Key takeaways

  • Choose the facility on power and cooling capacity per rack, network carriers on site, access and remote-hands terms, and the facility's design standard, not on price per rack alone.
  • Plan power in kilowatts per rack from the real load of the servers, and feed every dual-power device from independent A and B paths. Size each path so that either one can carry the whole rack alone.
  • Keep hot and cold air from mixing. Containment, blanking panels and brush strips cost little compared with the cooling they save. ASHRAE Technical Committee 9.9 publishes the thermal guidelines that equipment makers and facilities use.
  • Structured cabling follows standards: ANSI/TIA-568 for generic cabling, TIA-606 for administration and labeling, and ANSI/TIA-942 for data center infrastructure. Label both ends of every cable before it is pulled.
  • Burn-in finds early failures before production does. Build from a written baseline for firmware and settings, run memory, storage, network and thermal stress, and accept hardware against a checklist.
  • NIST SP 800-88 Rev. 2, published in September 2025, defines clear, purge and destroy methods for media sanitization and calls for records of each sanitized item. Plan for it on day one, not when the racks are being removed.

What an empty cage is

A colocation cage is a fenced area on a data center floor that you rent for your own equipment. The operator supplies the building, electrical power, cooling, physical security and often network carriers. You supply everything inside: racks or cabinets, power distribution units, switches, servers, storage, cabling and the people who understand them. Colocation suits companies that need low latency to a market or carrier, a second site for recovery, or control of hardware that a public cloud does not offer, and that do not want to build a facility.

Starting from an empty cage is a rare chance to do things in order. Most footprints grow by accident: a rack here, a switch there, a cable run that nobody labeled. This article is written for the people who will plan, build and operate a new deployment, and for the executives who will approve the budget. Our team has built such footprints for financial and technology clients and the order below is the one we follow.

Choosing the facility and the contract

Price per rack is the least informative number in a colocation quote. Start with capacity. How many kilowatts can the facility deliver to one rack, and to your whole cage, now and in two years? What cooling method does it use, and does the contract allow your planned density? A cage that cannot be cooled at the density you buy servers for will be left half empty. Ask how power is billed: by committed capacity, by metered use, or both, and what happens when you exceed the commitment.

Then look at connectivity and resilience. Which carriers and internet exchanges are in the building, and can you cross-connect to them directly? How are the cross-connect fees structured? Are there two independent fiber entrances? Does the facility describe its design against a published standard? ANSI/TIA-942 is the data center infrastructure standard from the Telecommunications Industry Association, and the association runs a certification program that independent auditors assess. A rating or certificate is a claim to be verified, not proof. Read the actual report and ask what maintenance can be done without taking systems down.

If the cage will talk to a public cloud, ask which cloud on-ramps the facility offers. The providers sell private links: AWS Direct Connect, Google Cloud Interconnect, Microsoft Azure ExpressRoute and Oracle Cloud Infrastructure (OCI) FastConnect. Each works through a connectivity provider or a nearby facility, so the answer depends on the building, not only on your carrier.

Finally, read the human terms. How do you get in, and how quickly? Can your staff bring tools, packaging and spares? What does the remote-hands service cost and what is its response time? What are the rules on shipping, storage, loading docks and hazardous materials? Who is liable for what if there is a failure? Ask for a site visit and walk the path your equipment will take from the dock to the cage. Also ask about the location's weather and flood risk, as we cover in our Florida cloud article.

Cage layout

Design the cage on paper before ordering. Decide the number of racks, the row direction and where the aisles run. Racks should face the same way in rows, so that cold air enters at the front of every server and hot air leaves at the rear into a shared aisle. Leave clear space in front of and behind each rack for service. Plan where network cross-connects will land, usually a meet-me area or the cage patch panel, and place the core network racks near it to keep cable runs short.

Power: kilowatts per rack, A and B feeds

Power is the constraint that decides everything else. Start from the equipment, not from the rack. Each server model has a measured power draw under typical and peak load, which the vendor's power calculator and your own tests will give you. Add switches, storage and power distribution losses. The sum, divided by the number of racks, is the planned kilowatts per rack. Be honest about density: a rack of general-purpose servers draws far less than a rack of GPU servers, which can require specialized cooling that many facilities do not offer.

