CoreWeave Deploys Liquid-Cooled 102.4 Tb/s Switch for Vera Rubin NVL72

A liquid-cooled switch for Vera Rubin NVL72

CoreWeave argues that networking has become a strategic part of AI cloud infrastructure, since it determines how well GPU compute turns into goodput for training and inference. As liquid cooling lets rack-scale systems such as NVIDIA Vera Rubin NVL72 deliver exaflops of compute, the company says the network has to keep pace, and that switches are now adopting the same cooling technique.

The SN6600-LD is built on the NVIDIA Spectrum-6 ASIC and is fully liquid-cooled. According to CoreWeave, it delivers 102.4 Tb/s of switching capacity across 64 x 1.6 Tb/s ports using 200G SerDes. That doubles the per-lane bandwidth of the previous generation, in a compact 2U form factor. In a standard 48U rack, 16 of the switches can be deployed, giving 1.64 Pb/s of total switching capacity.

CoreWeave says this lets it offer a fully non-blocking, multi-plane, multi-rail spine and leaf fabric connecting Vera Rubin NVL72 GPUs without oversubscription. For large scale training, the company says this means higher model FLOPS utilization (MFU), lower network latency and GPU idle time near zero. For inference, it points to consistently low latency, faster response times and higher throughput.

Three measurable gains over air cooling

CoreWeave compares the liquid-cooled Spectrum-6 based SN6600-LD with the air-cooled Spectrum-4 based SN5600/SN5610 in a topology connecting 16,000 Vera Rubin GPUs. It lists three gains:

• 100% more switching performance per rack. The SN5600/SN5610 offers 51.2 Tb/s across 64 ports of 800Gb/s on 100Gb SerDes. The SN6600-LD offers 102.4 Tb/s across 64 ports of 1.6Tb/s on 200Gb SerDes, in the same form factor.
• 62.5% reduction in rack footprint and floor space for 16,000 Vera Rubin GPUs, which the company says means more GPU density per square foot and more room for cluster expansion in a fixed data center space.
• 30% improvement in power efficiency versus the SN5600/SN5610 for the same 16,000 GPUs. CoreWeave says the saved megawatts reduce costs, simplify operations and free up power for higher GPU density.

The company adds that consuming less power, space and material per unit of compute lowers the carbon footprint per unit of AI compute delivered, and describes liquid-cooled switching as both a performance and a sustainability advantage as clusters grow.

One management domain for compute and network

CoreWeave extends its existing rack management tools to the switch rack. Valvey, its patent-pending programmable per-rack liquid cooling valve assembly, monitors and controls flow rate, temperature, pressure and leak detection. It now covers the NVIDIA Spectrum-6 SPX switch rack, bringing the network into the same thermal management loop as Vera Rubin NVL72.

Racky, the rack control manager, aggregates power, cooling and environmental sensors into one interface and surfaces telemetry such as leak detection, flow rates and temperature. The Rack Lifecycle Controller (RLCC) treats an entire rack, including GPUs, networking, power and cooling, as a single programmable resource and orchestrates firmware, configuration, health monitoring and recovery.

According to CoreWeave, customers get one control surface, one observability layer and one capacity model covering both compute and network fabric. The company also says this unified telemetry helps CoreWeave ARIA, a research and iteration agent built into Weights & Biases, give researchers cleaner answers when iterating on a model.

What it means for AI customers

For customers running trillion-parameter models, agentic systems or reinforcement learning pipelines, which CoreWeave says depend on fast, low-latency GPU-to-GPU communication, the pitch is a network that does not limit what the accelerators can do. The figures come from CoreWeave itself and are tied to its 16,000-GPU comparison.

The company frames the broader trend as compute, networking and cooling converging into a single, software-defined infrastructure layer, and says it already runs this model from NVIDIA GB200 NVL72 and GB300 NVL72 systems through Vera Rubin NVL72 and the SN6600-LD.