Google Open-Sources Brazos: Plug-and-Play Liquid Cooling for Air-Cooled DCs
Summary
Key Takeaways
Google introduces Brazos, a rack-mounted closed-loop liquid-to-air cooling system designed for existing air-cooled data centers, enabling high-density liquid cooling without facility retrofits. The modular system uses three cooling units occupying 11 OU height, compatible with OCP ORv3 racks, supporting 60 kW per rack. Coolant options include deionized water or 25% propylene glycol (PG25). Power input is 40-60V DC from rack busbars. Remote management uses Modbus over TCP. Safety features include UL/CSA/IEC 62368-1 certification, leak detection, and pressure relief valves. Key components are hot-swappable FRUs to minimize MTTR. Google will open-source the design via the Open Compute Project, inviting industry adoption.
Why It Matters
Google's open-source move disguises a lock-in strategy: by defining the design spec (pump type, heat exchanger geometry), it forces manufacturers to follow Google's standard, undermining competitors like CoolIT and Asetek. The Modbus over TCP control plane lacks fine-grained telemetry (per-node temp, flow rate) compared to Redfish or IPMI, limiting cooling optimization. The 60kW capacity per rack is insufficient for future 2000W+ TDP chips (e.g., Nvidia Rubin), requiring full unit replacement—a hidden asset depreciation trap. This defends against Nvidia's proprietary liquid-cooled racks, aiming to wrest control of the cooling layer. Physical limits (pump MTBF, fan noise) are downplayed, demanding extra redundancy in practice.
PRO Decision
【Vendors】 Nvidia should highlight its liquid-cooled racks supporting >100kW/rack with deep GPU integration and fine-grained telemetry via NVLink/DCGM. CoolIT and Asetek can attack Brazos's obsolete Modbus over TCP and offer OCP-compatible systems with Redfish and OpenBMC for real-time thermal management APIs.
【Enterprises】 CIOs and architects must demand detailed performance specs (pump MTBF, fan noise, real-world thermal capacity) and run independent benchmarks against Nvidia, Vertiv. Assess integration with existing DCIM tools to avoid operational blind spots. Adopt multi-vendor strategy to prevent lock-in to Google's standard.
【Investors】 See through Google's move: it lowers its own AI infrastructure cooling cost (boosting margins) while commoditizing competitors via an open standard. But the 60kW ceiling limits long-term applicability—watch for future high-TDP chip cooling. Beware of Google monetizing certification fees or supply chain control through OCP.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)