Servers with two power supplies should connect one to an A feed and one to a B feed. The two feeds should come from independent upstream paths, so that the loss of one does not take the other. Each feed must be able to carry the full rack alone, because when one fails the other takes over everything. That means that in normal operation each feed should run at no more than about half of its limit, and that the limit itself includes a safety margin set by electrical rules for continuous loads. Your colocation provider and a licensed electrician confirm the numbers for your circuits and breakers.

Use metered, switched or monitored power distribution units so that you can see the load on each outlet and each phase, set alarms and cycle a stuck device remotely. Balance the phases on three-phase feeds. Record which device is plugged into which outlet and which feed. When a feed goes down for maintenance, the record tells you in advance which servers will run on one supply. Test that claim during acceptance by pulling each feed in turn.

Cooling and containment

Servers pull cool air in at the front and push hot air out at the rear. Efficient cooling depends on keeping the two streams apart. When hot exhaust recirculates into the intakes, the cooling system works harder, hot spots appear and fans speed up. The standard tools are simple. Fill every empty rack position with a blanking panel. Seal gaps around and between racks, and use brush strips where cables pass through. Use the facility's hot-aisle or cold-aisle containment if it offers it, and place racks in the pattern the facility's design expects.

ASHRAE Technical Committee 9.9, which covers mission critical facilities, technology spaces and electronic equipment, publishes the thermal and environmental guidelines for data processing environments, now collected in its online Datacom Encyclopedia. Equipment makers state the inlet temperatures their hardware supports, and facilities state the conditions they deliver. Compare the two before you buy. We do not repeat the temperature and humidity ranges here, because they are revised and depend on equipment class, so use the current guideline and your vendors' specifications. Monitor inlet temperature at the front of several racks, not just the room's return air.

Structured cabling and labeling

Structured cabling means a planned, standards-based system of cables, panels and pathways, rather than point-to-point runs added over time. ANSI/TIA-568 is the family of standards for generic telecommunications cabling, covering copper and optical fiber components and performance. TIA-606 covers administration: how to identify and record cables, panels, pathways and spaces. ANSI/TIA-942 covers data center infrastructure, including spaces, cabling, pathways and ratings. These are paid standards, and your cabling contractor should work to the current editions.

For high-speed links inside the cage, many designs use multi-fiber push-on connectors, known as MPO or MTP, on trunk cables that carry twelve or more fibers in one jacket, with cassettes or breakouts at each end. This gives a neat way to run 40, 100 and 400 gigabit links, but polarity and connector gender must be chosen carefully, because a wrong polarity method means links that do not light. Our article on optics from 10 to 400 gigabits explains the transceivers. Test every fiber with a light source and power meter, and record the results.

Label every cable at both ends, every port on every panel, and every device. Use a scheme that includes rack, panel and port or device and port, and is the same in the drawing, the label and the inventory. Print labels before the cable goes in. Cables tied in rows with hook-and-loop straps, never tight plastic ties on fiber, can be changed without taking down neighbors. Keep an as-built record. When someone in a remote location is asked to move one cable, a clear label is the difference between a 5‑minute job and an outage.

Building the servers

Write a build baseline before the first server arrives. It lists the exact model and configuration, the firmware versions for the BIOS or UEFI, the baseboard management controller, network adapters, storage controllers and drives, and the BIOS settings you want. Each firmware version should be one you have tested together. Applying a baseline to every machine removes a class of intermittent faults that only appear on one machine.

Memory deserves care. Modern servers use several memory channels per processor, and performance depends on populating channels evenly. DDR5 platforms differ in how many channels each processor has and in how speed changes when you add more than one module per channel. Follow the server vendor's memory population guide for your exact model, install matched modules, and check in the firmware that every channel and module is recognized at the expected speed.

Management should be automated from the start. The Distributed Management Task Force's Redfish standard is an open interface for managing servers and other infrastructure out of band, over the network, through the management controller. A script that talks to the controller can set boot order, apply firmware, read sensors and power-cycle a machine, without a person standing at the rack. Put the management network on its own isolated segment with strong authentication, change every default credential, and keep management interfaces off the internet. Our microsegmentation article covers the network design.

