 |
| The ZGX Fury AI Station comes in a single colour option. |
HP has begun selling its ZGX Fury G1n AI Station in the United States, introducing a workstation built around NVIDIA's GB300 Grace Blackwell Ultra platform. With 748GB of combined CPU and GPU memory, a 72-core Arm processor, and configurations offering up to 276GB of total GPU memory, the system is designed for organizations running demanding artificial intelligence workloads locally. Its $132,000 starting price also makes clear that this is a specialized enterprise machine rather than a conventional high-end desktop PC.
The workstation targets AI developers, research institutions, and businesses that need substantial computing resources without relying entirely on cloud infrastructure. HP positions the ZGX Fury as an on-premises platform for large language model inference, fine-tuning, and AI agent development. Alongside the GB300 processor, buyers can select professional NVIDIA RTX graphics cards, expanded solid-state storage, and configurations intended to support multiple users working on the same system.
NVIDIA GB300 Grace Blackwell Ultra Powers the HP ZGX Fury
At the center of the ZGX Fury G1n is NVIDIA's GB300 Grace Blackwell Ultra Superchip, which combines a 72-core Arm Neoverse V2 CPU with a Blackwell Ultra GPU. The architecture is designed to handle workloads involving large AI models and substantial memory requirements. Unlike traditional desktop configurations that rely on a separate x86 processor and conventional graphics card, the GB300 platform brings CPU and GPU resources together through NVIDIA's high-bandwidth computing architecture.
The system includes 496GB of LPDDR5X-6400 CPU memory and 252GB of HBM3e GPU memory, providing 748GB of combined memory capacity. NVIDIA's HBM3e implementation delivers up to 7.1TB/s of GPU memory bandwidth, an important characteristic for workloads that repeatedly move large quantities of model data through the accelerator. This memory capacity can help accommodate models that would otherwise require multiple conventional graphics cards, although actual performance depends on model architecture, numerical precision, and software optimization.
Why 748GB of Combined Memory Matters for AI
Memory capacity is one of the main limitations when deploying large language models on local hardware. A model containing hundreds of billions of parameters can require enormous amounts of memory, particularly when running at higher numerical precision or serving multiple users. HP states that the ZGX Fury platform can host models with up to one trillion parameters locally under supported configurations and operating conditions. That figure should be understood as a platform capability rather than a guarantee that every trillion-parameter model will run at full precision or achieve a particular inference speed.
It is also important to distinguish the workstation's 748GB combined memory capacity from its dedicated GPU memory. The GB300 includes 252GB of HBM3e memory associated with its Blackwell Ultra accelerator, while the remaining 496GB is LPDDR5X memory on the CPU side. These resources are connected through NVIDIA's architecture, but they do not have identical bandwidth or performance characteristics. Consequently, the advertised combined capacity should not be interpreted as 748GB of conventional graphics VRAM.
Optional NVIDIA RTX PRO Graphics Expand GPU Memory to 276GB
HP offers the ZGX Fury G1n with optional discrete NVIDIA RTX PRO Blackwell graphics cards. Buyers can choose the RTX PRO 2000 Blackwell Generation with 16GB of GDDR7 memory or the RTX PRO 4000 Blackwell SFF Generation with 24GB of GDDR7. When combined with the GB300's 252GB of HBM3e memory, these configurations provide an aggregate of 268GB or 276GB of GPU memory, respectively. However, this total represents memory distributed across separate GPUs rather than a single, automatically pooled VRAM allocation.
The additional RTX PRO hardware provides flexibility for workflows that combine AI computation with professional graphics, visualization, or other GPU-accelerated applications. Organizations considering these configurations should evaluate software compatibility and how their applications distribute workloads across accelerators. The presence of an additional graphics card does not necessarily mean that a single AI model can transparently access the combined memory capacity as one continuous memory pool.
Available HP ZGX Fury G1n Configurations
HP currently lists several versions of the workstation, all based on the same NVIDIA GB300 platform. The primary differences involve the optional professional graphics card and installed storage. Each of the following configurations is listed with a three-year warranty, while the storage arrangements vary according to the selected model.
- Base configuration: The HP ZGX Fury G1n without an additional graphics card includes the GB300 Blackwell Ultra accelerator and four 2TB self-encrypting SSDs.
- RTX PRO 2000 configuration: The HP ZGX Fury G1n with RTX PRO 2000 Blackwell adds a professional GPU with 16GB of GDDR7 memory and four 2TB SSDs.
