CoreWeave Powers Agentic AI, HPC, and Data Analytics Workloads with NVIDIA HGX Rubin NVL8

Harsh Singh Banwait

The arrival of the NVIDIA Vera Rubin architecture raises the bar for compute performance and energy efficiency, with the architectural flexibility next-generation workloads demand.

For agentic AI, large-scale analytics, and HPC, memory capacity and bandwidth are often the limiting factors. NVIDIA HGX Rubin NVL8 brings HBM4 and sixth-generation NVIDIA NVLink together in an eight-GPU node built for those workloads, and it's coming this quarter to CoreWeave Cloud. Here's what it delivers, where it fits, and how we're preparing to run it.

NVIDIA HGX Rubin NVL8 delivers performance gains at lower operating costs

HGX Rubin NVL8 integrates eight NVIDIA Rubin SXM GPUs onto a single baseboard. NVIDIA indicates the system delivers up to 10x higher token throughput than previous architectures while matching training performance using 4x fewer GPUs. For engineering teams, this efficiency translates directly into lower operating costs, faster iteration cycles, and the ability to serve real-time, multi-step reasoning workloads at scale.

The key hardware specifications enabling these performance gains include:

Increased memory: 2 TB of total HBM4 and 130 TB/s of aggregate memory bandwidth, enabling real-time processing of massive datasets in memory for accelerated data analytics and high-throughput agentic workflows.
Low-latency GPU interconnect: Sixth-generation NVLink at 24 TB/s of within-node NVLink Switch bandwidth, driving ultra-low latency inter-GPU communication for complex tensor-parallel agent inference and tight HPC simulation loops.
High-bandwidth scale-out networking: Eight NVIDIA ConnectX-9 SuperNICs provide up to 800 Gb/s of backend connectivity per GPU, supporting communication across nodes for distributed training, HPC, and large-scale data analytics workloads.
Infrastructure offload: NVIDIA BlueField-4 provides frontend connectivity and supports networking, storage, and security offloads, helping preserve host CPU resources for application workloads.

For agentic AI, a 10x higher token throughput means more agents and users can be served on the same infrastructure, so every developer gets fast, multi-step help even during peak hours. With this ultra low latency and high-bandwidth networking, HPC simulations spend less time waiting on GPUs to share data, so a drug discovery team can screen more candidate compounds per day and get results to researchers sooner. And data analytics workloads can move large datasets between nodes faster, so a financial services firm can rerun risk analytics on the day’s transactions and give analysts updated views before markets open.

Line chart of tokens per second per megawatt versus tokens per second per user for Kimi K2 Thinking. HGX Rubin NVL8 stays well above HGX B200 across the curve, with a 10x gap marked at about 175 tokens per second per user. — NVIDIA HGX Rubin NVL8 is projected to deliver up to 10x more tokens per second per megawatt than NVIDIA HGX B200 when running Kimi K2 Thinking (4K input/4K output).

Line chart of training throughput versus number of GPUs for DeepSeek-R1. HGX Rubin NVL8 reaches the same throughput as HGX B200 with one-quarter the GPUs. — NVIDIA HGX Rubin NVL8 matches NVIDIA HGX B200 training throughput on DeepSeek-R1 using 4x fewer GPUs.

Unlocking peak performance with the CoreWeave advantage

CoreWeave completed the industry's first bring-up, validation, and benchmarking of NVIDIA Vera Rubin NVL72, and customers are already seeing results. On Vera Rubin NVL72, Cognition achieved 4.8x higher SWE-2 inference throughput and 3.8x faster output token speed for RL.

Building on this momentum, CoreWeave offers a specialized platform with features and tools designed to maximize efficiency and scale the performance of HGX Rubin NVL8 across agentic AI, HPC, and data analytics workloads:

CoreWeave Innovation	How It Unlocks Peak Performance
CoreWeave Mission Control®	CoreWeave Mission Control helps HGX Rubin NVL8 sustain peak performance by treating infrastructure as one continuously monitored system, from rack power and cooling to GPU behavior and workload performance. HGX Rubin NVL8 is 100% liquid cooled. To manage the cooling system, CoreWeave custom-built Racky and Valvey. Racky provides the rack-level control point, while Valvey executes cooling and isolation actions. Rack Life Cycle Controller (RLCC) orchestrates bring-up, health, and recovery workflows. Together, these capabilities keep agentic AI serving responsive, HPC jobs synchronized, and data analytics pipelines moving by identifying and resolving infrastructure bottlenecks before they occur.
CoreWeave Kubernetes Service (CKS)	CKS gives customers the managed cluster, performance, security, and operational foundation they need so they can focus on their agentic, HPC, or data analytics workloads instead of Kubernetes infrastructure.
CoreWeave SUNK	SUNK combines Slurm’s resource-aware scheduling with Kubernetes’ containerization and elasticity, allowing teams to run HPC and data-processing jobs on shared infrastructure instead of maintaining separate environments.
CoreWeave Direct-to-Expert Support	CoreWeave customers work directly with the engineers who brought up Rubin-generation systems, giving them hands-on guidance to validate workload fit, tune deployments, and troubleshoot performance across agentic AI, HPC, and data analytics workloads.

Use case fit: where NVIDIA HGX Rubin NVL8 excels

Memory capacity, memory bandwidth, and fast GPU-to-GPU communication shape how each workload performs on HGX Rubin NVL8. Here's how that plays out for agentic AI, data analytics, and HPC.

Agentic AI

Memory footprint and KV cache: With 2 TB of HBM4 memory, HGX Rubin NVL8 comfortably hosts large key-value (KV) caches for million-token context windows, so multiple agent sessions can stay active at once. Single-node deployments execute models and caches seamlessly, and 24 TB/s NVLink bandwidth eliminates fabric latency during tensor-parallel serving.
Throughput efficiency: With up to 10x higher token throughput per megawatt than previous architectures, you can run concurrent lightweight agents and multi-step reasoning workflows at a lower cost.

Data analytics

Scale-out pipeline acceleration: GPU-accelerated ETL, dataframe query processing, vector indexing, and embedding generation scale horizontally across nodes using high-bandwidth backend networking.
Node economics: With 130 TB/s of aggregate memory bandwidth, HGX Rubin NVL8 delivers superior cost-per-row processing and throughput for large-scale data pipelines.

HPC

Double-precision density: At 265 TFLOPS of FP64 double-precision performance per eight-GPU node, the system efficiently handles scientific simulations, molecular dynamics, and computational finance workloads.
Unified simulation and AI workloads: High memory bandwidth and fast inter-GPU communication let you run hybrid pipelines that combine classical HPC simulations with AI surrogate models on unified hardware.

Get ready for NVIDIA HGX Rubin NVL8

Coming to CoreWeave Cloud this quarter, NVIDIA HGX Rubin NVL8 brings teams running agentic AI, data analytics, and HPC workloads Rubin-generation performance on an AI cloud purpose-built to put it to work.

Start planning for NVIDIA HGX Rubin NVL8. Request a briefing for capacity planning, onboarding timeline, and workload fit — or request an ARENA Pass to run your workload on HGX Rubin NVL8 yourself.

Want to go deeper? Explore our related resources:

Published on

October 21, 2026

CoreWeave Powers Agentic AI, HPC, and Data Analytics Workloads with NVIDIA HGX Rubin NVL8

Harsh Singh Banwait

Copied

Eight NVIDIA Rubin GPUs, 2 TB of HBM4, and NVIDIA NVLink in one node. See how NVIDIA HGX Rubin NVL8 scales agentic AI, data pipelines, and HPC simulations.

Copied