IBM has made Nvidia’s HGX B300 GPU systems available on IBM Cloud, giving regulated industries access to some of the most powerful AI training hardware on the market through a managed cloud environment. The systems went live around April 29, marking the latest step in an expanded partnership between the two companies that was first announced at GTC 2026 in March.
What’s inside the box
Each HGX B300 system packs eight Blackwell Ultra GPUs into a single node. The raw compute numbers are staggering: up to 144 PFLOPS in FP4 precision and 72 PFLOPS in FP8. Each system offers more than 2 TB of shared HBM3e memory, which matters enormously for large language model training where the entire model needs to fit in GPU memory to avoid performance-killing data shuffling. The networking side runs at 800 Gb/s per GPU via Nvidia’s ConnectX-8 adapters, ensuring that multi-node training jobs don’t bottleneck at the interconnect layer.
The compliance angle
The HGX B300 deployment integrates with Red Hat OpenShift and IBM’s watsonx platform, creating what IBM frames as a complete stack for governed AI workflows. OpenShift handles the container orchestration that lets workloads move between on-premises and cloud environments. Watsonx provides the AI development tooling and governance layer. Together, they allow enterprises to train and deploy models in a hybrid-cloud setup where sensitive data stays within approved boundaries.
Timing and rollout
The initial availability targets select customers and regions, with a broader rollout expected to follow. IBM has also signaled that serverless fleet access for the HGX B300 systems is anticipated in May 2026, which would let customers consume GPU compute without managing underlying infrastructure.
The HGX B300 systems complement IBM Cloud’s existing H200 instances, which offer a lower performance tier for workloads that don’t require Blackwell Ultra-class hardware.