- RTX PRO 4000 configuration: The HP ZGX Fury G1n with RTX PRO 4000 Blackwell SFF includes 24GB of additional GDDR7 memory, two 2TB SSDs, and two 4TB SSDs, providing 12TB of installed storage.
PCIe 5.0 Storage and High-Speed Networking for Enterprise Workloads
Storage expansion is another important part of the ZGX Fury platform. The workstation supports multiple M.2 SSDs and configurations reaching up to 32TB of storage, giving organizations room for model files, training datasets, application environments, and project archives. PCIe 5.0 connectivity provides a modern foundation for high-performance solid-state storage, although actual transfer speeds will depend on the installed drives, storage configuration, and workload.
The workstation also provides PCIe Gen 5 expansion and QSFP112 connectivity supporting 400Gbps-class links. These capabilities are particularly relevant for AI development environments where large datasets must move between systems or where organizations want to connect multiple workstations. High-speed interconnects can reduce communication bottlenecks in supported multi-system workflows, but the performance of distributed AI applications still depends on the software stack, networking configuration, and how computing tasks are divided.
Ubuntu and HP Z Software Support Local AI Development
HP lists Ubuntu as the operating system for the ZGX Fury G1n and provides software intended to simplify AI development and deployment. The company highlights its HP Z Toolkit for setting up open-source AI tools, alongside a preinstalled HP Z Runtime environment. This approach is designed to help development teams move from hardware installation to running supported AI applications without having to assemble every component of the software environment independently.
HP also promotes the workstation as a shared computing resource for multiple developers or teams. Its office-deployable, rack-mountable design gives organizations an alternative to placing all their AI workloads in a remote data center. Local deployment may be particularly attractive for businesses handling proprietary datasets, confidential documents, or applications with strict data-governance requirements. Nevertheless, operating a workstation of this class still requires appropriate infrastructure planning, including power, cooling, security, maintenance, and software administration.
HP ZGX Fury AI Station Price and Availability
The HP ZGX Fury G1n has reached the US market with a reported starting price of $132,000. The HP ZGX Fury G1n listing at B&H Photo Video confirms a $132,000 price for product E70LKUT#ABA. At the time of the listing, the retailer identifies the workstation as a special-order product with an estimated availability window of six to ten weeks. Delivery estimates may change depending on inventory, configuration, and purchasing arrangements.
Reported HP pricing places the RTX PRO 2000 configuration approximately $1,000 above the starting model, while the RTX PRO 4000 version carries a roughly $4,000 premium. These differences reflect the listed configurations rather than a universal upgrade price, since storage arrangements also vary. HP's online product pages currently display $0.00 in their extracted pricing fields, so buyers should confirm final quotes directly with HP or an authorized reseller instead of treating that placeholder as a valid selling price.
How the ZGX Fury Compares With Conventional AI Workstations
The ZGX Fury occupies a different category from conventional professional desktops equipped with one or two discrete NVIDIA graphics cards. A typical GPU workstation may provide strong performance for smaller models, engineering applications, and content creation, but its local AI capabilities are often constrained by the VRAM available on each accelerator. HP's GB300-based system instead emphasizes a much larger accelerator memory capacity and a tightly integrated CPU-GPU architecture intended for demanding AI inference and model development.
That distinction does not automatically make the ZGX Fury the most economical choice for every AI project. Smaller organizations working with compact language models, quantized models, or occasional inference tasks may find conventional workstations or rented cloud GPUs more practical. The ZGX Fury becomes more relevant when workloads consistently require substantial local memory, when sensitive data must remain on premises, or when predictable access to dedicated compute resources is more important than minimizing the initial hardware investment.
Is the HP ZGX Fury AI Station Worth Considering?
HP's latest AI workstation brings NVIDIA's GB300 Grace Blackwell Ultra technology into a commercially available desktop-class system, combining a 72-core Arm CPU, 252GB of HBM3e GPU memory, and substantial LPDDR5X capacity. Optional RTX PRO graphics, expandable storage, and high-speed networking make it a flexible platform for enterprise AI development. Its strongest selling point is not general-purpose desktop performance, but the ability to run memory-intensive AI workloads on infrastructure controlled by the organization.
With a starting price of $132,000, the ZGX Fury G1n is clearly intended for professional buyers who can justify its capabilities through sustained AI workloads and operational requirements. Organizations should compare the cost of ownership with cloud-based alternatives while considering model compatibility, power requirements, and the benefits of keeping data and compute resources local. For teams that need a dedicated environment for large-scale AI inference and development, the ZGX Fury represents a significant addition to HP's professional workstation portfolio.
You can also check the related product on Amazon for additional details and availability.
