HPE is deepening its collaboration with NVIDIA to deliver a next-generation AI infrastructure platform, designed to push the boundaries of large-scale deployments. The focus is on systems capable of handling frontier-scale models—those exceeding one trillion parameters—while significantly improving data center efficiency.
A Rack-Scale Powerhouse
The partnership introduces the NVIDIA Vera Rubin NVL72 by HPE, a rack-scale system engineered for neo-cloud environments. It packs 36 NVIDIA Vera CPUs and 72 Rubin GPUs, leveraging sixth-generation NVLink scale-up networking and NVIDIA BlueField-4 DPUs to optimize performance. Liquid cooling integration ensures thermal efficiency while maintaining high throughput.
Doubling GPU Density
Complementing the NVL72 is the HPE Compute XD700 server, which delivers double the GPU density of its predecessor. It supports up to 128 Rubin GPUs per rack, addressing critical challenges in space, power, and cooling costs. This marks a significant leap forward for AI training and inference workloads.
Enhanced Performance with Blackwell GPUs
A key innovation is the introduction of NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs across HPE's AI factory portfolio. These GPUs, paired with enhanced software and services from both companies, streamline large-scale AI deployments while maintaining enterprise-grade reliability.
Multi-Tenancy and Enterprise Support
The AI Factory portfolio also supports multi-tenancy models through GPU passthrough for virtual machines and secure Kubernetes namespaces via NVIDIA Multi-Instance GPU (MIG). This flexibility allows service providers to choose between hard and soft tenancy deployment models, catering to diverse customer needs. Additionally, the platform integrates Red Hat Enterprise Linux and OpenShift, providing a robust enterprise-grade foundation for AI workloads.
Exascale Vision for Unified Computing
Looking ahead, HPE's next-generation exascale-class supercomputing platform, the Cray Supercomputing GX5000, will incorporate NVIDIA Vera CPU compute blades. Each blade can house up to 16 Vera CPUs, scaling to 40 blades with 640 Vera CPUs and 56,320 Olympus Arm-compatible cores per rack. This platform is designed to unify AI and high-performance computing (HPC) workloads, enabling faster scientific discovery.
Advanced Connectivity for Efficiency
The GX5000 will also integrate NVIDIA Quantum-X800 InfiniBand switches, featuring 144 ports of 800 Gb/s connectivity. These switches include power efficiency enhancements such as low-power link state and power profiling, further reducing operational costs.
Broadening AI Adoption
The solutions introduced by HPE and NVIDIA are part of a broader effort to accelerate AI adoption in critical fields like medicine, life sciences, engineering, and manufacturing. The partnership leverages HPE's decades-long expertise in building large-scale supercomputers and its leadership in data center design and liquid cooling.
Availability and Impact
The NVIDIA Vera Rubin NVL72 by HPE is set to be available in December 2026, while the HPE Compute XD700 will follow in early 2027. The NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs are already accessible through HPE's AI factory portfolio. With these advancements, the partnership is poised to set new benchmarks in AI infrastructure, offering higher performance and efficiency for enterprise and research applications.
