CoreWeave has deployed a multi-rack NVIDIA Vera Rubin NVL72 cluster on CoreWeave Cloud, interconnecting hundreds of Rubin GPUs into a unified scale-out environment optimized for agentic AI workloads. Each Dell-manufactured rack holds 72 Rubin GPUs and 36 Vera CPUs, adopting NVLink 6 as its scale-up interconnect fabric, paired with two ConnectX-9 SuperNICs per GPU. Racks are linked via NVIDIA Spectrum-X Ethernet within a two-tier non-blocking fabric, which CoreWeave states can scale to roughly 128,000 GPUs. CoreWeave is the first cloud provider to successfully validate and activate a Vera Rubin NVL72 system, and the first to publish its measured performance. Its initial peer-reviewed benchmark results were released last week, based on a GB300 NVL72 submission. Alongside the compute upgrade, CoreWeave AI Object Storage adds cross-region write acceleration and a brand-new Archive tier.

Multi-Rack NVIDIA Vera Rubin NVL72 Architecture

This multi-rack deployment targets agentic execution workflows, where sequential reasoning loops and external tool calls amplify data-access latency across distributed infrastructure. The dual ConnectX-9 SuperNICs fitted to each Rubin GPU deliver up to 1.6 Tb/s backend network bandwidth per GPU over multi-rail, multi-plane links. The fabric uses a modular design, enabling new NVL72 racks to join the same non-blocking topology as capacity expands. The racks operate with 45°C liquid cooling.

CoreWeave unifies these racks using its Mission Control software suite. A Rack LifeCycle Controller handles rack provisioning and lifecycle management, including hardware detection via firmware flashing, validation, power and thermal monitoring. The Racky rack-management module controls power and infrastructure, while the programmable Valvey cooling controller delivers software-defined visibility and management for liquid cooling circuits and environmental sensors.

Racks enter production only after phased validation testing. At node level, this includes hours of repeated GPU diagnostics, CPU-GPU transfer verification, interconnect and thermal load tests, plus realistic training workloads. At rack level, synchronized jobs across all 72 GPUs validate NVLink GPU-to-GPU throughput; any hardware falling below performance thresholds goes for troubleshooting. For cross-rack verification, CoreWeave runs distributed workloads, deliberately disabling NVLink to force traffic across the backend network, and monitors the physical fabric for unstable links, rising error rates, overheating components and imbalanced traffic. Test loads mimic agentic applications, featuring sudden demand spikes, variable concurrency and bursty communication patterns.

AI Object Storage Updates: Cross-Region Writes and Archive Tier

The storage enhancement builds upon its Local Object Transport Accelerator (LOTA), a caching proxy deployed on every GPU and CPU node within CoreWeave Kubernetes Service, storing object data on node-local NVMe. CoreWeave reports cached reads reaching up to 7 GB/s per GPU, with p99 read latency over 8x lower than direct bucket access. One leading model provider runs LOTA across more than 15,000 GPUs and 20PB cache, achieving a 99.7% hit rate, with LOTA offered at no extra charge.

Cross-region write acceleration allows applications to write to a bucket in a remote region with local latency, without modifying APIs or SDKs. Applications send standard S3 writes to the LOTA endpoint; object data is persistently stored locally while metadata commits to the remote region. The object becomes instantly readable, even for the writing job, and data migrates to the remote region in the background. Applications see a single bucket namespace spanning all regions with consistent IAM policies, lifecycle rules and access controls, eliminating separate regional buckets and replication pipelines typically required for cross-site checkpoint distribution.

“Our datasets span multiple regions, and we can’t afford to have our training schedule dictated by cross-region retrieval delays,” said Cécile Robert-Michon, director of internal infrastructure at Cohere. “CoreWeave AI Object Storage gives us a unified dataset footprint across regions with reads cached locally, so nothing waits on the network.”

The Archive tier is the fourth storage class alongside Hot, Warm and Cold, designed for write-once, rarely-read data such as historical training checkpoints and raw datasets. It features lower base capacity pricing and waives retrieval, cache, per-request, early-deletion, tiering and egress fees. CoreWeave recommends retaining recent checkpoints on faster tiers for quick restarts, with automatic tiering to Archive after 60 days with no reads. Cross-region writes and the Archive tier are now generally available.

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!