Cooling & heat removal

Cooling decisions at a glance

Cooling depends on the capacity envelope, the resilience target and the power design, and it fixes the density ceiling, the rack layout and how ready the hall is for liquid. Onsite you design and run the plant; in colo you work within the provider’s plant and rules. Site model and ratings are on Facility: Onsite vs colo and Facility: Resilience; power density on Power.

Containment and airflow management

Containment separates the cold supply air from the hot exhaust so that equipment inlets see the supply temperature. It is the main tool for raising density on air cooling. Cabinet parts (blanking panels, brush strips, seals) are on Racks: Airflow parts; row and aisle layout on Racks: Aisles. Sources checked 5 Oct 2026.

What containment fixes

Two airflow faults waste cooling. Bypass is supply air that returns to the cooling units without passing through IT equipment. Recirculation is hot exhaust that finds its way back into equipment inlets. Containment and blanking reduce both, so supply setpoints can rise without hot spots. Measure against the equipment inlet, not the room return: ASHRAE’s thermal envelopes are defined at the inlet, and ASHRAE advises designing and operating to the recommended range (18–27 °C for classes A1–A4); the wider allowable range describes what equipment is tested to function in, not reliable operation (ASHRAE TC 9.9 reference card, 5th ed.).

Hot-aisle vs cold-aisle containment

Containment roofs and doors interact with fire detection and suppression, so the allowed method depends on local building and fire codes. One large colo provider states that its containment method varies by site for this reason (Equinix: Colocation cooling). Fire systems are on Ops: Fire & life safety.

Equipment airflow direction

Containment only works if every device takes air in on the cold side and exhausts to the hot side. Network switches mounted port-side to the hot aisle often need a reversed fan and PSU option, so order the airflow direction that matches the row layout.

Rear-door and liquid cooling

Liquid takes over where air runs out of headroom; rack-scale AI systems are already liquid-cooled. Options range from liquid-to-air at the rack, where IT stays air-cooled, to liquid at the chip or full immersion. GPU-specific requirements (coolant, inlet temperatures, flow per module) are on AI Factory: Liquid cooling. Fast-changing: recheck this section at least every six months. Sources checked 5 Oct 2026.

Terms

FWS (facility water system) is the building’s chilled- or warm-water loop. TCS (technology cooling system) is the secondary loop that serves IT equipment. A CDU (coolant distribution unit) separates the two with a heat exchanger and pumps (ASHRAE TC 9.9 liquid cooling white paper, 2021). ASHRAE classes liquid-cooled equipment by facility water supply temperature, from W17 up to W45 and W+ (above 45 °C) (ASHRAE reference card).

Options

DMTF’s Redfish 2023.1 models rear-door and in-rack heat exchangers alongside CDUs and immersion units as liquid-cooling equipment (DMTF Redfish 2023.1 overview). The Open Compute Project’s Cooling Environments project publishes open specifications for cold plates, CDUs, door heat exchangers, immersion, coolants, quick connectors and leak sensors, useful as RFP references (OCP Cooling Environments).

Liquid-ready design

ASHRAE TC 9.9 recommends that any new data center include the capability to add liquid cooling later. It also expects the facility water temperature that liquid-cooled equipment can accept to fall as chip power rises, so a liquid-ready loop should not be designed only around high-temperature water (ASHRAE TC 9.9 white paper, 2021). Day-1 provisions for a hall built on air:

• FWS tap-offs or capped headers sized for a future liquid zone
• Floor space reserved for in-row CDUs
• Overhead pipe routes above aisles, with drip pans, leak detection and drains under joints over critical equipment, coordinated with Racks: Pathways (ASHRAE white paper)
• A wetted-materials list and a clean-and-flush plan in the design (ASHRAE white paper)
• Floor loading for filled racks and CDUs, on Facility: Floor loading
• CDU pump power on a protected supply, on Power
• Leak-detection points in the BMS points list (Cooling controls)
• Air-plant headroom for the residual air load of liquid-cooled racks

What a cooling topology means for a resilience rating (concurrent maintainability, continuous cooling, CDU and pump failure) is on Facility: Resilience.

Colocation cooling limits

In colo the provider designs and runs the plant. You work inside its density ceiling, containment rules and liquid offering. Contracted kW per cabinet is on Power; cage fit-out and provider installation rules on Racks: Colo fit-out; provider evaluation and SLAs on Facility: Colo SLA. Sources checked 5 Oct 2026.

What providers typically impose

One large provider’s published documentation shows the kinds of rules to expect (Equinix: Colocation cooling):

• Each cabinet has a defined cold and hot side; equipment installed backwards breaks the provider’s service guarantees
• Containment is audited, and some containment parts are chargeable
• The containment method varies by site with local building and fire codes, so a cage design may not carry over between sites
• Supplemental cooling (rear doors, in-row coolers) is offered only at some sites or in cages
• Liquid is delivered to the cage, with the demarcation at the tap-off to the CDU; you bring or buy the CDU

ASHRAE noted in 2021 that colo SLAs may exclude water near the racks, which limits tenants to room air cooling unless liquid-to-air units are used (ASHRAE TC 9.9 white paper, 2021). Provider liquid offerings have changed since; ask each provider.

Questions for the provider

• Cooling ceiling per cabinet and per cage, and how it is enforced
• Inlet temperature SLA: range, where it is measured (cabinet inlet or room), and which ASHRAE envelope
• Containment and blanking rules, and who supplies and pays for parts
• Whether rear doors or in-row coolers are allowed at this site
• Liquid: demarcation point, TCS ownership, water-quality spec, leak response and damage liability
• Remote-hands scope for cooling incidents, and what cooling telemetry you get (portal or API)

Cooling controls and connectivity

CRAHs, chillers, CDUs and leak-detection panels are networked controllers. The BMS as a platform is on Ops: BMS. Network placement and isolation are on Security: Zones and conduits and Security: OT; telemetry collection is on Ops: Monitoring. Sources checked 5 Oct 2026.

• List the protocol each unit speaks in the RFP: BACnet/IP, BACnet/SC, Modbus RTU/TCP (common on CRAHs and chillers), SNMP or Redfish. Read-only collection rules, including blocking Modbus write function codes, are on Security: OT placement
• BACnet/IP (UDP 47808) broadcasts do not cross IP routers, so a BMS spanning subnets needs BBMDs or foreign-device registration (BACnet International). BBMD and foreign-device traffic that crosses zone firewalls must appear in the conduit rules. A collector in the monitoring DMZ should register as a foreign device, or use BACnet/SC, so broadcasts are not extended into the DMZ. BACnet/IP writes are unauthenticated, so a read-only collector needs a protocol-aware firewall or BACnet/SC (Security: OT). BACnet/SC adds TLS-secured, certificate-based connections (BACnet International: BACnet/SC), which brings certificate lifecycle into scope
• Placement is set on the security page (Security: Cooling placement): facility CDUs, CRAHs, chillers and leak-detection panels sit in the OT mechanical zone, and only rack-integrated cooling managed through the rack’s own management controller (such as in-rack CDUs in rack-scale AI systems) goes on Network: OOB
• CDUs may be managed like IT equipment: Redfish defines schemas for cooling units, loops, pumps and leak detectors (DMTF Redfish 2023.1). Vendor Redfish support varies; ask for it. Collection should use a read-only device account role, because a firewall cannot tell a Redfish read from a write inside TLS (Security: OT placement)
• Plant controllers should keep running on local control if the BMS or network fails
• In colo the provider’s BMS is not yours: your own CDUs and sensors need their own connectivity in the cage