CoreWeave Launches NVIDIA Grace Blackwell GB200 Systems, Becoming First Cloud Provider with General Availability

News Overview

CoreWeave, an AI-focused cloud provider, announced it is one of the first to bring NVIDIA’s powerful GB200 NVL72 Grace Blackwell systems online and make them generally available at scale.
Major AI companies, including Cohere, IBM, and Mistral AI, are already utilizing these thousands of newly available Blackwell GPUs on CoreWeave’s platform to train and deploy next-generation AI models and applications.
This deployment positions CoreWeave as a key infrastructure provider for advanced AI development, offering cutting-edge compute resources designed for demanding tasks like large-scale model training, reasoning, and agentic AI.

NVIDIA GB200 NVL72 System: This is a rack-scale system connecting 72 NVIDIA Blackwell GPUs with 36 Arm-based NVIDIA Grace CPUs, designed to function as a single, massive GPU. It features a large unified memory pool and a 72-GPU NVLink domain for high-speed interconnectivity.
Performance Leap: The Blackwell architecture introduces significant performance gains. CoreWeave demonstrated a 2.86x per-chip performance increase on the Llama 3.1 405B model compared to NVIDIA H200 GPUs in MLPerf Inference v5.0 benchmarks, achieving over 800 tokens per second (TPS) per GB200 instance (2 Superchips / 4 GPUs). This is partly attributed to new features like FP4 precision support in the second-generation Transformer Engine.
Capabilities: The platform is engineered for up to 4x faster training and 30x faster real-time inference on trillion-parameter models compared to the previous generation (Hopper). It excels at AI reasoning, agentic AI development, and large-scale data processing.
Infrastructure: CoreWeave’s deployment utilizes NVIDIA Quantum-2 InfiniBand networking for high-throughput, low-latency communication between potentially up to 110,000 GPUs. They also offer managed services like CoreWeave Kubernetes Service and Slurm on Kubernetes (SUNK) optimized for these systems. IBM is notably pairing the GB200 systems with IBM Storage Scale System for high-performance AI storage within CoreWeave.

Market Impact: CoreWeave’s rapid deployment of Blackwell GPUs solidifies its position as a leading specialized “AI Hyperscaler,” directly competing with larger cloud providers by offering first-to-market access to NVIDIA’s latest, most powerful hardware. This availability accelerates the development and deployment of cutting-edge AI models by major players.
Competitive Positioning: Being among the first to offer GB200 NVL72 at scale gives CoreWeave a significant advantage in attracting AI labs and enterprises needing massive computational power. This move underscores the intense demand for state-of-the-art AI infrastructure.
Strategic Considerations: This highlights the crucial role specialized cloud providers play in the AI ecosystem, often moving faster than traditional hyperscalers to adopt new hardware. It also reflects NVIDIA’s strategy of enabling partners like CoreWeave to bring its latest technology to market quickly, potentially creating more competition and options for AI developers beyond the largest cloud platforms. The focus on rack-scale, integrated systems like the GB200 NVL72 signals a shift towards more holistic, powerful AI “factories.”