What does "cross connect" means in Datacenters?

A cross-connect is any connection between facilities provided as separate units by the datacenter.

In other words, if you rent a cage, you can run cables betweeen your various racks and they are not really considered "cross connects".

But typically a user facility is a rack. Consider the uplink to a network provider. The line that runs from you to the telco's rack is a cross-connect. Any other between-rack cable run falls in this category. Even if the datacenter is bundling network access, it's still a cross-connect. (It's also typical to have a meet-me room that is the center of all the cross-connects, it's a patch-panel room or cage.)

This matters a lot to the datacenter because the crossconnect uses their overhead cable trays and typically comes with both a setup and a monthly charge.

Update: If itemized, the charge can be $100 - $300, and can depend on various things like the type and speed of the line.

I should also add that sometimes people will unfortunately use "cross-connect" in its literal english sense when referring to any kind of a network connection, even if it's in a datacenter context. I've seen fiber metro links between different buildings and uplink bandwidth contracts quoted as "cross-connects.

Cross Connects are simply a link from a carrier or ISP to customers suite or rack to provide services. Most DC's have more information on them on the web page. Cross connects had mainly been on copper (Cat5e/6 etc.) but recently a large amount are on OM3 50/125 fibre cables to carry the bandwidth.

Many of these answers contain some excellent information, but several of them contain what seems to me like an element of confusion.

Cross-connnects are the physical or virtual connections from one facility to another.

Note that facility, here, does not mean a building or partition, and does not refer to a collocation enclosure (rack or cage)... it refers, instead, to a layer 0 entity, such as a single run of fiber, coax, or copper, that spans the desired endpoints of one leg of the complete circuit (physical cross-connect)... or to another low layer entity, such as an Ethernet port or VLAN on a port, a SONET or DSX timeslot, an ATM PVC, etc. (virtual cross-connect).

In a collocation environment, the collo vendor typically has home-run facilities connecting its distribution frame to all the tenant enclosures, and when a cross-connect is ordered, this involves the vendor installing a patch on their frame to connect two existing facilities -- it does not involve running an actual new line all the way from customer A to customer B. If no spare facilities are in place to support the desired circuit, those would typically need to be ordered separately. Circuits on facilities that run directly, physically, from tenant to tenant and are physically installed and reconfigured would be a problematic disarray of problems waiting to happen. Facilites terminating to a common frame under the collo vendor's control allow proper recordkeeping, security, and test access for troubleshooting.

If tenants A and B both have copper or fiber Ethernet facilities from patch panels in their enclosures to the data center operator's frame, and they want the operator to connect them together, only the data center operator's patch cable is considered the cross-connect. The wire or fiber (facilities) used by those circuits between the tenant patch panels and the vendor patch panels are part of the circuit that is completed by the cross-connect, but they are not properly considered part of the cross-connect. Those individual customers will also need to install patch cables within their enclosures if their equipment is not already pre-wired, but while these might also be referred to as cross-connects, they are not cross-connects in the sense that is relevant, here.

Similarly, if both companies A and B have existing Ethernet connections to ports A and B on the vendor's Ethernet switch, and they request that the vendor enable VLAN 694 on both port A to port B, then that is also a cross-connect, but a virtual one.

Typically, when two different companies own the facilities involved in the proposed cross-connect, one company provides a letter of authorizarion (LOA) authorizing the other company to order the cross connect to a specific connecting facility assignment (CFA, which is the patch panel connection, port, channel, VLAN, etc., where the other side of the proposed circuit terminates). The vendor will not create a cross-connect that is not authorized by the owner of record of the individual facilities involved (the copper, fiber, or coax of lower layer interconnections... or the ports/channels/timeslots/VLANs, etc. of higher-layer interconnections).

If you want some background reading, there's a lot about this from p98 of the book Tubes: A Journey to the Center of the Internet by Andrew Blum.

Essentially it's called cross-connect because it's a dedicated cable from one rack to another - the idea is it doesn't go through any other network switches or anything that could introduce extra latency, a single point of failure or risk impacting the service to other users. Cross-connects are used for Peering.

Besides being able to share bandwidth / potentially save money, another advantage of cross connects is reducing the number of hops between carriers.

The book describes how a technician will have to climb a ladder and carefully run a fibre ethernet cable (it doesn't have to be fibre, but nowadays..) through the special cable trays around the top of the racks.