Data center digital twin: what it is, and what it isn't

1. A definition that survives a real room

Vendors use “digital twin” for anything from a rendered floor plan to a physics simulation. A useful definition has three tests. A data center twin:

• Knows where everything is: every device in its rack and U, every rack in its row, every cooling unit and PDU in its place.
• Knows what depends on what: which UPS feeds which panel, breaker and PDU; which power supply plugs into which feed; which switch carries which server; which cooling unit serves which racks.
• Answers “what if” before you touch the room: fail a UPS, a feed, a switch or a cooling unit, and it tells you what goes dark, what runs on one feed, what overloads and how fast the room heats up.

A model that passes the first test is an inventory. One that passes the first two is good documentation. Only the third makes it a twin.

2. What it is not

	Answers	Needs	Typical user
3D visualisation	What does the room look like?	Positions	Sales, visitors
DCIM	What do we own, where, and who changes it?	Asset records, workflows	Facilities and IT asset teams
Monitoring	What is happening right now?	Sensors, agents, polling	Operations, on-call
Computational fluid dynamics	Exactly how does air move in this room?	Geometry, an engineer, hours per run	Mechanical engineers
Digital twin	What happens if this fails, or if we add that?	Positions, dependencies, ratings; readings optional	Operations, design, integrators

Three common confusions:

• A 3D view is not a twin. A beautiful room with no power chain cannot tell you what a UPS failure does.
• “Live” is a data source, not a definition. A twin can read measured values from monitoring, and it gets more accurate when it does, but a model built from a spreadsheet already answers most failure questions. Ask what the numbers are based on, not whether they move.
• A twin does not run the room. It should not switch equipment. It is where you try the failure first.

3. The layers that matter in a small or mid-size room

DC Sim models nine layers on one model: Floor (space, U, weight), Air (inlet temperatures, airflow), Water (liquid loops, manifolds, leaks), Fire (zones, detection, suppression), Power (the chain from utility to PDU, per-phase load), Cabling (every cord, traceable), Network (topology, utilisation), Device (health, parts, warranty) and Token (tokens per second and per kWh for AI rooms). Not every room needs all nine. Most rooms start with Floor and Power, because a power failure is the question people ask first.

4. Where the data comes from

• A spreadsheet first. Rack, U and model per device is enough to draw the room; power cords and a Power sheet add the failure answers. The Excel template has one sheet per layer, and only Devices is required.
• The tools you already run. Zabbix and NetBox connect directly with a read-only token; any other tool’s export goes into the Readings sheet.
• Defaults, marked as such. What the file does not say (aisle direction, a cooling zone, typical device power) is filled with defaults labelled assumed, so you know what to replace.

Every projected number (temperatures, runtimes, minutes to a threshold) is labelled estimated. The air model is zonal, not CFD.

5. Design, run, fail it first

The same model works before and after the room exists. In Design mode you place racks and devices, with U, weight, breaker and cooling checks as you go. When the room is built, you import the as-built and keep the same model as the running twin. In Simulate you fail anything: a feed at peak, a UPS, a cooling unit, a manifold leak, more load on a row. Each run sits next to an untouched baseline.

6. How to judge one in 30 minutes

1. Import your own file, not the vendor’s demo. How long until you see your room?
2. Fail a UPS at peak. Does it list dark devices, devices on one feed, stopped services and overloads? Or does it just turn things red?
3. Ask where each number comes from. Measured, from a sheet, from a vendor table, or estimated? It should say.
4. Re-upload a changed file. Do you see exactly what changed before it applies?
5. Check the thermal claim. A result in seconds is not CFD and should not be sold as one; an honest, labelled zonal estimate is fine for operational questions.
6. Look at the price unit. Per asset, per rack or per project? Compare it with what DCIM costs for your room size.

7. Build a first twin in an afternoon

The four steps below are the whole method: Devices, the power chain, import and read the checks, fail the first node. The clip shows the same steps on the 13-rack demo hall.