooling — it's getting enough utility power to the site. Some datacenter developers are now building their own power substations.
UPS batteries are the most common single point of failure in datacenter power¶
Despite being designed to prevent power failures, UPS (Uninterruptible Power Supply) systems are themselves a leading cause of datacenter outages. The Uptime Institute found that UPS failures caused 33% of all reported datacenter power outages. Battery degradation, capacitor failures, and transfer switch malfunctions are the most common failure modes. This is why Tier IV datacenters require 2N UPS configurations.
Power Usage Effectiveness (PUE) was created by The Green Grid in 2007¶
PUE, defined as total facility power divided by IT equipment power, was introduced by The Green Grid consortium in 2007. A PUE of 2.0 means the facility uses as much power for cooling and overhead as it does for computing. Google achieved a fleet-wide PUE of 1.10 in 2022, meaning only 10% of their power goes to non-compute overhead. The industry average in 2023 was approximately 1.58.
Diesel generators in datacenters are the same engines used in railroad locomotives¶
Large datacenter backup generators typically use engines rated at 2-3 MW each — the same class of diesel engine found in EMD and GE railroad locomotives. A hyperscale datacenter might have 50-100 of these generators, collectively capable of producing 100-200 MW. The fuel tanks store enough diesel to run the facility for 24-72 hours. Testing these generators is so loud it can be heard miles away.
Power distribution units (PDUs) in modern racks can measure per-outlet power draw¶
Modern intelligent PDUs (from companies like Raritan, Server Technology, and APC) can measure voltage, current, and power factor per outlet. This granularity allows operators to calculate exact per-server power costs, identify underutilized servers, and detect hardware problems (a server drawing 20% less than its peers likely has a failed power supply or disabled components).
The shift from AC to DC power distribution can improve efficiency by 10-15%¶
Traditional datacenter power goes through multiple AC-DC-AC conversions (utility AC to UPS DC to UPS AC to server PSU DC). Each conversion loses 3-5% efficiency. Facebook and Google experimented with distributing 48V DC directly to servers, eliminating two conversion stages. This approach, adopted in the Open Compute Project, can improve total power efficiency by 10-15%.
A single rack in a modern datacenter can draw 30-50 kW — up from 5 kW in 2005¶
Server power density has increased dramatically. A rack of GPU servers for AI training can draw 40-50 kW or more — equivalent to 30-40 average American homes. This density increase has made cooling the primary engineering challenge in modern datacenters, driving adoption of liquid cooling and rear-door heat exchangers. Some AI-focused racks are projected to reach 100+ kW by 2026.
Power cord locking mechanisms exist because servers used to vibrate their own cords loose¶
Server fans and disk drives create vibration that, over months, can gradually walk power cords out of their sockets. C13/C14 locking power cords (with a retaining clip) were developed after several notable outages caused by servers losing power due to vibration-loosened connections. The fix costs less than $2 per cord but prevents potentially catastrophic single-server failures in otherwise redundant configurations.
Japan's 100V power standard forces different server PSU configurations¶
Most of the world uses 200-240V power distribution in datacenters, but Japan's standard voltage is 100V. This requires different power supply configurations and results in higher current draw for the same wattage (P=VI), which means thicker cables and different circuit breaker ratings. Server manufacturers typically offer Japan-specific PSU variants, and deploying to Japanese datacenters requires different power planning.
A single lightning strike can carry 300 million volts and 30,000 amps¶
Datacenters in lightning-prone areas (like Northern Virginia and central Florida — both major datacenter markets) must have robust lightning protection. A direct strike to a building can induce voltage surges in every conductor in the facility. Modern datacenter lightning protection includes building-level surge suppressors, isolated grounding systems, and shielded power distribution. The Empire State Building is struck about 25 times per year.
Power factor correction saves datacenters millions of dollars per year¶
Utility companies charge commercial customers for both real power (watts) and reactive power (volt-amperes reactive). A poor power factor (below 0.9) means the facility draws more current than necessary for the actual work done, and utilities impose financial penalties. Modern server power supplies achieve power factor correction (PFC) above 0.99, but older equipment can drag facility-wide power factor below penalty thresholds.