Google Brazos is more like an L2A CDU reference design for the inference retrofit ecosystem than direct competition to existing vendors

Overview

This report analyzes the impact of Google's early Brazos L2A CDU specification on the CDU ecosystem. Bernstein's core view is that Brazos is merely a reference design opened to the OCP vendor ecosystem, intended to enable existing hyperscale data centers to support inference demand through relatively simple L2A retrofits, rather than to create direct product competition for existing CDU vendors.

Core views

The report argues there is no need for excessive concern in the short term: CDUs will remain supply-constrained, and vendors will stay focused on L2L and higher-spec products. But two medium-term risks warrant attention: first, standardization and commoditization of inference CDUs could pressure margins on lower-end products; second, if new data center builds are delayed, hyperscale customers may shift some demand from new builds to retrofits of existing facilities. Over the long term, CDUs are unlikely to be displaced, as complexity, service requirements, and ongoing technological evolution should continue to support some pricing power for leading vendors.

Analysis framework

The report derives Brazos' impact on CDU vendors' revenue quality, margins, and service stickiness by comparing L2A versus L2L architectures, Brazos versus Deschutes specifications, inference versus training rack power density, and hyperscale data center needs for new builds versus retrofits.

Methodology notes

• Technical architecture comparisonComparison of L2L and L2A liquid cooling architectures

  L2L exchanges heat between the technology cooling loop and the facility water system, making it suitable for new builds or scenarios where rack density exceeds roughly 150kW; L2A has only the technology cooling loop and rejects heat into the hot aisle, making it more suitable for retrofitting existing data centers and for cooling demand of about 40kW to 150kW.

  Based on this, the report concludes that Brazos is better suited to inference workloads and retrofit scenarios rather than cutting-edge training clusters.

• Open hardware ecosystemOCP reference design analysis

  Google released CDU technical specifications through OCP, allowing ecosystem vendors such as Vertiv, nVent, and Boyd to manufacture equipment that complies with the specifications.

  This means Brazos is not a competing product manufactured by Google, but rather a standardization requirement for the supplier ecosystem.

• Market stage assessmentNear-, medium-, and long-term outlook for CDUs

  From now through 2028, CDUs will remain a growth driver for companies exposed to liquid cooling; by 2030, standardization risk rises; after 2030, CDUs should remain important due to flow, pressure, and service requirements.

  The report uses a time-based framework to distinguish short-term supply-demand tightness, medium-term commoditization risk, and long-term service and innovation premiums.

• Valuation methodologyMulti-company valuation framework

  Vertiv is valued using an NTM+1 EBITDA multiple, nVent using sum-of-the-parts valuation, Schneider using DCF, and Eaton using a forward P/E multiple discounted back.

  These valuation methods support the report's Outperform ratings and target prices for the relevant electrical equipment and data center infrastructure companies.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Vertiv Holdings Co

  A beneficiary in data center power, thermal management, and CDUs, rated Outperform with a target price of $416.

  Strengths

  Has exposure to high-end liquid cooling and data center infrastructure, and the report believes the innovation premium of flagship CDUs will persist.

  Weaknesses

  If lower-spec inference CDUs become standardized, some products may face lower margins and weaker service moats.

  Comparison

  Relative to lower-spec reference designs like Brazos, Vertiv benefits more from high-end training CDUs and complex system integration demand.

  Risks

  Major improvements in cooling efficiency reducing demand, slower data center expansion, faster customer migration from NVIDIA chips to custom silicon, and technology pathway shifts such as DTC or 800VDC.

• nVent Electric PLC

  A vendor related to the OCP ecosystem and CDU products, rated Outperform with a target price of $218.

  Strengths

  Has opportunities to participate in CDU products and services, and can still generate revenue from installation and maintenance services.

  Weaknesses

  The Brazos specification is easier to deliver than Deschutes and may not require the high engineering premium of vendors such as nVent.

  Comparison

  If inference rack power remains around 60kW, Brazos-style L2A CDUs could replace some higher-spec solutions; if power density rises, Deschutes-class or higher-end CDUs would still be needed.

  Risks

  Accelerating commoditization of CDUs or OCP specifications, slower-than-expected CDU production ramp, and longer hiring cycles for key roles.

• Schneider Electric SA

  A name tied to data center electrical infrastructure and energy management, rated Outperform with a target price of €310.

  Strengths

  Stable cash flow, with the energy management business supported by data center construction and electrification trends.

  Weaknesses

  Relatively sensitive to changes in the demand mix for data center electrical infrastructure.

  Comparison

  Schneider's investment case is more centered on electrical infrastructure and energy management, making it less directly affected by Brazos than pure-play CDU vendors.

  Risks

  A decline in data center demand, technology changes altering the composition of electrical infrastructure, and additional U.S. tariffs weakening end demand.

• Eaton Corp PLC

  A name related to data centers and electrical infrastructure, rated Outperform with a target price of $534.

  Strengths

  Long-term structural factors support earnings resilience, backed by electrification, data centers, and utility capex trends.

