Matilda Bailey has spent her career at the intersection of high-speed networking and thermal management, witnessing firsthand how the heat generated by modern AI chips threatens to outpace the evolution of the data centers housing them. As a specialist in next-gen solutions, she advocates for a shift away from traditional cooling overhead toward integrated systems that treat thermal management as a primary lever for performance. This conversation explores the transition to adaptive phase cooling, the operational realities of “waterless” data centers, and why the future of compute density might rely on decentralized, mid-scale facilities rather than massive gigascale projects. We delve into the mechanics of fanless server architecture, the significant efficiency gains found in reducing chip leakage, and the collaborative manufacturing efforts required to keep pace with the rapidly shortening hardware cycles of the AI era.
Adaptive phase cooling techniques often draw from thermal engineering used in nuclear reactors. How do you successfully remove server fans while maintaining thermal stability, and what specific mechanical adjustments are required to repackage standard servers into these specialized rack-mounted units?
The shift from traditional forced-air cooling to adaptive phase cooling is a fundamental reimagining of how we manage the “thermal heartbeat” of a server. In a standard setup, you have banks of high-speed fans consuming roughly 10% of the server’s power, creating a constant, high-pitched mechanical whine and a lot of turbulent air that isn’t particularly efficient at moving heat. By drawing on techniques used in nuclear reactors, we replace that chaotic airflow with a passive, rack-mounted system that utilizes phase change to move heat away from the silicon with surgical precision. To successfully remove the fans, we have to repackage the server into a specialized unit where the heat-generating components are in direct contact with the adaptive phase cooling hardware, which has no moving parts. This mechanical adjustment is a meticulous process that happens at the pilot stage over a few weeks, ensuring that every thermal interface is optimized so that the coolant can absorb energy as it transitions from liquid to vapor. It’s a transition from the mechanical noise of a wind tunnel to the quiet, steady reliability of a closed-loop system that keeps the silicon at a much more consistent, lower temperature than air ever could.
Many operators face rigid power envelopes and rising energy costs. Since removing fans and reducing chip leakage can yield a 15% efficiency gain, how does this translate into actual compute density for a facility, and what specific metrics should managers track to validate these performance improvements?
When you are trapped within a fixed power envelope, every watt spent on a fan is a watt that isn’t performing a calculation. We consistently see that about 10% of a server’s power is simply feeding those mechanical fans; by eliminating them, you immediately free up that power to support more compute nodes in the same rack. Furthermore, because we run the chips at a significantly lower and more stable temperature, we reduce “leakage” current, which accounts for another 4% to 5% of efficiency, allowing the chips themselves to run faster and more reliably. For a facility manager, this translates into being able to feed more servers from the same utility drop, essentially getting roughly 35% more total compute when you factor in facility-level overhead reductions. To validate this, managers should move beyond simple PUE and start tracking “Compute per Watt” at the rack level, while also monitoring the reduction in chip throttling events that typically occur during peak thermal loads. It is incredibly satisfying to look at a power meter and realize that almost every amp entering the building is actually reaching a processor rather than being wasted on spinning plastic blades.
Transitioning to waterless cooling eliminates the need for heavy infrastructure like chillers and towers. What are the primary CAPEX trade-offs when shifting to this model, and how does reaching a 1.03 PUE change the long-term sustainability profile and water consumption of a typical 10 MW facility?
The CAPEX shift is one of the most compelling parts of this story because you are essentially deleting the most expensive and heavy parts of a traditional data center build. By moving to a waterless, rack-mounted adaptive phase system, you can strip out the massive cooling towers, the complex piping for chilled water loops, and the heavy chillers that usually sit on the roof or in a separate mechanical gallery. While the specialized rack-mounted units themselves have a base cost comparable to high-end direct-to-chip liquid cooling, the savings realized by not building a massive water-intensive infrastructure are enormous. When you reach a PUE of 1.03 or 1.04, you aren’t just saving money; you are fundamentally changing the facility’s relationship with its environment, as a 10 MW facility can operate without draining millions of gallons of local water annually. This sustainability profile is becoming a “must-have” rather than a “nice-to-have” as local governments increasingly push back against the water demands of massive new data center developments.
