It deliberately does not re-explain what each component is — the beginner guide does that — or the fundamentals of data-centre tiers and redundancy notation, which the pillar covers. It starts where those leave off: turning a parts list into an availability guarantee.
The power chain
Power enters the site from the utility at medium voltage, passes through transformers and switchgear, is conditioned by UPS systems, distributed through PDUs and busways, and finally reaches rack PDUs and server power supplies. Backup generators and fuel systems support extended utility outages.
In resilient facilities, this chain is duplicated or made maintainable so equipment can be serviced without shutting down IT load.
What to check at each stage
For what each of these components does in plain terms, see the power systems guide. The design question is different at every stage: what single failure could take the load down, and can that stage be maintained without a shutdown?
| Component | Role | Risk to check |
|---|---|---|
| Utility feeds | Primary grid supply into the site. | Single substation or truly diverse feeds? |
| Switchgear | Controls and protects electrical distribution. | Can breakers be maintained without load impact? |
| UPS | Bridges outages and conditions power. | Battery autonomy, bypass design and maintenance process. |
| Batteries | Short-duration backup until generators start. | Testing, ageing, thermal management and fire safety. |
| Generators | Long-duration backup during utility loss. | Fuel contracts, start testing, emissions rules and runtime. |
| Rack PDUs | Final distribution to IT equipment. | Per-outlet monitoring, phase balance and breaker capacity. |
Redundancy language
N means the exact capacity needed to support the load. N+1 adds one spare component. 2N provides a fully independent second path. Concurrently maintainable means planned maintenance can occur without impacting the IT load. Fault tolerant means the system can survive a failure as well as maintenance.
These terms only matter when applied to a defined scope. A facility may have 2N UPS but only N cooling, or diverse utility feeds but a shared downstream switchboard.
Rack density and AI load
Traditional enterprise racks often sat between 3 kW and 10 kW. GPU and AI racks can push far higher densities, which changes cabling, busway, breaker, UPS, cooling and floor-loading assumptions. A hall designed for 8 kW racks cannot simply accept 80 kW racks because the contract says there is spare megawatt capacity.
Always distinguish between site capacity, committed IT capacity, rack density and available power path capacity. They are related, but not the same.
Operational controls
Many power incidents are procedural, not purely electrical. Good facilities maintain switching procedures, method-of-procedure approvals, infrared scanning, breaker testing, UPS battery tests, generator load-bank tests, fuel quality checks, change freezes and incident drills.
For buyers, the useful question is not only "what is the design?" but "how do you operate, test and evidence the design?"
Sources
- Uptime Institute Tier Certification overview
- Uptime Institute Tier Classification System
- ASHRAE TC 9.9 power trends white paper
- IMDA Green Data Centre Roadmap
Related resources
Go deeper on this topic
Knowledge base
Vendor directories
Research cluster
Related analysis
Recent TechDirectory Insights coverage from the same research cluster.