  Weaknesses

  Linked to overall load growth, the data center construction cycle, and the pacing of large projects.

  Comparison

  Compared with CDU vendors, Eaton is more of an electrical infrastructure beta play, with limited direct technological impact from Brazos.

  Risks

  Load growth below expectations, disruptions to global data center construction, slower utility capex growth, slower reshoring trends, or cancellation of large projects.

• Brazos-style L2A CDU

  A lower-spec L2A CDU reference design for the OCP ecosystem.

  Strengths

  Suitable for simple retrofits of existing data centers, able to support roughly 60kW inference workloads, and able to leverage existing DC infrastructure.

  Weaknesses

  Cooling capacity is insufficient for high-power training racks such as Blackwell, and its technical requirements are lower than those of flagship L2L CDUs.

  Comparison

  Compared with Deschutes-class L2L CDUs, Brazos is better suited to inference and retrofits, with lower specifications and potentially lower margins.

  Risks

  If inference rack power density rises, the applicable range of 60kW L2A solutions will become constrained.

• Deschutes-class L2L CDU

  A higher-spec CDU for new high-power-density training clusters.

  Strengths

  Higher cooling capacity and engineering complexity, potentially offering higher margins and stronger service stickiness.

  Weaknesses

  Deployment typically requires a facility water system, making it better suited to new builds rather than simple retrofits.

  Comparison

  Compared with Brazos, Deschutes is higher-end, better suited to cutting-edge training workloads, and better able to preserve vendor differentiation.

  Risks

  If hyperscale customers delay new-build projects and shift toward retrofits, near-term order mix could be affected.

Key data

• Brazos cooling capacity60kWBelow the roughly 120kW cooling requirement of a single Blackwell rack, and therefore unsuitable for cutting-edge training racks.
• Brazos power characteristicDC powerDesigned to draw direct current directly from the busbar, matching the DC infrastructure of some existing hyperscale data centers.
• Deschutes specification reference2MWAn L2L CDU specification released by Google through OCP in 2025, with cooling capacity significantly higher than Brazos.
• L2A applicable rangeAbout 40kW to 150kWThe report believes L2A is better suited to retrofits of existing data centers or to moderate power-density cooling.
• L2L applicable scenariosNew-build projects or rack densities above roughly 150kWBecause of its higher energy efficiency, L2L is generally better suited to high-density new-build data centers.
• Vertiv rating and target priceOutperform, $416Valuation uses 32x NTM+1 EBITDA, based on roughly $5.1B EBITDA.
• nVent rating and target priceOutperform, $218Valuation uses a sum-of-the-parts approach, with the EC business at 17x EV/EBITDA and the SP business at 28x EV/EBITDA.
• Schneider rating and target priceOutperform, €310The target price is based on DCF, with a WACC of 7.8% and a terminal growth rate of 2.5%.
• Eaton rating and target priceOutperform, $534The target price is based on 27x 2030 EPS and discounted back to the present, equivalent to 33x 2027 EPS.
• CDU market size assessmentCurrently low single-digit billions of dollars, potentially rising to mid- to high-single-digit billions over the next 5 yearsThe report believes this growth path is not unrealistic, though market forecasts vary widely.

Impact & implications

The direct implication of Brazos for the industry is that lower-spec inference CDUs may standardize more quickly, reducing vendor margins and differentiation at that tier; however, for flagship training CDUs and high-complexity services, innovation capability, reliability, and post-installation service can still support pricing power for leading vendors. From an investment perspective, the report is more inclined to view Brazos as a medium-term margin-structure risk rather than a short-term demand shock.

Risks

• Lower-spec inference CDUs may accelerate toward commoditization due to OCP and Brazos specifications, pressuring vendor margins.
• There is uncertainty around inference rack power density; if it stays near 60kW, Brazos-style L2A CDUs will have broader applicability, but if it rises materially, higher-spec CDUs will be required.
• Insufficient readiness or delays in data center projects may push some demand from new builds to retrofits of existing facilities.
• CDU service revenue still has value, but the service moat of lower-spec products may be weaker than that of flagship training CDUs.
• Technology pathway changes such as higher efficiency, custom silicon, DTC, and 800VDC may alter demand for cooling and electrical infrastructure.
• Company-specific risks also include production ramp, hiring for key roles, tariffs, load growth, utility capex, and cancellation of large projects.

What to watch

• When Google will formally open-source Brazos technical specifications, design principles, and visual assets.
• How OCP ecosystem vendors such as Vertiv, nVent, Boyd, and Motivair will respond with products based on the Brazos specification.
• Whether rack power density for inference-specific chips and hyperscalers' in-house silicon stabilizes around 60kW.
• Changes in the order mix between L2A retrofit demand and L2L new-build demand.
• Whether CDU lead times and backlog persist through 2028.
• Signals of idle capacity, project delays, and deferred customer deliveries in data center capacity tracking.
• Whether CDU standardization weakens pricing power, and whether service quality and uptime continue to support post-installation revenue.