AI workloads are driving a whole new level of heat and power demands. As compute continues to scale, thermal constraints and power consumption are becoming increasingly intertwined. Together, they represent one of the most significant limiting factors to data center growth.

This shift means efficiency can no longer be approached incrementally or addressed in isolation. Instead, it must be embedded from the outset, designed into the very foundations of data center infrastructure. For ZutaCore, this means fundamentally rethinking how heat is removed – starting at the source.

In a recent DCD>Talks episode, Brian Lillie, president and COO at ZutaCore, explores how rising AI-driven heat densities are reshaping infrastructure requirements, and why waterless, direct-to-chip cooling is emerging as a critical enabler of next-generation, high-density deployments.

Heat rises

“Heat is no longer a theoretical problem,” says Lillie. “It’s now the primary constraint on AI infrastructure growth. And on the other side of that same coin is power.”

The rack densities that once defined the typical data center have quickly been overshadowed by the scale of modern operations. Today’s AI workloads demand significantly higher power densities, pushing infrastructure far beyond the limits of long-held design assumptions.

As a result, legacy cooling approaches, such as air, are struggling to keep pace. Not only are they less effective at removing extreme heat from high-density racks, but they also introduce inefficiencies in terms of internal rack space and system design. Lillie believes that even some newer liquid cooling approaches may soon face similar challenges:

“Single-phase has the capability today, but it’s going to hit a thermal wall. The pace is accelerating and the heat flux is growing. It’s a problem now, but it’s going to be an even bigger problem in the next few years.”

A direct approach

As densities rise, removing heat directly from the silicon is becoming increasingly essential. Direct-to-chip liquid cooling offers a more efficient route, but not all approaches are created equal.

Traditional single-phase liquid cooling systems – typically using water or water-based solutions – work by flowing coolant across the chip to absorb and carry away heat. While this is up to scratch for current densities, Lillie explains why these systems are at risk of falling behind:

“As the heat flux increases, you have to move colder liquid faster,” explains Lillie. “That means higher pressure, higher flow rates – and eventually that just becomes unsustainable.”

Two-phase, direct-to-chip cooling offers an effective alternative. Rather than relying solely on liquid flow, it leverages phase change, where the coolant evaporates as it absorbs heat to dramatically improve heat transfer efficiency.

“We use a dielectric fluid, not water,” adds Lillie. “That avoids many of the challenges associated with water and electronics.”

In this approach, heat is absorbed at the chip and converted into vapor, which is then removed at low pressure, allowing for significantly lower flow rates and reduced mechanical demand on the system.

By removing heat at the source, two-phase systems eliminate many of the inefficiencies associated with indirect cooling methods. This not only enables higher-performance chips to operate effectively but also reduces overall energy consumption.

“It’s a fundamentally more efficient approach,” says Lillie. “And it’s one that can scale both technically and economically as densities continue to rise.”

Scaling within constraints

Beyond cooling performance, operators are increasingly focused on how to scale efficiently within existing footprints and set constraints – particularly when it comes to upgrading legacy infrastructure.

“It’s not just about cooling higher densities,” explains Lillie. “It’s about doing it at scale, with certainty.”

The challenge lies in deploying next-generation hardware without requiring a complete redesign of existing facilities. As a result, modularity and retrofit compatibility are becoming key priorities.

“We’re seeing strong demand for solutions that can be deployed in existing data centers,” he continues. “How do you reach these new thermal thresholds without starting from scratch? How do you extend the useful life of your assets? These are the kind of common questions we’re hearing right now.”

Given the vast amount of infrastructure already in operation, extending the lifespan of these assets is not only economically advantageous, but also critical from a sustainability perspective.

Thirsty facilities

Power isn’t the only resource under pressure. Increasingly, water availability is emerging as a key constraint on data center development – as well as a significant source of public and regulatory scrutiny.

Projects are at risk of being delayed not just due to power limitations, but because of concerns around water usage. This is particularly challenging in regions where water scarcity is already a critical issue.

“Cooling can represent a significant portion of total data center energy use,” says Lillie. “And when you look at overall efficiency via metrics like power usage effectiveness (PUE), it becomes clear how important it is to optimize the entire system.”

Improving efficiency isn’t just about reducing energy consumption, but also about minimizing resource usage more broadly across facilities.

“If you can use less energy and less water, that’s a win on both fronts,” explains Lillie. “For example, a 1MW single-phase CDU can require around 70kW just to operate. That’s a significant overhead.”

Reducing this auxiliary load has a direct impact on overall efficiency, freeing up more power for compute while lowering operational costs.

The water issue also points to a broader industry-wide conversation around how data centers can engage with local communities and stakeholders as they expand into new regions.

“When you’re building in a community, you need to be mindful of the resources you’re using,” says Lillie. “Water is for humans, not servers.”

Efficiency as standard

As the industry looks to the future, it’s clear that efficiency must come built-in from the earliest stages of design.

“For new builds, power and cooling have to be at the center,” says Lillie. “It’s not something you add later, it’s something you design around.”

This requires a holistic approach that spans the entire technology stack – from silicon to system architecture to facility infrastructure.

“It starts at the chip,” explains Lillie. “Silicon providers need to design with thermal considerations in mind. That then extends to how those chips are packaged, and ultimately how they’re integrated into the wider system.”

This end-to-end perspective is becoming essential as data centers evolve into increasingly complex, high-density environments that exist inside established communities.

The pace of change in AI infrastructure leaves little time or space for incremental adaptation purely at the component level. Instead, the industry must adopt a more integrated, forward-looking approach to design that prioritizes efficiency, scalability, and sustainability from day one.

Lillie concludes with his distinct vision of this path forward: “I don’t see any way around a fully liquid-cooled future. It starts at the chip and extends all the way through the system. That’s going to be the key factor in designing data centers for the next generation.”

To hear more about cooling for the high-density era, watch the full DCD>Talks episode with Brian Lillie, here.