Liquid cooling's next frontier: rethinking heat capture at the chip level

As AI workloads drive chip processing power and rack densities exponentially upwards, liquid cooling has quickly transitioned from an emerging technology to a critical part of modern data center design. For many operators, the conversation is now shifting toward how liquid cooling systems can evolve to support the next generation of AI compute. And a key focus is how efficiently heat can be captured at the source.

During a recent discussion with Data Centre Magazine, Bill Flaherty – PhD, Senior Director Data Center Thermal Management for Johnson Controls – explored how innovation at the cold plate level can help data centers support growing AI demands.

Throughout the conversation, Flaherty touched on several key topics, including:

• Why innovative thermal management means solving for the entire thermal chain – from chip to ambient – and the role of cold plates within that system
• The specific limitations of legacy cold plate technology and where those technologies are failing to meet the demands of modern data centers
• How microcapillary cold plates are used to target high density hot spots within a GPU and enhance heat transfer
• What single-piece manufacturing means for yield consistency and the ability to scale at pace with data centers that are constantly changing
• The single most important question that operators in hyperscale facilities should be asking about their cold plate technology

AI is pushing traditional cooling approaches to their limits

Modern GPUs generate significantly higher power densities than previous generations, and heat is often distributed unevenly across the surface of a chip. As a result, cooling systems that were effective just a few years ago can struggle to keep pace with today's AI environments.

According to Flaherty, one of the clearest signs that cooling infrastructure is reaching its limit is when operators are forced to throttle performance to keep chips within acceptable operating conditions. This means that operators slow down the operating speed and reduce power to lower temperatures before it gets too hot. When operators are forced to do this, organizations aren’t getting full value from their compute investments.

"When you talk about ‘what does a data center care about,’ today it's how many tokens am I generating at any one time,” Flaherty said. “Because that’s how they’re making their money – by selling those tokens to their customers. So, the faster you run your GPU, the faster you run your system, the more tokens you're creating for the same investment in that capital equipment."

The next opportunity may be at the point of heat capture

As liquid cooling adoption grows, industry attention is beginning to move further along the thermal chain. While cooling discussions often focus on CDUs, facility water loops or heat rejection equipment, every watt of heat begins its journey at the chip itself. The more efficiently that heat can be captured and transferred into the liquid cooling system, the more effectively the rest of the thermal chain operates.

This is where cold plate innovation is becoming increasingly important. Traditional cold plate designs were developed for an earlier generation of compute infrastructure. As chip architectures evolve and thermal loads become more concentrated, operators are looking for technologies that can target hot spots more precisely and support increasingly dense AI environments.

During the conversation, Flaherty spoke about how microcapillary technology is being used to create highly customized cold plates that target high density hot spots within the GPU. For starters, microcapillary cold plates use novel microgeometries to reduce thermal resistance by up to 35% while cutting pressure drop by up to four times compared with incumbent designs such as skived fin. This allows organizations to remove heat precisely where it is needed, maximizing cooling efficiency. 

Why the entire thermal chain matters

The webinar also highlighted a broader shift occurring across the industry – the move toward integrated thermal management solutions. Chip-level cooling, coolant distribution infrastructure, facility cooling systems and heat rejection equipment all influence one another. Optimizing one component in isolation is becoming increasingly difficult as AI workloads scale.

This, as Flaherty describes, comes down to simplifying your thermal management approach and gaining greater control of the entire thermal chain.

“When you talk about a modern data center, having a full chain of solutions is critical because it allows you to optimize,” Flaherty said. “When you're trying to cobble together solutions from multiple different vendors, you have to do that optimization yourself. You're never going to get it quite right because you don't have the same level of understanding of those systems that the people that designed them and built them do.”

Every watt of heat generated by a chip must travel through a broader thermal chain that includes the cold plate, coolant distribution infrastructure, facility cooling systems and – ultimately – heat rejection to the atmosphere. Improvements at any point in that chain can create operational benefits throughout the system.

This shift to integrated solutions reflects a broader change in how the industry approaches thermal management. Rather than viewing cooling as a collection of individual components, operators are increasingly evaluating how the entire thermal chain works together to support performance, efficiency and scalability.

FAQs

What is direct-to-chip liquid cooling?

Direct-to-chip cooling uses liquid-filled cold plates that are mounted directly onto CPUs, GPUs, and now peripheral devices like DIMMs, NICs, and QSFPs to capture heat at the component level. Heat is then carried along to an external heat exchanger or cooling unit. 

Why has liquid cooling become a critical part of data center design?

Liquid cooling has become a critical part of modern data center design because high-density AI workloads generate more heat than traditional air-cooling systems can handle. Air is limited to how much heat it can hold and carry. Once rack densities climb past 40kW, air alone struggles to keep component temperatures within safe operating ranges.  For this reason, liquid cooling has increased in prominence – especially when it comes to capturing heat from high-density GPUs. 

What is heat absorption in data centers?

Heat absorption in data centers is the process used to capture, remove and transfer thermal energy or heat generated by servers and IT equipment. The main types of heat absorption and cooling include air cooling, direct-to-chip liquid cooling and immersion cooling.