CoreWeaveIn Development
An eight-GPU Rubin node CoreWeave said on 21 Oct 2026 is coming to its cloud this quarter. The post lists 2 TB of HBM4, 130 TB/s of aggregate memory bandwidth, sixth-generation NVLink at 24 TB/s within the node, eight ConnectX-9 SuperNICs at up to 800 Gb/s per GPU, BlueField-4, and 265 TFLOPS of FP64 per node. NVIDIA's 10x token-throughput comparison and the 4x-fewer-GPU training comparison are not CoreWeave measurements of a shipped cluster.
Latest update Oct 22, 2026
01
Announced 21 Oct 2026 as coming this quarter. Not described as generally available. No product image was published in the post.
02
Announced next steps from CoreWeave that mention NVIDIA HGX Rubin NVL8.
Q4 2026
03
CoreWeave Cloud is the company's AI cloud. The 2025 Form 10-K says it covers large-scale training, inference, data movement, continuous iteration, and agentic workflows. Named software includes Mission Control and Slurm on Kubernetes, called SUNK. The company calls the service The Essential Cloud for AI.
The GPU Compute page offers NVIDIA GPUs on bare metal and names GB200 NVL72, H200, H100, and L40S systems. On 2 Oct 2026 CoreWeave said Vera Rubin NVL72 is in limited availability, with Cognition as the first production customer. That is not general availability.