In recent years, liquid cooling has undergone a significant transformation. What was once viewed as a specialist solution reserved for the highest-density computing environments, has rapidly become one of the most critical technologies enabling AI data center buildouts. As rack densities continue to rise at pace, the conversation around liquid cooling is intensifying and evolving, too.

Today, the challenge extends well beyond asking, ‘should I adopt liquid cooling?’ Attention is turning to how these systems can be produced and deployed at unprecedented scale.

Delivering tens of megawatts of liquid-cooled capacity across multiple campuses and geographies demands more than performance alone. Manufacturing capability, resilient supply chains, deployment speed, operational consistency, and long-term flexibility have all become equally important pieces of the puzzle.

"The magnitude and scale of these buildouts is historic in nature," says Joe Capes, vice president at Trane Technologies and general manager of LiquidStack. "Serving the global marketplace requires sufficient regionalized manufacturing and supply chain scale to build at the kind of volumes operators are looking for. Equally, you also need specialized global service capability to support everything from installation and commissioning through to long-term lifecycle management."

Against this backdrop, operators are increasingly looking beyond individual cooling products towards architectures capable of evolving alongside future generations of AI hardware. For Capes, the next chapter in liquid cooling will be defined by platforms that make large-scale deployment more repeatable, adaptable, and efficient.

Redefined by scale

Early liquid cooling deployments were typically designed around the compute requirements inside individual facilities. While this bespoke approach made sense at lower scale, AI infrastructure is forcing a more nuanced strategy.

"We're seeing a really wide range in what operators define as scalability," says Capes. "Some organizations prefer relatively low capacity technology cooling system loops, while others are now designing for 10MW, 20MW and even 30MW deployments. What scalability means to one operator or tenant doesn't necessarily mean the same thing to another."

Crucially, managing rising density isn’t just about adding more cooling distribution units (CDUs). Each additional standalone CDU introduces another set of controllers, pumps, expansion tanks, and associated infrastructure. Redundancy remains essential, but endlessly replicating discrete CDUs can quickly introduce unnecessary costs and complexity. Discrete CDUs can also consume an unnecessary amount of space. As Capes explains, it’s not always the more the merrier:

"One way to help solve larger capacity deployments is by running multiple one or two megawatt CDUs on a common TCS ring. But that means stacking up unnecessary duplication of CDU components that not only adds cost, it can also potentially add risk.”

The challenge is compounded by the differing lifecycles of infrastructure and hardware. Cooling systems are expected to remain operational for well over a decade, while AI accelerators continue to evolve at a faster pace. Each successive generation arrives with higher thermal design powers (TDPs), making it increasingly important to design infrastructure that can accommodate tomorrow's hardware, rather than simply meeting today's requirements.

Thinking platforms, not products

As operators look to simplify deployment without sacrificing flexibility, modularity has become a defining solution across the data center, including liquid cooling. Yet, the term itself can be hard to define.

For some vendors, modularity simply involves deploying multiple identical CDUs alongside one another. While this can offer greater flexibility than fully bespoke engineering, it still relies on replicating complete systems each time additional cooling capacity is required.

LiquidStack is approaching the concept of modularity with a different perspective: "We don't refer to GigaModular as a product," says Capes. "It's a platform, an architecture."

This distinction is an important one. Rather than treating every CDU as a standalone unit, the platform separates key functions into modular building blocks. Pump modules help provide scalable cooling capacity, while hydronic and control systems are shared across the wider installation, helping reduce unnecessary duplication without compromising resilience.

The result is an architecture capable of supporting cooling loops ranging from approximately 2MW to 14MW, all under the same unified platform.

"It provides the ability to solve different project needs using one platform rather than multiple products," explains Capes. "You may have one operator designing around a 6MW cooling loop and another building a 14MW loop, but rather than sourcing different CDU products for each project, they can standardize around the same platform."

For operators deploying AI infrastructure across multiple campuses, standardization can simplify engineering and potentially shorten commissioning times. It also enables teams to develop expertise around one common architecture, rather than managing a disjointed collection of individual systems.

Perhaps most importantly, a platform-based approach offers valuable future flexibility. It supports incremental expansion, helping to avoid the need to replace multiple disparate components each time thermal demands increase.

Predictability as a competitive advantage

The race to deploy AI infrastructure has turned speed into a commercial necessity. Delays could mean postponed revenue, placing renewed emphasis on both pace and reliability being built-in across design, manufacturing, and deployment.

While legacy custom engineering approaches increased flexibility, they also introduced variability into the supply chain and long-term operations.

"What we're seeing is an increasing desire for standardization, so deployments become predictable," says Capes. "A lot of the operators and tenants we're working with are deploying liquid cooling for the first time. Having that predictability and assurance can provide important benefits compared with deploying a bespoke solution that nobody has seen before.

“When you scale with GigaModular, it also means that you likely won’t be replacing your first-generation CDUs the next time a semiconductor company does a new GPU introduction."

