Rubin全液冷架构转型与散热分析
Rubin全液冷架构转型与散热分析
Conclusion: Rubin’s move to all-liquid cooling is driven by rising chip and rack heat loads: air cooling is becoming a constraint as AI systems pack more compute into each rack. Capturing heat directly at the components—and using warmer coolant—can make dense systems more practical within a data center’s power and cooling limits. That supports NVIDIA’s rack-scale platform strategy and creates opportunities for liquid-cooling suppliers, but it does not guarantee higher profits for every company labelled a “cooling play”: realized demand will depend on Rubin deployments, supplier awards, execution and facility design.
Data as of October 30, 2026. The market outlook below includes a TrendForce forecast published August 17; supplier forecasts are not treated as confirmed shipment results.
What is driving the change
- More heat in less space. TrendForce says individual high-end AI chips have moved above 1 kW of thermal design power, while rack-scale systems consume hundreds of kilowatts. Rubin’s cooling requirement also extends beyond GPUs and CPUs: TrendForce describes liquid cooling reaching networking, power and other rack components.
- Air cooling is no longer enough for the whole rack. NVIDIA describes Rubin’s design as liquid-cooled and fanless, including components that earlier liquid-cooled servers left air-cooled. This is a platform-level redesign, not simply adding cold plates to the GPUs. NVIDIA’s Rubin cooling explanation
- Warmer coolant can reduce facility cooling overhead. NVIDIA says its MGX racks are designed for 45°C inlet coolant. The company’s rationale is that warmer water can allow ambient-air heat rejection and closed-loop dry coolers in many climates, reducing reliance on chillers and freeing more facility power for computing. NVIDIA says this can enable up to 10% more Vera Rubin NVL72 racks in the same power budget. These are vendor claims; results will vary with climate and facility design.
- Cooling is part of a broader power-and-density problem. The design aims to remove heat while enabling higher rack density and more compute from constrained power capacity. It does not eliminate the need to reject heat at the facility: CDUs, manifolds, piping and heat-rejection equipment remain important, and some sites may still need chillers.
The broader shift is not unique to NVIDIA. TrendForce projected liquid-cooling penetration among AI chips at 53% in 2026, rising from about 33% in 2025. That is an industry forecast, not a measure of actual Rubin shipments.
What it could mean for NVIDIA
- A stronger rack-scale proposition. If operators can fit more compute into a fixed power envelope, Rubin’s cooling design could improve the practical economics of deploying the platform—not just its chip-level performance.
- More system integration, and more execution responsibility. Cooling, power delivery and serviceability are tightly coupled in a dense rack. That makes coordinated rack design valuable, but also raises the importance of reliable components, installation and maintenance.
- No automatic revenue or margin uplift. The cooling transition supports the case for higher-density AI systems, but public evidence cited here does not establish how much incremental NVIDIA revenue, profit or customer savings it will produce. Those depend on system shipments, pricing, customer deployment pace and the cost of the full infrastructure.
Supplier implications—and the limits of the thesis
TrendForce identifies cold plates, manifolds and coolant distribution units (CDUs) as key elements of liquid-cooling systems. It names Cooler Master, AVC, BOYD and Auras among major cold-plate suppliers, and Jentech as the Rubin heat-spreader supplier. These are industry-research attributions, not a complete, NVIDIA-confirmed list of contract awards or market shares. TrendForce said AVC was expected to begin shipping Rubin cold-plate modules in the third quarter; I could not verify whether that forecast had been realized by October 30.
The potential opportunity spans more than GPU cold plates: rack-level coolant distribution, connectors, pumps, manifolds and heat rejection may also matter as more components move onto liquid cooling. But supplier exposure is uneven. A company’s participation in the broad cooling market does not by itself prove it has a Rubin design win, meaningful volumes or attractive margins. And while warmer coolant can reduce chiller use in suitable sites, that may temper demand for some conventional cooling equipment even as direct liquid-cooling infrastructure grows.
Investment read-through: The durable thesis is that AI rack thermal management is becoming a core infrastructure requirement, with value potentially shifting toward suppliers qualified for dense, integrated liquid systems. The key checks are evidence of actual Rubin production and deliveries, named supplier qualifications or orders, and whether reported sales and margins confirm that increased cooling content is translating into financial results. Until then, treat supplier names and penetration estimates as indicators of potential exposure—not proof of realized earnings.