Google unveils Brazos: liquid cooling for AI without rebuilding the data centre

On 17 June 2026, Google introduced Brazos, a rack-mounted, closed-loop liquid-to-air cooling system that lets operators run today's hottest AI chips inside data centres that were only ever built for air cooling — without ripping out the building's plumbing.

Why it matters (the plain version): the chips powering modern AI run blisteringly hot. Next-generation AI and high-performance computing processors now routinely exceed 1,000 watts of thermal design power each — far more than a fan and cold air can carry away. The usual fix is to retrofit an entire facility with chilled-water loops, which costs a fortune and takes months. Brazos sidesteps that: it bolts liquid cooling onto a single rack at a time, so an existing "air-cooled" hall can start hosting dense AI hardware almost immediately.

How it works (the technical version): Brazos is a self-contained liquid ecosystem. It captures heat with liquid right at the chip, then rejects that heat into the data centre's hot aisle through high-efficiency liquid-to-air heat exchangers — crucially keeping its internal coolant loop separate from the facility's water supply. That separation is what gives it "plug-and-play" simplicity: no re-plumbing the building.

The headline specs:

• 60 kW nominal thermal load per rack, delivered across three modular cooling units (each chassis occupies 11 OpenU and fits standard Open Compute Project ORv3 racks)

• Runs on either deionised (DI) water or a 25% propylene-glycol mix (PG25)

• 40–60V DC input wired straight to standard rack busbars

• Built-in leak detection and pressure-relief valves; certified to UL/CSA/IEC 62368-1

• Local monitoring via an on-board human-machine interface, remote management over Modbus/TCP

• Pumps and fans are hot-swappable field-replaceable units on low-friction slides, to minimise repair time

The strategic angle: Brazos is generally available now, and Google says its manufacturing suppliers are ready to produce the design for the wider industry. In the coming months Google will open-source the technical specifications, design principles and visual assets through industry forums such as the Open Compute Project — positioning the company not just as a buyer of compute, but as an infrastructure innovator shaping how the whole sector cools the AI build-out.

Read more: Google Cloud blog