HPE expands NVIDIA-based Cray & AI Factory systems

HPE has expanded its AI and supercomputing portfolio with new NVIDIA-based compute and networking options for its Cray systems, plus updated configurations for its AI Factory range aimed at large-scale deployments.

The additions include new CPU compute blades, high-speed InfiniBand networking, and new rack-scale and server platforms based on NVIDIA's latest roadmaps. HPE also outlined updates to software support and operational tools for its AI Factory products.

Cray updates

The main change is a new compute blade option for the HPE Cray Supercomputing GX5000 platform. The HPE Cray Supercomputing GX240 Compute blade includes up to 16 NVIDIA Vera CPUs and uses liquid cooling. HPE positioned it as part of a second-generation exascale-class platform designed to run AI workloads alongside more traditional high-performance computing.

In a rack configuration, the design can scale to 40 blades with 640 NVIDIA Vera CPUs per rack and up to 56,320 NVIDIA Olympus Arm-compatible cores per rack.

The GX5000 line also adds NVIDIA Quantum-X800 InfiniBand. The switches provide 144 ports with 800 Gb/s connectivity per port, along with power-efficiency features such as low-power link state and power profiling.

HPE said organisations are increasingly blending AI with established simulation and modelling workloads, and positioned the new options for research laboratories, sovereign entities, and large enterprises running large-scale systems.

“Having built the three most powerful, exascale supercomputers in the world, HPE is at the forefront of innovation that brings together cutting-edge AI workloads with traditional HPC to accelerate scientific breakthroughs,” said Trish Damkroger, Senior Vice President and General Manager, HPC & AI Infrastructure Solutions at HPE. “Our continued collaboration with NVIDIA helps customers tap into the high-performance density they need to push the boundaries in the fields of medicine, life sciences, engineering, manufacturing, and more.”

AI Factory systems

HPE also refreshed its AI Factory portfolio with systems based on NVIDIA Vera Rubin and NVIDIA Blackwell, targeting service providers, sovereign deployments, and large enterprises.

For large, shared infrastructure providers, it introduced the NVIDIA Vera Rubin NVL72 by HPE rack-scale system. HPE described it as engineered for frontier-scale models exceeding one trillion parameters, with 36 NVIDIA Vera CPUs and 72 NVIDIA Rubin GPUs.

It also includes sixth-generation NVIDIA NVLink scale-up networking, NVIDIA ConnectX-9 SuperNIC, and NVIDIA BlueField-4 DPUs. HPE said the system integrates liquid cooling and draws on its data centre design services for large AI installations.

Another addition is the HPE Compute XD700, an Open Compute Project-inspired GPU server built on NVIDIA HGX Rubin NVL8. HPE said it is designed to deliver higher GPU density per rack while reducing space, power, and cooling costs for training and inference.

Each rack of XD700 servers supports up to 128 Rubin GPUs, according to HPE-double the GPU density of the previous generation, the HPE ProLiant Compute XD685.

On the Blackwell side, the NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs are now available across HPE's AI factories.

Software and operations

HPE also outlined software and services updates for the AI Factory portfolio, including alignment with the NVIDIA Cloud Partner programme. Its AI factories are ready for NVIDIA Cloud Provider certification, which HPE said streamlines validation for cloud service providers.

Multi-tenancy is another focus. HPE plans to support multi-tenancy models for virtual machines with GPU passthrough and secure Kubernetes namespaces via NVIDIA Multi-Instance GPU. It said the approach is enabled by SUSE Virtualisation and SUSE Rancher Prime Suite, offering a choice between hard and soft tenancy models.

The AI Factory portfolio also supports Red Hat Enterprise Linux and Red Hat OpenShift as part of Red Hat AI Enterprise. HPE said these integrate with NVIDIA AI Enterprise.

For operational control, HPE's AI Factory at scale and AI Factory sovereign offerings will add support for NVIDIA Mission Control software, including workload orchestration with NVIDIA Run:ai and monitoring and autonomous recovery tied to NVIDIA Dynamo.

NVIDIA described the collaboration as a response to demand for large-scale AI and HPC infrastructure across commercial and public-sector environments.

“To realise the potential of AI, enterprises and nations require infrastructure that can handle massive-scale model training and HPC workloads,” said Chris Marriott, Vice President, Enterprise Platforms at NVIDIA. “Together, HPE and NVIDIA have developed full-stack AI infrastructure that unite accelerated computing, advanced networking and liquid cooling for faster time-to-insight in at-scale and sovereign environments.”

HPE said the HPE Cray Supercomputing GX240 Compute blade with NVIDIA Vera CPUs and NVIDIA Quantum-X800 InfiniBand for the GX5000 are scheduled for 2027. The NVIDIA Vera Rubin NVL72 by HPE rack-scale system is due in December 2026, while the HPE Compute XD700 is due in early 2027.

NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs are available now across the AI Factory portfolio, according to HPE. Multi-tenancy with GPU passthrough is due in spring 2026, Red Hat integration is available now, and support for NVIDIA Mission Control is planned for 2026.