- Demand for liquid cooling is rising even faster than anticipated thanks to AI
- But supply of critical CDUs for cooling systems is lagging
- Rising capacity demands could change what kinds of cooling solutions are in demand in the coming years
The first AI-driven data centers are coming online by mid-year, and demand for high-performance liquid cooling is skyrocketing. But cooling vendors are struggling to keep pace as hyperscalers and enterprises race to upgrade their data center cooling infrastructure.
LiquidStack CEO Joe Capes told Fierce in an exclusive interview that demand for coolant distribution units (CDUs) is currently 4-5X greater than supply, leading to year-long equipment lead times from some competitors.
To meet this challenge and get a leg up on the competition, LiquidStack recently doubled its U.S. manufacturing capacity, joining competitors like CoolIT and Vertiv in expanding operations.
The present dynamic isn't likely to change anytime soon given there are plenty of new data center facilities on the way – at least 505, according to Synergy Research Group. That means more demand for liquid cooling.
CDU supply crunch: Why it’s happening
Dell’Oro Group estimated data center capex will climb to $1 trillion by 2029, with half of that spending going towards accelerated servers. In terms of practical implications, Dell’Oro Group VP Baron Fung noted in a recent blog that Meta, Microsoft and Nvidia have all either emphasized the need for or rolled out products that require liquid cooling.
At last check, Dell’Oro said it expected liquid cooling to generate $15 billion in cumulative revenue by 2028. But in February, the firm raised its forecast for overall data center physical infrastructure, citing higher than expected demand for liquid cooling and rack power distribution components.
The problem, Capes said, is that the supply of CDUs is already lagging demand and dragging out lead times for equipment. “One of our biggest competitors we’ve heard is announcing one-year lead times on CDUs,” he noted, without naming names.
✔ $1 Trillion in Data Center CapEx by 2029 – Half of this spending will be focused on accelerated servers, many of which require liquid cooling. (Source: Dell’Oro Group)
✔ 505+ New Data Centers In the Pipeline – AI and cloud expansion is fueling growth. (Source: Synergy Research Group)
✔ Nvidia’s Shift to Higher-Power Racks – The GB200 system requires 120kW per rack, but upcoming 600kW racks demand even more advanced cooling.
LiquidStack is one of several key vendors in the nascent liquid cooling market, competing with the likes of Vertiv, Schneider Electric, CoolIT and Asperitas. Though it also provides immersion cooling systems, LiquidStack’s CDU for direct-to-chip systems is one of its most popular products.
Capes said the short-term goal of its recent expansion is to be able to produce enough CDUs at its two plants to cover 3GW of data center capacity. It is also eyeing an international expansion to meet rising demand in the EMEA and Asia Pacific regions.
The future of liquid cooling: Immersion vs direct-to-chip
While direct-to-chip cooling leads the pack today, Capes said the types of products needed are changing just as fast as demand is.
Late last year, LiquidStack unveiled a CDU capable of serving up to 1MW of capacity. That was more than enough to meet the needs of Nvidia’s top-tier GB200 accelerated system, which requires a whopping 120kW of power per rack. But just last month, Nvidia changed the game by unveiling plans to launch a 600kW rack – aka a rack that draws more than half a MW of power.
“Everything is going to continue to scale in capacity,” Capes said, noting that semiconductor roadmaps seem to call for chip power levels increasing up to 3,000, 4,000 and even 5,000 watts. As thermal design power needs increase, there could be a swing toward immersion cooling.
Sure, you probably could cool a 5,000-watt chip with a direct-to-chip system – as JetCool is aiming to do – but Cape said whether you should is a different question.
“There are always aspects of heroic engineering that can be done to take a particular technology and extend it. But that doesn’t make it efficient or scalable or cost effective,” he argued.
Capes hinted LiquidStack plans to announced a new product to meet rising capacity demands later this year.
Standards vs scale
There’s one other undercurrent complicating this whole situation: a lack of standards.
“The investment community wants to know how your CDU is different[iated],” Cape said. “But when an industry is changing this rapidly and this dynamically, customers are not looking for proprietary, they’re not looking for exotic…they’re looking for something that is predictable, reliable and standard. That’s how you scale any technology.”
Top takeaways:
✔ CDU supply is falling behind demand, setting the stage for delays.
✔ AI workloads and Nvidia’s 600kW racks are reshaping the cooling market.
✔ Vendors like LiquidStack, CoolIT, and Vertiv are scaling up to meet industry needs.
✔ Immersion cooling may become essential as chip power levels reach 5,000W+.
While a handful of entities – like The Green Grid and Open Compute Project – have been attempting to make progress on this front, Cape said the aforementioned pace of change in the space is in conflict with those efforts.
Cape said perhaps the best approach for now is to build bigger – that is, standardize designs that leave plenty of room for growth. That way, CDUs and cooling systems at large will have room to grow with the industry.
Stay ahead of the cooling revolution—subscribe for exclusive insights!