Nvidia launches Rubin servers with 100% liquid cooling

Nvidia unveiled the Rubin server family this week, successors to Blackwell, with a radical design change: 100% liquid cooling. No fans. The announcement was made during Computex 2026 in Taipei, where Jensen Huang confirmed the first Rubin Ultra systems will reach market in Q4 2026.

Why air no longer works at high density

A Blackwell server reached 35-40 kW per rack. Rubin, per preliminary Nvidia data and SemiAnalysis estimates, should operate at 60-80 kW per rack. The physics is unforgiving: water conducts heat 25 times more efficiently than air. Liquid cooling routes heat directly from chips to external chillers. Data center PUE can reach 1.02-1.05 with liquid cooling, versus 1.4-1.6 typical for air.

What changed under the hood

• **GPU Rubin Ultra R100**: TSMC N3 process, 80 billion transistors, estimated 2.5-3x H100 training performance in FP8
• **NVLink 5.0**: 3.6 TB/s bandwidth between GPUs, versus 900 GB/s on Blackwell
• **HBM4**: first production GPU to use HBM4 with 8 TB/s memory bandwidth, double H200

The real cost of the transition

Adopting Rubin servers is not just a hardware swap — it is an infrastructure retrofit. Per Uptime Institute, the average retrofit cost is $15-25 million per 10,000 m² data hall. This creates a real barrier for regional cloud providers, concentrating access to Rubin chips in AWS, Azure, GCP and niche players like CoreWeave.

Sources

• Nvidia — Computex 2026 presentation (investor.nvidia.com)
• SemiAnalysis — *Rubin Architecture Deep Dive: What We Know* (Jun. 2026)
• Uptime Institute — *Liquid Cooling Retrofit Cost Guide* (2025)
• Gartner — *Hype Cycle for Data Center Infrastructure, 2025*