CoreWeave Makes NVIDIA HGX B300 Generally Available for Agentic AI

What CoreWeave is offering

CoreWeave argues that the rise of agentic AI is pushing cloud requirements toward inference. In its words, long context windows need more memory, multi-step reasoning keeps GPUs busy for longer, and agent coordination depends on low-latency, high-bandwidth interconnects across large clusters.

NVIDIA HGX B300 is the company's answer to those demands. According to CoreWeave, it doubles interconnect speed through NVIDIA Quantum-X800 InfiniBand networking, NVIDIA BlueField-3 data processing units (DPUs) and 800 Gbps NVIDIA ConnectX-8 SuperNICs. It also raises GPU memory capacity by 50% over the NVIDIA HGX B200 and improves NVFP4 inference performance.

CoreWeave says Cursor is already moving toward production workloads on the new hardware.

Hardware specifications and network design

The HGX B300 has 270 GB of memory, up from 180 GB on the HGX B200. CoreWeave says this lets larger models and longer context windows run directly in high-bandwidth memory, which reduces reliance on external memory tiers. NVFP4 performance rises from 9 PFLOPs to 14 PFLOPs, according to the company.

For training, the HGX B300 keeps the same high-precision performance profile as the HGX B200 across FP32, FP16, BF16 and FP8. Both systems offer 1.8 TB/s per-GPU interconnect bandwidth, 14.4 TB/s aggregate NVSwitch bandwidth and 7.7 TB/s of memory bandwidth.

CoreWeave's comparison table lists total GPU memory of 2.1 TB for the HGX B300 against 1.4 TB for the HGX B200, and networking bandwidth of 1.6 Tb/s against 0.8 Tb/s. Attention performance is listed as 2x versus 1x.

On the network side, CoreWeave says its infrastructure uses non-blocking NVIDIA Quantum X-800 InfiniBand and RoCE fabrics in a fat-tree architecture. The company claims this provides 2x the node-to-node bandwidth across multi-thousand GPU clusters, with no oversubscribed links.

Platform tools and customer views

CoreWeave pairs the hardware with its own software stack. Mission Control is meant to shift responsibility for cluster health from the customer to CoreWeave. The company says HGX B300 instances are integrated with its fleet lifecycle controllers, so teams can begin training or fine-tuning "within hours, not weeks."

• CoreWeave Kubernetes Service (CKS): Kubernetes-native orchestration for deploying, scaling and managing agent workflows.
• Slurm on Kubernetes (SUNK): lets users run more experiments in parallel while maximizing GPU efficiency.
• W&B Weave: end-to-end tracing of agent runs to diagnose failures in intermediate reasoning steps.
• CoreWeave AI Object Storage: speeds of up to 7 GB/s per GPU, intended to keep NVIDIA Blackwell Ultra GPUs from idling on storage or networking.

The company also cites 96% goodput, which it says means nearly all requested compute turns into useful work.

Aman Sanger, co-founder of Cursor, said the company has already run production workloads with CoreWeave on NVIDIA HGX B200. Moving toward B300, he said, that "proven operating model and continuity in execution" lets the team focus on building AI code generation systems. Orian Leitersdorf, chief scientist at Decart, pointed to the HGX B300 and the upcoming NVIDIA Vera Rubin platform as enabling real-time generative and agentic AI at scale.

What it means for AI teams

For teams building agent systems, the release gives access to a GPU generation with more memory per accelerator and higher NVFP4 throughput, delivered with CoreWeave's managed operations and tooling. The performance figures come from CoreWeave's own benchmarks against the older HGX H200, so they are the company's claims rather than independent results.

The release also fits a broader message CoreWeave has promoted around NVIDIA GTC 2026, where it described the HGX B300 launch as one of its major infrastructure announcements for teams moving from experimentation into production.