Maintaining specialized hardware can be complex for data center staff. If a rack-mounted cooling unit experiences fluid loss, what is the step-by-step protocol for servicing it without disrupting the entire rack, and how do you monitor gradual temperature shifts to prevent sudden hardware failure?
One of the biggest fears in liquid cooling is a catastrophic “spray” or a sudden failure, but adaptive phase cooling is designed to fail gracefully, which is a relief for on-site technicians. Because the system is rack-mounted and modular, the protocol for a unit experiencing fluid loss is quite similar to swapping out a standard server: you identify the specific unit, slide it out of the rack for service, and replace it, all while the surrounding units continue to operate unaffected. We don’t see the sudden thermal spikes that you get when an air-cooled fan dies; instead, because the fluid has high thermal mass, the temperature rises very gradually over time. Technicians monitor these thermal trends through a centralized dashboard that flags any deviation from the baseline long before the chip reaches a critical threshold. This allows for scheduled maintenance rather than emergency “fire drills,” and the simplicity of having no moving parts means there is much less to go wrong in the first place compared to traditional pump-heavy systems.
Scaling hardware solutions requires deep collaboration with OEMs and ODMs. What does the integration process look like for high-volume manufacturing, and how do you ensure these customized cooling systems remain compatible with the rapidly evolving hardware cycles and power demands of modern AI chips?
To move beyond pilot programs and reach hyperscale, we have to integrate directly into the production lines of the world’s major Original Equipment Manufacturers (OEMs) and Original Design Manufacturers (ODMs). The integration process involves a deep dive into the physical layout of the server motherboards to ensure our adaptive phase units can be mated to the latest GPUs and CPUs right at the factory. This collaboration is essential because AI chip cycles are moving at a breakneck pace, with power demands increasing from 300W to 700W or even 1000W per chip in just a few generations. We design our cooling interfaces to be adaptable, ensuring that the thermal “envelope” we provide can handle the increased flux of next-gen silicon without requiring a total redesign of the rack infrastructure. By working with ODMs, we can ensure that these customized cooling solutions are built with the same rigorous quality controls as the servers themselves, making the transition to advanced cooling feel like a standard hardware refresh rather than a science experiment.
There is a growing interest in utilizing smaller, distributed 5-10 MW data centers rather than focusing solely on gigascale facilities. Why might a decentralized approach be more viable for AI workloads, and what logistical hurdles must be overcome to integrate these smaller sites into a network?
The industry has been obsessed with gigascale facilities, but we are reaching a point of diminishing returns where grid capacity simply cannot keep up with the demand for 100 MW or 500 MW sites in a single location. There is a massive opportunity in utilizing the “leftover” 5-10 MW data centers that are already scattered across the country, many of which have existing grid connections that are currently underutilized. A decentralized approach is viable for AI because many training and inference workloads can be distributed across a network, provided the cooling and power efficiency at these smaller sites are modernized to handle high-density racks. The logistical hurdles are primarily focused on networking and orchestration—ensuring that these disparate sites can act as a single, cohesive compute fabric with low latency and high reliability. If we can retrofit these smaller sites with efficient, rack-mounted cooling, we can unlock a massive amount of compute power without waiting years for a new gigascale facility to be permitted and built.
What is your forecast for data center cooling?
I believe we are rapidly approaching a “thermal wall” where the current method of using air to cool silicon will be viewed as a relic of a less efficient era. In the next few years, the focus will shift entirely away from facility-level cooling toward chip-level thermal management that is integrated directly into the compute stack. We will eventually realize that our current silicon-based computing is inherently unsustainable because of how much energy it wastes as heat compared to biological systems like the human brain. My forecast is that we will see a mandatory transition to “waterless” and fanless architectures, not just for the efficiency gains, but because the power densities required for the next generation of AI will make traditional air cooling physically impossible to implement. Ultimately, the winners in this space will be those who treat heat not as an annoying byproduct to be exhausted, but as a primary constraint that must be engineered out of the system from the very beginning.
