Canninah Dladla | Cluster President for English-speaking Africa | Schneider Electric | mail me |
Artificial Intelligence (AI) infrastructure has arrived, and with it comes a new reality. Power-hungry chips like NVIDIA’s Blackwell have pushed traditional air cooling beyond its limits.
Indeed, rack densities are rising from 20 kW to more than 140 kW. The 1 MW milestone is also approaching. As a result, the industry is now moving into the liquid cooling deployment phase for AI workloads.
Challenges of direct liquid cooling deployment
Direct-to-chip liquid cooling (DLC) answers the call. It delivers coolant directly to processors and enables efficient heat removal in high-density AI environments. However, moving from a pilot project to full-scale production introduces complex engineering challenges.
For one, liquid cooling requires much closer integration between facility infrastructure and IT hardware. Together, they become a single interconnected system.
Also, unlike air cooling, where airflow adjustments can often resolve thermal issues after deployment, liquid cooling requires precision from specification through operation. Small mistakes in fluid chemistry, pressure management or system design can result in hardware damage, warranty issues and reduced efficiency.
As organisations scale these systems, managing direct liquid cooling deployment risks becomes essential. Operators must address these risks before they affect performance, reliability or equipment.
Material compatibility and corrosion risks
Material incompatibility between the Cooling Distribution Units (CDUs) and connected components increases the risk of galvanic corrosion. This corrosion can create debris that clogs cold plates and damages servers.
When two or more metals with different electrochemical potentials connect in the presence of a fluid, they trigger galvanic corrosion. These metals become incompatible when the difference becomes large enough to cause issues. Such issues include material degradation and redeposition.
Operators must minimise the difference in the anodic index between the materials used for all wetted surfaces. This means strictly following the IT manufacturer’s guidelines for compatible materials. Operators should also refer to acceptable metals listed by organisations like the Open Compute Project.
Operators should also keep a comprehensive registry of all materials used. They should avoid materials like aluminium unless they use proper inhibitors. In addition, they should consider separate cooling loops if cold plate materials vary across different IT vendors.
Balancing efficiency gains with infrastructure costs
Liquid cooling allows higher chilled water temperatures. This improves chiller efficiency. However, achieving maximum energy savings may require investment in a dedicated high-temperature chilled water plant.
Operators typically choose between:
- Using existing low-temperature chillers and sacrificing some efficiency gains.
- Building a dedicated high-temperature chiller plant for liquid-cooled loads.
A total cost of ownership analysis over 10 years or more can help determine the best approach. Climate conditions also play a major role.
Increased system integration complexity
Direct liquid cooling physically connects servers to cooling infrastructure. This introduces new considerations around piping, coolant flow, leak prevention and cooling distribution.
Successful deployments, however, require collaboration between IT and infrastructure suppliers. They also require compatible connections, dedicated cooling loops and advanced leak detection systems. Pressure control and flow balancing are essential for maintaining stable cooling performance.
This integration makes managing direct liquid cooling deployment risks a shared responsibility. IT and facilities teams must therefore coordinate throughout design, installation and operation.
Lack of liquid cooling standards
Unlike air cooling, liquid cooling does not yet have widely adopted efficiency and capacity standards for CDUs. Performance ratings can vary depending on fluid type, temperature and flow conditions. This makes comparisons difficult.
Until formal standards become established, operators should request vendor testing data based on emerging industry guidelines. They should also limit CDU loop capacity to reduce the impact of potential failures.
Designing for unknown future IT loads
Data centres increasingly need to prepare for liquid-cooled IT without knowing the exact future equipment mix. Because DLC connects cooling directly to each server, systems must support a wide range of rack densities.
A flexible design approach includes sizing the main piping for overall capacity. At the same time, operators should design branches and manifolds for maximum rack density. Developing adaptable liquid cooling “pods” can also help manage future requirements.
Managing contamination risks during installation
Contamination risks weigh far heavier on the Technology Cooling System (TCS) loop than on the Facility Water System (FWS). This is because contamination can directly affect IT equipment.
During installation, vulnerabilities can arise from mismatched materials in connectors, piping and manifolds. They can also result from unsuitable fluid sealants on threaded joints. Microbial growth can develop at stagnation points within the TCS coolant.
Operators can mitigate these risks through proactive design and disciplined installation practices. Keeping cooling loops smaller helps limit the spread of contamination. Components should also be assembled in clean environments. Operators should add biocide before the initial fill. The system should then be thoroughly flushed, with all air purged before connecting IT equipment.
If installation is delayed, operators must keep the coolant in motion. They should circulate and filter it continuously. This prevents stagnant fluid from becoming a breeding ground for bacteria.
These measures form an important part of managing direct liquid cooling deployment risks. They also reduce the likelihood that installation-related contamination will compromise IT equipment.
Warranty and responsibility challenges
As DLC connects cooling systems directly to IT hardware, warranty responsibilities can become unclear. Server manufacturers, cooling suppliers and installers may each have different responsibilities.
Operators should therefore establish agreements on cooling temperatures, flow rates and pressures. They should define maintenance responsibilities early. They must also ensure that warranty conditions do not conflict across suppliers.
Managing GPU power spikes
AI workloads drive rapid changes in GPU power consumption. These changes can often outpace the response time of conventional cooling systems.
Software-imposed power limits can reduce this risk, but they can also reduce performance. A stronger safeguard is to stage workloads gradually. Operators should monitor GPU temperatures closely. They should also fine-tune coolant flow or temperature as needed.
Building the future of AI-ready infrastructure
Moving from liquid cooling pilots to production environments requires a fundamental change in how data centres are designed, installed and operated. AI-ready infrastructure leaves little room for disconnected teams or isolated systems.
Ultimately, success depends on integrating IT requirements with facility capabilities from the beginning. Organisations must also maintain a disciplined approach to specification, installation and operations.
By doing so, organisations can manage the complexity of liquid cooling. They can also strengthen managing direct liquid cooling deployment risks as a core operational discipline. This approach can unlock the performance and efficiency needed for the next generation of AI workloads.
