液冷：谷歌Brazos对更广泛的CDU生态系统意味着什么

Liquid Cooling: What does Google's Brazos mean for the broaderCDU ecosystem? Google recently released early specifications for a liquid-to-air (L2A) CDU named “Brazos”.Like project “Deschutes” that came before it in 2025, Brazos simply announces a CDUreference design for the Open Compute Project’s vendor ecosystem (which counts Vertiv,nVent, and Boyd as members) to produce.It is not a competing product to Vertiv/Boyd/ Varun Govindaraj+1 917 344 8543varun.govindaraj@bernsteinsg.com Chad Dillard+1 917 344 8469chad.dillard@bernsteinsg.com Prima facie, the specifications on Brazos-class CDUs seem relatively low. At 60kW ofcooling capacity, it cannot even cool one Blackwell rack (which requires ~120kW of coolingcapacity). Interestingly, Brazos-class CDUs are designed to draw DC power (vs. many other Alasdair Leslie+44 20 7762 4952alasdair.leslie@bernsteinsg.com Madison Rezaei+1 917 344 8622madison.rezaei@bernsteinsg.com We believe these CDUs are designed to be OCP compliant, and fit into legacy hyperscalerdeployments which tend to have DC power infrastructure already. The cooling capacityof 60kW does not work well for frontier training models, but can support inference at thatrange. Given concerns around how quickly data center capacity can get up and running,creating an inference ecosystem that can run in existing facilities with simple L2A CDUretrofits seems like a good move (especially when inference looks set to grow faster than Steve Song+1 917 344 8401steve.song@bernsteinsg.com While hard to quantify, we see two mid-term risks for CDU manufacturers.First,there is real commoditization risk of the inference CDU ecosystem. Brazos specs are meaningfully easier to deliver than Deschutes. While there will be a service attach, marginswill be lower vs. flagship training CDUs. The deciding factor here is inference rack densities;we are unsure what this looks like for purpose-built inference silicon from hyperscalers. If itremains around 60 kW/rack, then a meaningful share could be cooled by L2A Brazos-style Second, we may also see some hyperscaler demand shift from greenfield to retrofits,especially if projects get delayed. While we do not see that in stranded capacity numbers inour data center capacity tracker, there has been some anecdotal input from companies thatthey are seeing a few customers (both hyperscalers and colos) push out deliveries becauseof a lack of readiness / project delays. We’re not saying thisWILLhappen, but we think We do not view this announcement by Google as imminently alarming for CDUmanufacturers, especially those that focus on technical excellence at the top of the stack. Some level of CDU commoditization was always expected by the market. While we thinkthis announcement accelerates the commoditization trend for lower spec. products, we still BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS We rateVertivOutperform with a target price of $416. We rateEatonOutperform with a target price of $534. DETAILS Context Hyperscalers have historically pushed the “Open Compute Project” or OCP, a non-profit that published open-sourcespecifications for data center hardware. Founded by Meta, it now has participation from Google, Microsoft, and Oracle (althoughAmazon is a notable missing name) and a host of other vendors. In 2025, Google published details of Project Deschutes as apart of the OCP; this was a set of technical specifications and standards for compliant Liquid to Liquid (L2L) CDUs (with 2MW of cooling capacity at a 30C approach temperature). Vertiv, nVent, and Boyd all now have Deschute-compliant units available intheir product catalogs. Yesterday, Google announced that they would specify another CDU design - this time for Liquid to Air(L2A CDUs) named project Brazos (in line with Google naming these projects after rivers). This note walks through implications L2L vs. L2A: What’s the difference? In our CDU primer earlier this week, we focused mostly comparing on flagship L2L CDUs. L2L units are named as such becausethey reject heat between two liquid loops (the technology cooling system and facility water system). In contrast, the L2Aconfiguration we are talking about today only has one liquid loop (the technology cooling system) which rejects heat as air intothe “hot aisle” of a data center. A CRAH or CRAC (Computer Room Air Handler or Computer Room Air Conditioner) then blowsthe hot air out of the hall / building. Generally, L2L cooling is preferred for all greenfield builds or situations where rack densitiescross 150kW (it is much more energy efficient). In contrast, L2A builds are preferred when retrofitting an existing data center Thoughts on Brazos specifications When we look at the specifications offered by Brazos, prima facie it does not look that impressive or demanding. 60kW ofcapacity cannot cool even a single Blackwell rack (which requires ~120kW of cooling power) - players like nVent, Motivair, andCoolIT go far higher on their cooling power for L2A CDUs.