TechPrufer
Get started
Coreweave
Coreweave logoCoreweave · Feature trail

Liquid Cooling Infrastructure

13 posts · May 2026 – Sep 2026

What this trail covers

CoreWeave enhances its AI infrastructure by deploying NVIDIA Spectrum-X SN6600-LD liquid-cooled Ethernet switches, doubling per-rack network bandwidth to 1.64 Pb/s. This enables a non-blocking, multi-plane, multi-rail spine and leaf fabric for Vera Rubin NVL72 GPUs, reducing network latency and keeping GPU idle time near zero. The deployment integrates with CoreWeave Mission Control's Valvey and Racky for unified thermal management and operational control, and the Rack Lifecycle Controller (RLCC). This post details customer validation of the Vera Rubin NVL72 platform, showing a 4.8x increase in total token throughput for SWE-2 inference workloads and a 3.8x boost in output token throughput for Reinforcement Learning (RL) compared to previous generation GB200 NVL72, demonstrating the performance benefits of the new hardware and liquid cooling for agentic AI workloads.

All Coreweave trails
  1. First Vera Rubin NVL72 Customer Sees 4.8x Throughput | CoreWeaveThis post details customer validation of the NVIDIA Vera Rubin NVL72 platform on CoreWeave Cloud, specifically highlighting Cognition's experience. It reports a 4.8x increase in total token throughput for SWE-2 inference workloads and a 3.8x boost in output token throughput for Reinforcement Learning (RL) when comparing Vera Rubin NVL72 against NVIDIA GB200 NVL72. The post emphasizes how these performance gains, enabled by the new hardware and liquid cooling infrastructure, directly benefit agentic AI workloads by accelerating reasoning, retrieval, and iteration cycles.
  2. A Deep Dive on CoreWeave Innovations | CoreWeave BlogThis post details the engineering innovations behind CoreWeave's deployment of NVIDIA Vera Rubin NVL72, focusing on rack-scale AI infrastructure. Key contributions include the patent-pending Valvey system for software-defined liquid cooling with per-rack isolation, Racky for unified rack management and sensor aggregation, and the integration of liquid-cooled NVIDIA Spectrum-X SN6600 Ethernet switches to achieve 1.64 Pb/s per rack bandwidth. It also discusses the choice of multi-rail, multi-plane networking (InfiniBand and RoCE) with ConnectX-9 SuperNICs for large-scale training and inference clusters, and the overall software stack engineering to maximize FLOPs utilization.
  3. CoreWeave Brings Liquid Cooled Switching to AI Networking | CoreWeave BlogThis post details the deployment of NVIDIA Spectrum-X SN6600-LD liquid-cooled Ethernet switches, which double per-rack network bandwidth to 1.64 Pb/s. It explains how this enables a non-blocking, multi-plane, multi-rail spine and leaf fabric for Vera Rubin NVL72 GPUs, reducing network latency and GPU idle time. The integration with Valvey and Racky for unified thermal management and the Rack Lifecycle Controller (RLCC) for automated rack management are also highlighted.
  4. First-Ever Measured Vera Rubin NVL72 Silicon Performance Stats | CoreWeave BlogThis post provides the first measured silicon performance statistics for the NVIDIA Vera Rubin NVL72 on CoreWeave, demonstrating a 10x improvement in tokens per megawatt for the DeepSeek R1 inference workload compared to NVIDIA GB200 NVL72. It details the architectural advantages of Vera Rubin NVL72, including its GPU count, CPU count, NVLink fabric bandwidth, and NVFP4 support, and highlights the impact of these on inference performance and cost-efficiency for agentic AI workloads. The post also notes that this is an initial measurement and further optimizations are expected.
  5. This post details the performance gains and architectural advantages of deploying NVIDIA GB300 NVL72 instances with NVIDIA Blackwell Ultra GPUs for AI inference. It quantifies a 6.5x performance improvement on the DeepSeek R1 model by comparing a 16-GPU H100 system to a 4-GPU GB300 system, attributing the uplift to the GB300's superior memory and interconnect bandwidth, enabling 4-way Tensor Parallelism instead of 16-way. The post also outlines CoreWeave's infrastructure enhancements, such as a topology-aware scheduler and automated Rack LifeCycle Controller, which optimize the utilization of the GB300's NVLink and networking capabilities for AI workloads.
  7. CoreWeave Data Center Operations: Built for AI | CoreWeaveThis post details CoreWeave's investment in direct-to-chip liquid cooling for its data centers, enabling higher rack power densities (up to 130kW per rack) and improved thermal management for AI workloads. It highlights the sustainability benefits and the integration of liquid cooling with existing air cooling infrastructure to support NVIDIA GB200 NVL72 racks. The post also touches upon the operational aspects of data center technicians (DCTs) and their role in maintaining this advanced infrastructure.
