Google’s Brazos Sidecar Brings Liquid Cooling to One AI Rack at a Time

Brazos Is Built for Air-Cooled Data Centers

Google has introduced Brazos, an open-source rack-mounted liquid-to-air cooling system for existing air-cooled data centers.

It targets a real data-center headache: AI racks are getting hotter faster than many facilities can rebuild their mechanical rooms.

Brazos is a closed-loop sidecar.

It captures heat through liquid at the component level, then rejects that heat into the hot aisle through liquid-to-air heat exchangers.

The practical appeal is straightforward: operators can add high-density liquid cooling one rack at a time instead of converting an entire hall in a single project.

Google said Brazos is already generally available, and its manufacturing suppliers are ready to work with the broader industry to produce the design.

The company also plans to open-source the technical specifications, design principles, and visual assets.

The open-source approach is not just about sharing.

Wider adoption could help standardize the retrofit layer around AI cooling before every operator builds a slightly different box.

The Numbers Are Rack-Level, Not Facility-Level

The system uses three cooling units and integrated rack manifolds.

Each chassis takes 11 Open Units of rack height and is built for standard ORv3 racks from the Open Compute Project.

At rack level, Brazos is rated for a 60kW nominal thermal load across its three modules.

Operators can run it with deionized water or a 25 percent propylene glycol mixture, while power comes through a 40-60V DC input connected to standard rack busbars.

The safety and operations details matter more than the usual launch language.

Brazos is certified to UL/CSA/IEC 62368-1 standards, includes leak detection and pressure relief valves, and has a built-in human-machine interface with remote management through Modbus over TCP.

Those are the details facility teams will care about before they let liquid cooling sit next to production AI hardware.

Sidecars Are Becoming AI Infrastructure Tools

Google is not introducing the sidecar in isolation.

Brazos extends work around Open Compute Project rack adaptations, where power and cooling are being separated from the compute rack to handle AI power densities.

In 2024, Microsoft and Meta announced Mount Diablo, an open rack design for AI data centers that separates power and compute into different cabinets.

Google has also worked on Mount Diablo with Meta and Microsoft, including an AC-to-DC sidecar power rack that leaves the main rack space for GPUs, TPUs, and CPUs.

AWS has explored a similar path with its In Row Heat Exchanger, a custom liquid cooling design that can be installed without changing air-cooled mechanical designs.

The industry signal is hard to miss: AI clusters are forcing operators to upgrade old rooms in smaller, modular steps because full rebuilds are expensive, slow, and disruptive.

The Hard Part Is Installation at Scale

Brazos gives data center operators a defined retrofit option: open specifications, OCP rack compatibility, 60kW rack-level cooling, and a supplier path for production.

That is useful for facilities that cannot wait for entirely new AI halls.

The real test is less elegant than the diagram.

Operators will need to show that Brazos can be installed with limited disruption, maintained safely, integrated with existing airflow, and produced at a cost that makes one-rack-at-a-time liquid cooling worth the trouble.

If those economics work, Brazos becomes more than a Google design note.

It becomes a practical bridge between yesterday’s air-cooled rooms and the AI racks already arriving at the door.