The excitement surrounding liquid cooling continues to grow with the burgeoning demand for AI infrastructure across the world. With rack densities increasing and AI factories being planned across all major markets, we caught up with Matt McGuff, Senior Director of Technology Strategy and System Architecture, Intelligent Infrastructure division, Jabil, to take a deep dive into the future of liquid cooling.
Q-1) How have you seen liquid cooling technology evolve? Will it gradually phase out traditional cooling methods?
Liquid cooling has moved from a niche high-performance computing technology to a foundational requirement for AI infrastructure. Over the past several generations, xPU (GPU, CPU and custom accelerator) power has climbed from a few hundred watts to more than 1 kW, with next-gen designs enabling 5kW. Rack densities have risen from 10–20 kW to well over 100 kW in today’s AI racks, with roadmaps projecting several hundred kilowatts and potentially 1 MW.
At those densities, cooling can no longer be treated as a standalone subsystem. It must be designed alongside compute, networking, power, rack mechanics and the data center environment.
That said, we don’t expect liquid cooling to replace air cooling across the board. Even liquid-cooled racks reject some heat to air, and general-purpose compute will remain largely air-cooled for years. The reality is hybrid, with liquid becoming the default for AI compute.
Q-2) Given increasing rack densities, how is liquid cooling changing rack design, validation, deployment readiness, and operations?
“Liquid cooling is not just changing how we cool a server. It is changing how we design, manufacture, test and deploy the entire rack.”
Higher-density racks turn cooling from a component challenge into a system-level one. Effective design starts with the data center: how facility water is delivered, available flow rates, connection requirements, coolant type, and the location and specification of the CDU. At the rack level, the design must account for flow balancing, pressure drop, leak detection and how the system responds if a leak occurs.
Engineering labs and manufacturing lines must adapt too. They need facility water and equipment to fill, purge and service coolant loops. Testing covers pressure drop, leak detection, flow balancing across manifolds and coolant cleanliness, since microchannel cold plates are sensitive to particulates, followed by full-power thermal testing under real or simulated workloads. Manufacturers also need to plan how racks are shipped, filled or dry, and protected from shock and vibration in transit.
With liquid cooling, the risk moves upstream. Issues that might once have been fixed on site must be eliminated before the rack ships, and the facility’s cooling loop must be clean, commissioned and ready to receive it.
In operation, serviceability and visibility must be designed from the start. That means telemetry and leak detection integrated with facility monitoring, disciplined coolant management, and components that can be serviced without disrupting neighboring systems.
Q-3) In light of the growing pressure on cooling infrastructure, how are hyperscalers adopting integrated rack-level solutions?
Rack-scale procurement is not new for hyperscalers. What has changed is how interdependent the components inside the rack have become, and how quickly new capacity must come online.
Many hyperscalers own their rack and cooling designs, often building on open standards such as OCP. What they need from manufacturing partners is the ability to take those designs into volume production quickly, rather than a fixed off-the-shelf product. Custom silicon adds to this: each new accelerator brings its own rack configuration, with its own thermal and power requirements.
Hyperscalers also ramp capacity across several regions at once, so the same validated configuration must be built identically at every site. And while new liquid-ready facilities are under construction, many are using sidecar and liquid-to-air approaches to bring liquid cooling into existing halls.
Q-4) As rack-level integration becomes critical to AI infrastructure deployment, what are some of the challenges operators face in making this happen?
The biggest challenge is managing system complexity at speed and scale, while the technology keeps moving.
AI platforms now evolve roughly annually, so operators are building facilities today for racks two or three generations away, without knowing exactly what power and cooling those racks will require. Many are also retrofitting air-cooled halls that were never designed for heavy, liquid-cooled racks. Meanwhile, every week of delay leaves expensive compute idle, so time to capacity is critical.
These issues are interconnected. A change in cooling architecture affects power, mechanics and serviceability, and with standards still maturing, operators often must coordinate many suppliers and work out who is accountable when something goes wrong. Those expanding across several markets also need racks that perform and are serviced consistently everywhere. In Asia-Pacific, grid capacity, permitting and a shortage of experienced liquid-cooling technicians add further pressure.
That’s why integration matters. Pre-validated rack configurations, consistent manufacturing and traceability across regions, and a single accountable partner across cooling, power and mechanics help operators reduce risk and deploy faster.
Q-5) How important is it to validate power, thermal management, networking, and optics as a complete system?
It is critical. AI infrastructure is increasingly a system of systems, and components that pass individual testing don’t necessarily perform as expected once everything runs together.
The interactions are where problems hide. Power spikes from AI workloads stress both power delivery and cooling response at the same time. Changes in cooling design can push heat-sensitive optics out of their operating range. And a single xPU throttling because of uneven coolant flow can slow an entire training job across a cluster. Found on site, these issues mean delays, idle compute and costly rework.
Validation therefore needs to progress from component and subsystem testing to rack-level and, where applicable, cluster-level validation, so integration problems are found in a controlled engineering and manufacturing environment rather than on the customer’s data center floor. As co-packaged optics and liquid-cooled networking become more common, these interdependencies will only increase.
Jabil’s rack integration capability brings cooling, power and networking together, so complete rack configurations are validated before deployment and then built consistently at volume.
Q-6) How can infrastructure providers reduce deployment risks and accelerate time-to-capacity?
The biggest opportunity is to remove iterations, handoffs and surprises between design and deployment.
That starts with engaging early, co-designing with chip vendors and operators so systems are built for manufacturability and serviceability from day one rather than redesigned later. It continues with standardization: reference designs and rack platforms that carry across chip generations mean each deployment isn’t a fresh engineering project. And it means completing integration in the factory, so racks arrive as tested systems rather than parts to assemble on site.
The facility side matters just as much. Prefabricated power and cooling infrastructure can be built and tested in parallel with site construction rather than after it. Securing long-lead components early, and building validated configurations identically everywhere, removes common causes of delay.
Q-7) What trends and innovations do you foresee with respect to liquid cooling adoption in 2027?
2027 looks like an inflection point. Five trends stand out:
- Rack power will step up sharply. Announced industry roadmaps point to racks of several hundred kilowatts and 800 VDC power distribution in 2027. At those levels, power and cooling can no longer be designed separately.
- Liquid cooling will extend across the rack. Beyond xPUs, power components, busbars, memory and networking will increasingly be liquid-cooled, and co-packaged optics will add to that.
- Warmer coolant and new cold plate approaches will gain ground. Designs running at coolant temperatures of 45°C or higher reduce or remove the need for chillers, a significant advantage in hot, humid climates across Asia-Pacific. For the highest-power chips, direct-to-die and two-phase cooling will move closer to mainstream adoption.
- Adoption will broaden beyond hyperscalers. Neoclouds, sovereign AI programs, colocation providers and enterprises will adopt liquid cooling at scale, many through sidecar and retrofit approaches that bring liquid into existing facilities. Greater standardization of interfaces will make that easier.
- Manufacturing and diagnostics will get smarter. Digital traceability, AI-assisted diagnostics and real-time process data will support the manufacturing of complex rack-scale systems, and will increasingly follow the rack into operation, helping operators diagnose and resolve issues faster.
* Jabil is a global provider of engineering, supply chain, and manufacturing solutions. Founded in 1966 in Detroit, Michigan, Jabil established its cloud and data center business in 2014.