Brazos liquid cooling system for air-cooled data centers | Google Cloud Blog<br>Contact sales Get started for free

Systems

Introducing Brazos: Bringing liquid cooling to air-cooled data centers

June 16, 2026

Jorge Padilla<br>Senior Staff Product Design Engineer, Google

Madhusudan Iyengar<br>Distinguished Engineer, Google

Try Gemini Enterprise Business Edition today<br>The front door to AI in the workplace<br>Try now

Next-generation artificial intelligence (AI) and high-performance computing (HPC) chips routinely exceed 1000 W Thermal Design Power (TDP). Simply put, standard air cooling cannot manage these extreme heat loads. The alternative — retrofitting entire data center facilities with chilled water loops — requires extensive amounts of capital and time. To solve this problem, Google developed Brazos, a rack-mounted, closed-loop liquid-to-air cooling system that lets you deploy high-density, liquid-cooled equipment inside existing air-cooled environments. Brazos is generally available, and our manufacturing suppliers are ready to engage the broader industry to market and produce the Google Brazos design.

Data center facility updates can take months. Brazos breaks with this by allowing simple, one-rack-at-a-time installations. By separating the internal-to-IT liquid loop from the facility water supply, Brazos delivers high-performance liquid cooling with the operational simplicity of standard air systems.

Figure 1: Brazos OCP ORV3 Sidecar Configuration showing three units providing cooling to an adjacent IT rack.

Brazos functions as a self-contained liquid ecosystem, capturing heat via liquid at the component level and rejecting it into the data center's hot aisle using high-efficiency liquid-to-air heat exchangers. This plug-and-play architecture can be rapidly installed in any legacy facility that has sufficient power and standard air handling.

System design and technical specifications<br>Brazos is a modular system that includes three cooling units and integrated rack manifolds, all engineered for high reliability. Each modular chassis occupies 11 Open Units (OU) of rack height and interfaces with standard Open Compute Project (OCP) ORv3 form-factor racks. Key design and performance parameters include:<br>Rack thermal capacity : Supports a 60 kW nominal thermal load per rack across three modular units<br>Coolant compatibility : Runs using either deionized (DI) water or a 25% propylene glycol mixture (PG25)<br>Power delivery : Operates on a 40–60 V DC input designed to connect directly with standard rack busbars<br>Safety features : Certified to UL/CSA/IEC 62368-1 standards and features built-in leak detection alongside pressure relief valves<br>Control plane : Local monitoring uses a built-in human-machine interface (HMI), while remote management connects via Modbus over TCP<br>The mechanical design prioritizes field serviceability. The chassis sits on low-friction slides so it can easily be extended for rapid component access. Crucial components like pumps and fans are designed as hot-swappable, field-replaceable units (FRUs) to minimize mean time to repair (MTTR).

Rapid deployment and industry adoption

In the coming months, we will formally open-source the technical specifications, design principles, and visual assets of Brazos through industry forums. As part of a broader infrastructure portfolio that continues to leverage waterless air-cooled systems alongside liquid cooling, Brazos represents one of many innovations we are contributing to the open hardware ecosystem. We invite system architects, manufacturers, and thermal engineers to evaluate these designs to scale rack-mounted cooling infrastructure for the high-power computing demands of the future.

Next steps

To optimize your legacy data center infrastructure for liquid cooling, follow our upcoming open-source design submissions through the Open Compute Project forum.

Posted inSystems<br>Infrastructure<br>Sustainability

Related articles

Networking<br>How we evolved Google’s global and data center networks for the AI era<br>By Bikash Koley • 7-minute read

Systems<br>How Google Does It: Fleet-wide, large-scale A/B experimentation<br>By Nilay Vaish • 8-minute read

Networking<br>Introducing Virgo Network, Google’s scale-out AI data center fabric<br>By Benny Siman-Tov • 5-minute read

Systems<br>AI infrastructure efficiency: Ironwood TPUs deliver 3.7x carbon efficiency gains<br>By Keguo (Tim) Huang • 6-minute read