液冷 Primer：冷却液分配单元（CDU）

Liquid Cooling Primer: Coolant Distribution Units (CDUs) As rack thermal densities in data centers continue to increase, air is no longer powerfulenough to handle cooling requirements. And so liquid cooling, once a fridge modalityrelegated to only the highest-performance computing use cases, increasingly finds itselfin the mainstream. At the core of liquid cooling, two components drive the narrative (CDUsand Cold Plates). In this note, we offer perspectives on CDUs, why they are critical, how toevaluate them, key players in the market and their offerings, and a longer term view on whatthe future for the equipment could look like when the market eventually cools down (punintended). Varun Govindaraj+1 917 344 8543varun.govindaraj@bernsteinsg.com Chad Dillard+1 917 344 8469chad.dillard@bernsteinsg.com Alasdair Leslie+44 20 7762 4952alasdair.leslie@bernsteinsg.com We think CDUs are a great business to be in; the equipment is complex enough to not beentirely commoditized and has a material service attach. They are mission-critical; if a CDUfails you are looking at multiple racks burning out (which is a huge issue today when rackvalues continue to increase). This also creates a technical moat where customers will wantreliable service; they are unlikely to go to the lowest cost provider in the market because thecost of failure far outweighs the near-term savings a cheaper service contract can deliver. Madison Rezaei+1 917 344 8622madison.rezaei@bernsteinsg.com Specialist Sales Steve Song+1 917 344 8401steve.song@bernsteinsg.com While we have not opined on the market size of CDUs, there is clearly debate on both sizeand growth rates. We are comfortable with an LSD $B market size for 2026, growingdouble digits (mid-teens+) over the next 5 years to get to MSD-HSD $B by 2030. We havedeveloped a proprietary liquid cooling model for this to be modeled; but market size ishighly sensitive to GW added, cost per kW of cooling and CDU useful life (all of which areseeing debates). Reach out to the authors or your Bernstein sales contact if you’d like awalkthrough of how to use it. Looking at the actual CDUs launched in the market today, there are many players, butnot everyone is innovating or has a large scale unit. Players in the NVIDIA liquid coolingecosystem (Vertiv, nVent, Boyd, Motivair) all have great products. Trane Technologiespunches above its weight. Carrier and JCI have CDUs but given they are more focused onchillers, we found their CDU breadth, specifications, and level of detail provided to not be atthe same level as the other names in this list. CoolIt seems to be more of a cold plate name;their approach temperature lags competitors. As we think about the next five years, we believe that technology roadmap visibility (whichcomes from being a partner of NVIDIA since they really set the direction of change) andparticipation in the Open Compute Project (OCP) create a right to win (because it deepenshyperscaler relationships and creates a pathway for long-term demand generation); and notmany companies can say they have both. Lastly, we think the shift to two-phase DTC cooling (from single-phase) should be closelywatched. Only Vertiv and Accelsius (from the companies we have mentioned in this note)have actually announced products / published detailed perspectives. While we’re still atleast a year out from commercially scaled offerings (if not more), and most other companiesserious about liquid cooling are likely working on a product, the step-change in engineeringfrom this shift has the potential to disrupt the market and position occupied by key players. BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS We rateVRTOutperform with a target price of $416.We rateNVTOutperform with a target price of $218.We rateCARRMarket-Perform with a target price of $75.We rateTTOutperform with a target price of $550.We rateJCIOutperform with a target price of $176.We rateSchneiderOutperform with a target price of €310.We rateEatonOutperform with a target price of $534. INTRODUCTION In the past, a continuous flow of cool air over servers was sufficient to cool data centers. With time, chips and racks have grownincreasingly dense, to the point where even freezing cold air blowing at gale force speeds wouldn’t do the job. Needless to say,finding an alternative cooling modality became critical. And so, liquid cooling technology, which long sat at the fringe of datacenter infrastructure, suddenly found itself thrust into the spotlight. Looking ahead, rack densities and chip TDPs seem set tokeep growing, forcing liquid cooling solutions to evolve with more rigorous server demands. OVERVIEW OF LIQUID COOLING Liquid cooling encompasses multiple different types of technology, with varying levels of efficiency and at distinct stages ofmaturity. At their core, they share one common principle; liquid flows over a surface (directly or indirectly) and uses principles ofconvection to extract heat. We describe five modalities of liqu