The Operating Model for
Liquid-Cooled AI Infrastructure
Is Overdue for a Rebuild.
When a single rack carries $3.5M in GPU assets and runs 24/7, your old operating model is the risk. This guide was written for the teams responsible for keeping those systems up and performing.
This guide covers
The Operating Environment
Most incidents start as something small.
A reading no one acted on. A procedure that hadn't been updated. An alarm that went to the wrong team. In liquid-cooled AI environments, those small misses compound fast. Here's what that looks like in practice.
The Midnight Thermal Drift
CDU telemetry showed a 2°C rise in coolant supply temperature, but only Facilities had visibility into that data. IT saw GPU throttling and assumed it was the workload. By the time both teams figured out what was actually happening, performance had dropped 18% and the training window was gone. Forty-seven minutes of drift that nobody caught.
Thermal drift hits performance long before it triggers an alarm.
The Procedure That Changed at 3 A.M.
A tech pulled up the SOP for a coolant filter change, ran the procedure, and restarted the rack. What he didn't know was that the updated version, sitting in a different system, added a 10-minute stabilization step before restart. The pressure imbalance tripped a CDU alarm. Ninety minutes of downtime and a full cold plate inspection followed.
A missing step in a liquid-cooled environment is a failure mode.
The Ownership Gap That Cost an Hour
A rack thermal alarm fired during an active training run. Facilities assumed IT was handling it. IT assumed it was a cooling issue. The vendor didn't get pulled in until 30 minutes later. With no defined escalation path, the rack throttled twice and the job had to restart from a checkpoint.
In AI/HPC, response time is revenue.
The Tool Problem
DCIM. CMMS. Ticketing systems. None of them were built for this.
These tools were designed for environments where IT and Facilities could work independently. That model is gone. Liquid-cooled AI infrastructure needs one system where telemetry, procedures, and team coordination actually connect.
DCIM shows you what's happening. It doesn't tell you what to do.
When an alarm fires, your team needs a guided response, not a dashboard.
CMMS manages work orders. It can't run a cross-team incident response.
Scheduled maintenance and real-time coordination are different problems entirely.
Ticketing systems record what went wrong after the fact.
You can't ticket your way out of a cooling failure that's actively degrading a training run.
"By 2027, more than 60% of new enterprise-class AI infrastructure deployments will require some form of direct liquid cooling. That's a complete reversal from where things stood just five years ago."
Mike Parks, CEO, MCIM
A single liquid-cooled AI rack now represents:
Operational Maturity
The infrastructure isn't the problem. The operating model is.
Most teams running liquid-cooled AI infrastructure are doing so with processes and tools built for air-cooled rooms. The guide walks through a five-level maturity model so you can see exactly where your operation stands and what needs to change.
Running liquid cooling with an air-cooled playbook. Siloed teams, tribal knowledge, no shared view.
SOPs exist but enforcement is inconsistent. Monitoring is in place but nobody's connecting the signals.
IT and Facilities are working from shared workflows. Response is faster, but still reliant on manual steps.
Telemetry feeds directly into operational workflows. Problems surface before they escalate into incidents.
One system of record across IT, Facilities, and vendors. No silos, no blind spots, nothing left to chance.
Get the full guide
The complete framework for deploying and operating liquid cooling at scale.