HP Inc. (NYSE: HPQ) has publicly demonstrated Windows running on its ZGX Fury workstation with NVIDIA’s (Nasdaq: NVDA) GB300 Grace Blackwell Ultra Desktop Superchip. HP is positioning the system for developers, data scientists, researchers and enterprise teams that need to develop, test and deploy advanced AI workloads on premises, closer to their data, applications and users.
The demonstration establishes a Windows option for a class of workstation hardware designed to support demanding AI work. HP said the ZGX Fury is intended for inference workloads, concurrent users and AI calls, and experimentation with larger or more complex models. The company did not disclose final system configurations, pricing, general availability or supported model sizes for the product.
Local AI capacity can be important for organizations that want more direct control over where models and data run. A workstation placed near an engineering team can reduce reliance on remote cloud capacity for selected development or inference tasks, but it also places responsibility for endpoint security, patching, model governance, data access and physical infrastructure with the organization operating it.
HP’s announcement spans both enterprise workstations and consumer systems. The company said it will offer RTX Spark-based OmniBook devices starting Oct. 16, while the ZGX Fury is the portion aimed at enterprise AI work. It also said it demonstrated the workstation at a Microsoft event, but the announcement provides no timeline for commercial availability of the ZGX Fury and directs prospective buyers to local sales representatives.
The distinction matters because a local AI PC and a workstation that supports higher-concurrency development or inference are not interchangeable. HP described the ZGX Fury as a system for organizations building and deploying sophisticated AI applications. It said the system would allow teams to run workloads closer to users, data and workflows while retaining greater control over their AI infrastructure.
The company also described its RTX Spark offerings as running larger models locally without cloud dependency. Those claims are product descriptions, not published benchmarks. Actual capacity will depend on the model architecture, parameter count, quantization, memory footprint, data pipeline and the other applications running on the system.
Microsoft (Nasdaq: MSFT) and NVIDIA were involved in the demonstration, according to HP. The company also said its Windows-based approach places AI work on the platform where many users, applications and workflows already reside. For IT teams, that can reduce some adoption friction, but it does not eliminate the need to decide how a high-performance workstation is enrolled, segmented, monitored and connected to enterprise data sources.
HP’s update reflects a growing deployment option between ordinary client devices and centralized AI clusters. A GB300-based workstation may give teams a local environment for development and limited deployment without first provisioning a data center service. Whether it becomes a practical enterprise platform will depend on final specifications, management tooling, security controls and availability—details that HP has not yet disclosed.
On-premises AI systems can also be useful where data-residency rules, latency targets or disconnected operations limit a cloud-first approach. Those requirements do not automatically make a workstation the right architecture, however. Teams must still size the system for demand, plan backups and recovery, and ensure that access controls cover both models and the data used to ground them.