Data center infrastructure: the 7 layers that support your operation and who answers for each
A colocation contract hands you a single availability number: 99.982 %, 99.995 %, whatever it is. Behind it there are at least sixteen subsystems, each with its own standard, its own contractor and its own point of failure. And no document says which layer is responsible for which when something breaks at 3 AM.
This is the map that is missing: the seven layers that hold up your operation, what the standard says about each one, and where the contract breaks down. The difference between a data center that delivers its availability nines and one that only promises them is a document.
The 7 layers, one by one
ANSI/TIA-942 covers, in a single standard, architecture and topology, environmental design, power, cooling, telecommunications, redundancy, fire protection, security and monitoring, across four rating levels. It is astonishing how much a standard covers and how little it says about who is responsible for each thing.
The standard covers everything, the contract covers nothing
In the ANSI/TIA-942 standard, the Rated-3 level requires an active route and a backup route for power, cooling and telecommunications, and Rated-4 requires dual active routes. Compliance with a standard does not allocate responsibilities: it describes capabilities. That is why Uptime Institute split its certification into Topology and Operational Sustainability, and warns that long-term availability is not guaranteed by the Tier alone.
And the same source tightens the screw with data from its own incidents: the leading cause of reported outages is attributed to management failures, personnel errors and operating procedures. The most expensive layer of the contract is the one that gets audited the least and causes the most outages.
The strongest SLA is defined by the layer nobody supervises
Google SRE separates three things that colocation contracts mix into a single number: the service indicator, the service objective and the service agreement. Its warning is direct: 100 % availability is the wrong objective, because the chain of systems between you and your customer is long and any component can fail. Its table gives the size of the problem: at 99.999 % you tolerate 5.26 minutes a year, and a single event in an unmonitored layer consumes the whole budget.
From there comes the rule that serves you in the contract: if any component of the chain can fail, every layer needs its own service objective and its own owner. Applied to your room, you can ask in writing who handles 3 AM in the UPS, in cooling, in the network and in the control system.
The three mistakes that cost the most
Uptime Institute documents three topology failures that cost more than any brand difference. The first: a UPS configuration with common input and output switchgear is almost always unmanageable without taking the room down. The second: feeding the mechanical components in bypass, which forces shutting down the whole mechanical system for electrical work. The third: poor coordination of the chilled water isolation valves.
All three are design mistakes, not brand mistakes, and they are committed before anyone quotes a price. A provider who quotes you 99.995 % but cannot isolate a branch of chilled water without closing the whole circuit is selling you a number, not a service.
The cost of the layer sits in energy and maintenance
The 2025 report from the U.S. Department of Energy and Lawrence Berkeley National Laboratory calculates that data centers could consume 11.8 % of the country’s electricity by 2030, and breaks down the weight of each layer: infrastructure (cooling plus power distribution) accounts for 31 % of a data center’s energy in 2024, and the network 4.5 %. Average PUE fell from 1.55 in 2018 to 1.45 in 2024, and facilities with AI loads sit at 1.145. That 31 % is operating expense (OPEX): it is controlled with infrastructure management (facility management), not with the switchgear brand.
Your list for requesting the contract by layers
Before signing, ask for a seven-row matrix. On each layer: who owns the equipment, who operates it, who answers at 3 AM, what service objective applies to that component separately, what evidence there is of the commissioning tests, and what exclusion exists in writing.
And end with an uncomfortable question: if tomorrow your Tier IV degrades to Tier III because a branch failed, does the contract tell you who gets up at 4 AM? The standard does not say. The layer does not either. Only the document you demanded says it.
Sources
Want to master this?
Noxtel Academy →