Is a Tier III Data Center Enough for Liquid-Cooled AI?

A Tier III data center can support liquid-cooled AI racks, but the facility must be able to maintain cooling during a power outage or equipment failure. Schneider Electric's Reference Design 99 targets Tier III availability for AI racks requiring up to 73 kW each. The design includes 5 minutes of thermal storage to maintain cooling during a power outage, enough to cover the 3-minute chiller restart time.

Whether Tier III is sufficient also depends on the AI workload. Training jobs can recover from node failures using checkpoints, while production services without failover may require Tier IV or additional cooling safeguards.

What Each Data Center Tier Guarantees

The Uptime Institute tier standard defines four levels of data center infrastructure reliability. Each tier incorporates the requirements of the tiers below it.

1. Tier I: Basic infrastructure capacity. The facility must shut down for preventive maintenance and repairs.

1. Tier II: Redundant power and cooling capacity components. An unexpected shutdown can still interrupt IT operations.

1. Tier III: Concurrently maintainable infrastructure, with redundant components and distribution paths. Equipment can be maintained or replaced without shutting down IT operations.

1. Tier IV: Fault-tolerant infrastructure, with independent and physically isolated systems. A single equipment failure or distribution path interruption does not stop IT operations. Tier IV also requires continuous cooling.

The main distinction for liquid-cooled AI deployments is that Tier III does not require continuous cooling.

A Tier III facility may therefore experience an interruption in coolant flow during a power transfer unless the cooling system includes additional safeguards.

Tier III Data Center Requirements for Liquid-Cooled AI

In a liquid-cooled AI data center, concurrent maintainability must extend to the coolant distribution units (CDUs), pumps and piping. These systems need to support equipment maintenance without interrupting coolant flow to the GPU racks.

Schneider Electric's Reference Design 99 provides an example of how this can be achieved.

The design includes:

• N+1 CDU redundancy: Liquid-to-liquid CDUs share a common cooling loop, allowing one unit to be unavailable while the remaining units provide cooling.

• UPS-backed cooling equipment: CDUs and chilled-water pumps receive power from an uninterruptible power supply (UPS).

• Thermal storage: The cooling system provides 5 minutes of continuous cooling following a power outage or chiller restart, compared with a 3-minute chiller restart time.

UPS-backed pumps and thermal storage go beyond the minimum requirements of Tier III certification.

When evaluating a Tier III facility for liquid-cooled AI, the design documentation should confirm whether these additional cooling safeguards are installed.

Tier III vs. Tier IV: How AI Workloads Affect the Decision

The appropriate data center tier also depends on how the AI workload responds to equipment failures.

According to an Uptime Institute analysis of AI liquid cooling, redundant cooling equipment is often uneconomical in current liquid-cooling designs. AI training workloads can tolerate some node failures by recovering from checkpoints.

However, production AI services without software failover may not have the same recovery capability.

For these workloads, Tier IV infrastructure or a Tier III facility with additional safeguards, such as UPS-backed cooling and separate CDU loops, may be appropriate.

The decision therefore depends on both the facility's cooling infrastructure and the workload's ability to recover from interruptions.

Tier III Certification: TCDD vs. TCCF

A Tier III rating can refer to either an approved facility design or a completed, tested facility. Uptime Institute provides two separate certifications:

1. Tier Certification of Design Documents (TCDD): Confirms that the facility's design documentation meets the requirements of the claimed tier. This certification is issued before construction.

1. Tier Certification of Constructed Facility (TCCF): Confirms through on-site inspection and testing by Uptime Institute engineers that the completed facility meets the requirements of its certified design.

For example, in February 2026, Uptime Institute awarded Tier III TCDD certification to Khazna's 100 MW QAJ01 AI data center in Ajman, UAE. The facility was designed to support liquid cooling.

This certification confirms that the facility's design meets Tier III requirements. A separate TCCF certification would be needed to confirm that the constructed facility meets those requirements.

What to Check Before Deploying AI Racks in a Tier III Data Center

A Tier III rating confirms specific infrastructure reliability requirements, but it does not establish whether a facility is ready for a high-density, liquid-cooled GPU cluster.

Before deploying AI racks, the following six areas should be verified:

1. Constructed facility certification: Confirm that the facility has Tier III TCCF certification and that the certificate applies to the specific building housing the GPU racks. Certification of another building does not cover the proposed deployment.

1. Cooling ride-through: Verify how many minutes of cooling the thermal storage provides and compare this with the chiller restart and generator transfer times. Insufficient ride-through can interrupt cooling to the GPU racks.

1. Cooling failure domains: Determine how many GPU racks share each CDU loop. A failure affecting one loop can take all connected nodes offline. N+1 CDUs provide redundancy against a single CDU failure, but the piping also needs a redundant path.

1. AI workload recovery: Establish how frequently training workloads save checkpoints. Checkpoint frequency determines how much training progress may be lost following a cooling failure.

1. GPU infrastructure readiness: Conduct a GPU readiness audit to verify that the facility has sufficient power, cooling capacity, network infrastructure and floor load capacity for GPU racks. These requirements are not covered by the tier rating.

1. GPU cluster commissioning: Complete GPU cluster commissioning, including DCGM diagnostics, GPU burn-in and NCCL network fabric validation, before accepting the cluster into production. These tests are required regardless of the facility's tier.

A Tier III data center can therefore be suitable for liquid-cooled AI, provided its cooling infrastructure meets the requirements of the deployed GPU racks and workloads.

Zettabyte provides AI data center assessment, design and build services through its TITAN AI data center service line.