A Q&A on AI-optimised cooling systems with Phil Cullerton, Vice President of Services at Vertiv.
Executive Summary
- Vertiv’s Vice President of Services, Phil Cullerton, spoke to DCi in an interview about AI-optimised cooling systems, how Vertiv are using machine learning to predict thermal events and what other innovations can be used to optimise cooling.
- Cooling optimisation is more about maintaining equilibrium across energy, acoustics and reliability. Using AI-optimised cooling systems and predictive service models that close the loop between field data and design allows for continuous refinement of cooling logic based on real-world operational insight.
- The appropriate level of human involvement in these autonomous systems is insight; engineers maintain strategic oversight through dashboards and digital twins that show precisely what each algorithm is doing. Every automated action is logged, auditable and reversible, so accountability remains human even when optimisation is automated.
1. Traditional cooling systems react to temperature changes after they occur. AI workloads shift from idle to maximum power density in seconds, creating thermal transients that reactive systems cannot manage. How are you using machine learning to predict thermal events before they happen, and what data inputs enable effective predictive thermal management for AI manufacturing?
AI-intensive workloads generate rapid, uneven heat signatures that traditional reactive systems can’t handle. Our service teams use machine learning models trained on extensive data points (temperature gradients, compressor vibration, pump speed, rack load) to anticipate changes seconds before they occur. The system identifies subtle deviations from normal operation and pre-emptively adjusts fluid flow or fan speed to stabilise conditions. This predictive capability helps prevent hot spots before they form and reduces the energy penalty of over-cooling.
Machine learning lets us understand cause and effect, not just measure outcomes. If an algorithm recognises that a GPU ramp-up will lift pressure in a given loop, it makes that correction instantly rather than waiting for an alarm.
It’s a continuous learning process. The more data the system analyses, the more accurately it predicts future states – a perfect partnership between service expertise and smart automation.
2. Smart environmental controls for AI facilities must balance multiple competing objectives: minimising energy consumption, preventing hotspots, managing noise levels, and maintaining precise temperature ranges. How do AI-driven cooling optimisation systems make these trade-offs in real-time, and what happens when optimisation algorithms make the wrong decision in a mission-critical AI training environment?
In AI-scale environments, cooling optimisation isn’t just about temperature control – it’s about maintaining equilibrium across energy, acoustics and reliability. Our control logic evaluates thousands of variables in real time and calculates the most efficient balance between power draw and performance. If noise thresholds or efficiency limits are at risk, the system
automatically recalibrates within defined safety bands. Expert engineers remain in the loop, constantly looking to align optimisation with uptime, efficiency and regulatory priorities. Our systems are designed to comply with emerging AI governance and data-integrity standards, keeping the human firmly in control loop.
When algorithms misjudge, we have protective layers that limit their impact. Each decision is traceable, and operators can instantly override. Essentially it is autonomy supported by transparency.
Because Vertiv manages both infrastructure hardware and digital logic, we can simulate potential scenarios in advance. That learning allows AI-driven systems to correct themselves faster and with less energy waste.
3. AI infrastructure operates at power densities of 60-150kW per rack today, with some facilities planning for 200kW+. At what power density threshold do traditional cooling control systems become inadequate, and what specific AI-optimised capabilities are required beyond that point?
Typically, beyond ~50–60 kW per rack, air-only systems start losing efficiency because the thermal load fluctuates too quickly. At that point, predictive, AI-optimised control with liquid or hybrid cooling becomes essential. Beyond 100 kW per rack, sensors must feed live data to models that forecast fluid behaviour, compressor response and energy flow across loops.
These higher-density environments demand real-time control of pump speed, fluid chemistry and loop pressure. Our engineers can see these metrics remotely and act before performance drifts.
We design each installation around how AI workloads actually behave – dynamic, burst-based, unpredictable – rather than relying on static set-points.
4. Predictive maintenance for cooling systems promises to prevent failures before they occur, but AI training clusters cannot tolerate even brief thermal events. How do you balance predictive maintenance that might require taking cooling capacity offline against the zero-tolerance requirements of AI manufacturing operations?
Predictive maintenance is valuable only if it protects uptime, not disrupts it. We achieve this by combining asset health scoring with live capacity modelling, so service windows can be scheduled precisely when resilience is highest. If a compressor shows performance drift, we
transfer the load to redundant units and replace the component under protection. It’s maintenance in motion: continuous operation with predictive intelligence built in.
Condition-based maintenance eliminates unnecessary interventions. We act when data shows risk, not because a calendar says it’s time.
Every predictive decision is verified against real operating conditions, so we don’t sacrifice availability for diagnostics.
5. Most data centres have cooling systems from multiple vendors deployed over the years, creating a heterogeneous environment with disparate control systems. How critical is unified cooling control and AI-driven optimisation across this complexity, and what integration challenges prevent facilities from achieving optimal thermal management?
Many of the data centres we support have evolved through multiple generations of equipment. Our role is to bring that diversity under one intelligent control layer, giving operators a unified view of their entire cooling ecosystem. When AI-driven optimisation spans across disparate systems, it prevents inefficiencies such as simultaneous over-cooling and under-cooling, cutting waste and improving overall resilience.
Integration challenges are mainly about data integrity. Different protocols, calibration standards and sensor accuracy levels need harmonising before meaningful optimisation can occur. We achieve integration through open protocols and secure data sharing frameworks, enabling vendor interoperability without compromising system security.
With unified visibility, operators can finally make proactive, site-wide decisions rather than firefighting individual alarms.
6. Automated thermal management systems make thousands of control decisions per minute. What level of human oversight is appropriate for AI-driven cooling systems in mission-critical facilities, and how do you design systems that can operate autonomously while still allowing operators to intervene when necessary?
Autonomy works best when paired with human insight. Predictive service models close the loop between field data and design, allowing continuous refinement of cooling logic based on real-world operational insight. Automation manages the data volume and people manage the context. That combination delivers speed and safety in equal measure.
Our systems make thousands of micro-adjustments per minute, but engineers maintain strategic oversight through dashboards and digital twins that show precisely what each algorithm is doing. Every automated action is logged, auditable and reversible, so accountability remains human even when optimisation is automated.
We design layered control logic so that if external conditions change – a grid event or workload shift – operators can intervene immediately without losing the benefits of continuous optimisation.
7. What other AI-optimised cooling capabilities or thermal management innovations should we be considering that we haven’t addressed in these questions?
We’re moving toward a fully integrated lifecycle model, where insights from design, commissioning, operation and maintenance feed into each other through a continuous digital thread. By learning from live site data, we can refine fluid loop design, control logic and service scheduling for future builds. This closes the loop between engineering and operation – turning every project into a living feedback system.
Waste-heat recovery will also become central to the cooling conversation. AI sites generate vast, stable heat outputs that can be reused for district heating or industrial processes, shifting sustainability from obligation to opportunity.
The other major advance is the use of digital twins for cooling ecosystems. They let us test conditions, run fault simulations and optimise configurations without touching the live system, which is the safest way to manage complexity as densities climb.
Ultimately, the goal is not just smarter cooling, but a self-optimising thermal ecosystem – one that evolves with every workload generation. That’s where we’re heading…