Burn-in and acceptance

Hardware fails most often early in its life, and burn-in is the practice of running it hard before production does. Do this in the staging area or in the cage before go-live. Typical steps are a long memory test, a storage test that reads and writes every drive and reviews the drive's health data, a processor stress test that pushes thermal limits, and a network test that fills links in both directions and checks error counters. Watch temperature, fan speed and power throughout, and read the controller's hardware event log afterward.

Define pass and fail in advance. A server passes if it completes the planned duration with no errors in the event log, no corrected-error trend and temperatures inside the limits. Anything else goes back to the vendor under warranty before the machine enters production. Keep the results per serial number, because they become the baseline when the same machine misbehaves in year three.

Acceptance covers the whole cage, not just servers. Verify that each rack's power matches the plan and that feed failover works. Verify every cross-connect and uplink, with light levels recorded. Confirm that monitoring, alarms and remote access work. Test the out-of-band path as well, such as a cellular or separate carrier link, since it is the way in when the main network fails.

Remote hands and written procedures

Once the cage is running, most work happens at a distance. When something requires a hand, such as replacing a drive, moving a cable or reading a light, you will ask the facility's technicians or your own travelling staff. The way to get reliable results is to write a method of procedure, usually shortened to MOP, for each recurring task. A good MOP names the goal, the affected systems, the rack and device identifiers, the steps in order, the checks after each step, the point at which to stop and the way to back out.

Decommissioning and media sanitization

Equipment leaves a cage for many reasons: end of life, a lease return, a refresh, a closure. Whatever the reason, storage media may hold customer data, credentials and keys. NIST Special Publication 800-88 Revision 2, Guidelines for Media Sanitization, was published in September 2025 and supersedes Revision 1 of 2014. It defines media sanitization as a process that renders access to target data on the media infeasible for a given level of effort, and it helps organizations build a sanitization program with techniques and controls based on the sensitivity of the information.

The publication describes three methods. Clear applies logical techniques to all user-addressable storage locations, protecting against simple, non-invasive recovery. Purge applies physical or logical techniques that make recovery infeasible using state-of-the-art laboratory techniques while preserving the media in a potentially reusable state, and the guidance says to use purge instead of clear when possible. Destroy renders data recovery infeasible and leaves the media unusable. It also covers cryptographic erase, which sanitizes by destroying the keys of data that was encrypted, with a precondition: sensitive data must not previously have been stored in plaintext on that media. For some kinds of solid state drives, overwriting alone gives little protection, and the guidance recommends purge or destroy when more assurance is needed.

The program side matters as much as the technique. Sanitization should be verified, and a certificate of sanitization completed for each item, recording at least details such as manufacturer, model, serial number, the method and technique used, and who performed it. The publication notes that tracking should start when media is introduced to the environment, so you can show that everything that arrived also left clean. Decide early whether you will encrypt every drive, so that cryptographic erase is available later, and whether drives that fail under warranty will be returned or destroyed. Returning a failed drive to a vendor is a decision about media control that should be written into the plan.

Keep a parts kit on site and agree in the contract what remote hands may and may not do. For anything that could cause an outage, schedule it, announce it and watch from the other end. Update the inventory after each job.

Clayton Cybernetics plans and builds colocation footprints. A first call usually starts with load numbers and a floor plan. Get those on paper, and the rest of the project follows in order.

Sources

This article is general information, not an electrical, structural or fire-safety design and not legal advice. Power, cooling and cabling work must follow the local electrical code, the facility's rules and the current editions of the standards, which are paid documents that we cite by name and describe in general terms. On 2026-11-08 we read NIST SP 800-88 Rev. 2 in full, the TIA-942 program page, the ASHRAE TC 9.9 page through the Internet Archive and the DMTF Redfish Developer Hub; the ANSI/TIA-568 and TIA-606 texts and the ASHRAE thermal guideline tables were not available to us, so we give no distances, temperatures or humidity values. Server memory population and firmware baselines differ by vendor and model; follow your vendor's documentation. Sanitization requirements for your data depend on your regulator and contracts. Last reviewed .

Keep reading

Need help with this?

Cages, racks, power, cabling and servers, built by the people who will run them.

Datacenter & Colocation Builds