  8. CoreWeave Leads the Way with First NVIDIA GB300 NVL72 DeploymentThis post announces the first deployment of NVIDIA GB300 NVL72 instances on CoreWeave, detailing the hardware specifications (72 Blackwell Ultra GPUs, 36 Grace CPUs, 18 BlueField-3 DPUs), performance improvements for AI reasoning (up to 10x user responsiveness, 5x throughput per watt, 50x output for reasoning model inference), and enhanced memory (21TB HBM3e per rack). It also highlights platform-level optimizations like RLCC, Cabinet Wrangler, Cabinet Details dashboard, and integration with Weights & Biases for infrastructure observability. The post emphasizes the collaboration with Dell, Switch, and Vertiv for this deployment.
  9. Our Networking Services For GenAI | CoreWeaveThis post details CoreWeave's networking services for GenAI, focusing on the use of NVIDIA InfiniBand for high-throughput, low-latency GPU-to-GPU communication, including its SHARP and SHIELD features. It also covers direct connections for multi-cloud data transfer with no egress fees and the upgrade of Ethernet networking using NVIDIA Spectrum switches and BlueField DPUs for improved performance between GPU nodes and storage clusters.
  10. The Future of AI Clusters for Enterprise in 2025This post details the evolution of AI clusters for enterprises, focusing on power and efficiency challenges at the hardware, data center, and operational levels. It highlights advancements in GPUs (H100, Blackwell, GB200), DPUs (NVIDIA BlueField-3) for offloading and security, and BMCs for out-of-band monitoring. Data center evolutions include the widespread adoption of liquid cooling to manage higher power demands and enable denser GPU deployments, the critical role of Data Center Technicians (DCTs) in physical infrastructure deployment and maintenance, and closer collaboration with partners like NVIDIA and data center providers to standardize hardware and data center designs for future AI workloads.
  11. The Redesign of the Data Center Has Already Started | CoreWeaveThis post details the ongoing redesign of CoreWeave's data centers to specifically support AI workloads. It elaborates on the challenges of traditional data centers (architecture, cooling, software) and how CoreWeave is addressing them. Key technical contributions include: 1. **Power Management:** Transitioning from 2N to N+1 redundancy, implementing dynamic workload scheduling based on power infrastructure optimization, and preemptive adjustment of cooling and power distribution. 2. **Liquid Cooling:** Designing infrastructure for circulation of cooling liquid (water or specialized coolant) to GPU servers, including specialized plumbing, pumps, heat exchangers, and reservoirs, ensuring leak-proof and corrosion-resistant operation. 3. **Networking:** Building GPU clusters with technologies to optimize performance and reliability for massive-scale parallel processing, facilitating high-speed, efficient communication between GPUs. 4. **Software Stack:** Developing a purpose-built software stack for provisioning, hardware validation, health checking, and intelligent orchestration/scheduling to optimize FLOPs for production workloads 24/7/365, including tools like SUNK (Slurm on Kubernetes) and Tensorizer (near-zero-copy model loading).
  12. Why ClusterMAX 2.0 Validates CoreWeave’s EngineeringThis post elaborates on CoreWeave's engineering approach to AI infrastructure, as validated by the ClusterMAX 2.0 report. It details specific technical implementations including the SUNK framework for job queuing and resource management, a custom Rack LifeCycle Controller for managing GB200/GB300 systems as unified objects, the use of NVIDIA BlueField DPUs for workload isolation and performance, end-to-end monitoring leveraging DCGM, NVML, and interconnect telemetry, AI Object Storage with LOTA caching for data proximity, and a zero-trust security architecture. It also emphasizes the direct-to-expert support model.
  13. Why Liquid Cooling Matters for AI Scale | CoreWeaveThis post explains why liquid cooling is essential for scaling AI workloads, focusing on the increasing power density of modern GPU architectures like the NVIDIA GB200 NVL72. It details the limitations of traditional air cooling in high-density racks and highlights the benefits of liquid cooling for CoreWeave customers, such as enabling more GPUs per rack, faster inter-GPU communication, access to next-generation hardware, and improved power efficiency leading to better TCO. The post also touches on the operational aspects of implementing liquid cooling in data centers, emphasizing redesign over retrofitting and the continued role of air cooling for less intensive components.
  14. Liquid Cooling for AI Data Centers | CoreWeave BlogThis post details CoreWeave's adoption of direct-to-chip liquid cooling and closed-loop liquid recycling systems for AI data centers. It explains the technical rationale for liquid cooling over air cooling due to higher molecular density and heat absorption capacity. The post highlights how direct-to-chip cooling interfaces directly with AI chips to efficiently transfer heat, and how closed-loop systems recycle coolant for environmental benefits. It also mentions the use of proactive AI agents for thermal management at scale and the integration with CoreWeave Mission Control for unified thermal management. The core technical contribution is the operationalization of liquid cooling as a critical architectural pillar for enabling high-density AI capacity, reliability, and performance.