IBM has made Nvidia’s HGX B300 systems available on IBM Cloud, opening managed access to high-end AI training infrastructure for regulated industries. The systems went live around April 29 as part of the expanded IBM-Nvidia partnership announced at GTC 2026 in March. Each HGX B300 node includes eight Blackwell Ultra GPUs and delivers up to 144 PFLOPS in FP4 precision and 72 PFLOPS in FP8 precision, alongside more than 2 TB of shared HBM3e memory (high-bandwidth graphics memory), a setup designed to keep large language models in memory and reduce performance losses from moving data between devices. Networking runs at 800 Gb/s per GPU through Nvidia ConnectX-8 adapters, aimed at preventing interconnect bottlenecks in multi-node training. IBM is pairing the systems with Red Hat OpenShift (container orchestration platform) and watsonx to offer governed AI workflows in hybrid cloud environments, allowing enterprises to train and deploy models while keeping sensitive data within approved boundaries. Initial availability is limited to select customers and regions, with a broader rollout expected later. IBM also said serverless fleet access (on-demand compute without infrastructure management) is anticipated in May 2026. The new systems sit above IBM Cloud’s existing H200 instances, which serve workloads that do not need Blackwell Ultra-class performance.