HP ZGX Fury — GB300 Grace Blackwell Ultra AI Station with 20 PFLOPS FP4

Specifications

Overview

The HP ZGX Fury is a deskside AI system built on the NVIDIA GB300 Grace Blackwell Ultra superchip, rated at up to 20 petaFLOPS of FP4 compute with 748 GB of coherent memory. That memory figure is the design point: with the model harness quantising to FP4, the system is specified to run inference on trillion-parameter-class models locally, and to develop or fine-tune models in the 100-billion-parameter class — without a cloud round trip and without token-metered economics.

Architecturally it is a single board pairing a 72-core Grace Neoverse V2 CPU with a Blackwell Ultra GPU over NVLink-C2C, carrying 496 GB of LPDDR5x SOCAMM for the CPU and 252 GB of HBM3e at 7.1 TB/s for the GPU. Storage is deliberately split: OS drives hang off the CPU on PCIe Gen5 in a RAID 1 mirror, while data drives sit behind the ConnectX-8 on PCIe Gen6, both self-encrypting. An optional discrete NVIDIA RTX PRO GPU can take over display output so the GB300 GPU is never distracted by a desktop compositor.

The system ships on Ubuntu 24.04 LTS with NVIDIA's developer stack plus HP's own Z Runtime and Z Toolkit — the latter free of charge, with MLflow experiment tracking and Ollama testing built in. In September 2026 HP extended the platform commercially: a collaboration with Red Hat and NVIDIA delivers a Red Hat AI Factory with NVIDIA environment on ZGX Fury, so a local system can be evaluated in a sandbox and then moved into production with the same software foundation, workload isolation and governance controls.

For AI infrastructure buyers the ZGX Fury fills the gap between a DGX Spark-class personal AI computer and a rack-mounted DGX system. It is the correct shape for a lab that must keep sensitive data on premises — regulated industries, government and sovereign deployments, defence, and manufacturing lines running computer-vision inference next to production equipment with no tolerance for cloud latency.

Key Benefits

20 PFLOPS FP4 in a deskside chassis removes the data-centre requirement for large-model inference. 748 GB coherent memory keeps trillion-parameter models resident and cuts explicit data movement. Fixed-cost on-prem capacity replaces per-token cloud billing for steady workloads. Data never leaves the site, satisfying residency, sovereignty and air-gap constraints. Multi-user sharing with workload isolation lets a whole team justify one system. Red Hat AI Factory and RHEL certification give a supported path from evaluation to production.

Applications

On-premises LLM and VLM inference; local agentic coding assistants and developer sandboxes; regulated-industry AI in healthcare, finance and government; sovereign and air-gapped model deployment; manufacturing computer-vision inference near production lines; retail and branch edge inference; research fine-tuning and model evaluation; data-science teams replacing cloud GPU instances.

Related Products