Repeatable architectures deliver efficiencies throughout the project lifecycle. Manufacturers benefit from more consistent production, installation teams work with familiar deployment models, and operators gain confidence that systems will perform consistently regardless of location.

Building for a global ecosystem

Scalability is increasingly being defined by whether suppliers can deliver the infrastructure consistently across multiple regions, often at the same time.

AI investment is spreading rapidly beyond established data center markets, placing a new level of demand on manufacturing networks and supply chains across the globe. In this environment, delivery capability is a key differentiator.

Recognizing this shift, LiquidStack has developed its GigaModular platform around a global manufacturing strategy, with localization capability. Rather than shipping fully assembled systems across the globe from one manufacturing location, core modules can be constructed closer to the project destination before final integration is completed on site.

"The idea is to have a global supply chain, manufacture regionally and integrate locally," says Capes. “We can build where it makes the most sense for the customer and then complete the integration. This equates to a tailored supply chain that our customers can appreciate."

This approach delivers multiple advantages. Regional manufacturing can help reduce transportation costs, potentially shorten lead times, and strengthen supply chain resilience, while enabling operators to bring new capacity online more quickly.

It also helps allow suppliers to respond more effectively to local market requirements without sacrificing consistency across time zones. For Capes, this reflects a broader shift and focus on supply chain sustainability taking place across the industry:

"This is a global ecosystem. The companies that succeed will be the ones with the global footprint to deliver products and services wherever they're needed, with sustainability as a focus."

Designing for hardware that hasn't arrived yet

Perhaps the greatest challenge facing infrastructure designers is planning for hardware that remains several product generations away. While operators have a clear understanding of today's thermal requirements, predicting the demands of tomorrow's AI accelerators is much more difficult.

What’s certain is the direction of increasing densities. Power demands and heat outputs will inevitably continue to rise, and AI workloads may place greater stress on supporting infrastructure with each successive generation of hardware.

At the same time, few operators are designing facilities that rely entirely on liquid cooling. Instead, most new data centers are expected to operate hybrid environments for the foreseeable future, combining traditional air-cooled infrastructure with liquid cooling systems. Capes expands:

"Air-cooled heat loads are going to be here for a long time, if not forever. Designing for the future means understanding liquid cooling will become a much larger part of the infrastructure, while recognizing there will still be hybrid environments."

This introduces an additional layer of complexity. Operators must deploy infrastructure capable of supporting today's AI hardware while retaining the flexibility to help accommodate even more demanding systems as they emerge.

Platform-based architectures offer one potential solution. By allowing cooling capacity to be expanded incrementally, operators can adapt infrastructure alongside changing hardware requirements without replacing entire systems.

"What we do know is that chip TDPs and rack densities are going to continue increasing," continues Capes. "By designing around a platform, an architecture, rather than a discrete product, we believe we've created something with a much longer runway that can meet future design requirements for chips that haven't even been announced yet."

Intelligent infrastructure

As cooling systems become increasingly connected, the operational data they generate is expected to play a much greater role in improving efficiency, reliability and maintenance.

Today's platforms already collect significant amounts of performance data to monitor equipment health and support fault diagnosis. Over time, that same information could underpin predictive maintenance strategies capable of identifying potential failures before they get the chance to affect operations.

"We already collect operational data for fault analysis," says Capes. "But over time, there are greater opportunities that help to diagnose potential failures before they happen, for instance data we collect would help identify when a pump is likely to fail months before it actually does.

"When integrated with AI, proactive service can also be scheduled to help with uptime, and potentially reduce operational risks."

Cooling infrastructure itself could become increasingly autonomous, using real-time operational data to help improve energy efficiency and respond dynamically to changing workloads.

"As an industry, we're focused on delivering AI infrastructure today," adds Capes. "But the future holds a lot of opportunity for infrastructure that actually uses AI to become more efficient and more reliable."

Although these capabilities remain at an early stage, they point towards a future where cooling systems become active participants in wider data center operations, helping to improve resilience while reducing operational overheads.

Broadening horizons

Liquid cooling has firmly established itself as a cornerstone of AI infrastructure, but the industry's priorities are continuing to evolve. Today, cooling innovation must go hand in hand with streamlining its execution.

Meeting this challenge requires a broader definition of scalability. Thermal performance remains fundamental, but it now sits alongside manufacturing capability, supply chain resilience, operational consistency, and the flexibility to accommodate hardware generations that have yet to be announced.

For LiquidStack, and Trane Technologies, this means viewing modularity as the foundation of an architecture intended to scale with both customer requirements and the rapid evolution of AI hardware.

In an industry moving at breakneck speed, the ability to deliver consistent, adaptable infrastructure may prove just as valuable as achieving the next breakthrough in cooling performance.

For more information about the LiquidStack GigaModular platform, please visit https://liquidstack.com/cdu-direct-to-chip/giga-modular