HP ZGX Fury Review: A Shared AI Workstation, Not a Desktop GPU

HP's ZGX Fury G1 AI Station, built on Nvidia's GB300 Blackwell Ultra chip and a 72-core Grace CPU, delivers 20 petaflops of FP4 AI compute with 748GB of unified memory — 252GB of it HBM at 7.1 TB/s — to serve large open models to an entire team on-premises. In testing the workstation sustained over 3,000 tokens per second in aggregate and handled 64 simultaneous users, and HP's ZRT command-line toolkit automates pulling models from Hugging Face, configuring them for vLLM serving, and benchmarking throughput. The Fury is the first GB300 station certified for Red Hat Enterprise Linux, and a partnership with Infinity AI produced a custom quantization of Nvidia's Nemotron Super model that roughly doubled concurrent users and token throughput without extra hardware.

HP ZGX Fury Review: A Shared AI Workstation, Not a Desktop GPU HP's ZGX Fury pairs Nvidia's GB300 Blackwell Ultra chip with 748GB unified memory to serve large open models to an entire team on-prem. What is the HP ZGX Fury? The HP ZGX Fury G1 AI Station is HP’s version of Nvidia’s DGX Station reference design, built around Nvidia’s GB300 chip Blackwell Ultra paired with a 72-core Grace CPU. It delivers 20 petaflops of AI compute at FP4 and 748GB of memory shared between CPU and GPU. Unlike a personal AI box such as the DGX Spark or ZGX Nano, the Fury is designed to sit next to a desk or in a server rack and serve inference to an entire team at once, not just one person. TL;DR - The GB300 chip at the heart of the Fury is the same Blackwell Ultra architecture Nvidia ships in its data center NVL72 racks, not a scaled-down desktop part. - Total memory hits 748GB , including 252GB of HBM running at 7.1 TB/s, far beyond the 1.8 TB/s bandwidth on Nvidia’s RTX Pro 6000 workstation card. - HP ships its own ZRT command-line toolkit that automates pulling models from Hugging Face, configuring them for serving with vLLM, and benchmarking throughput and concurrency. - The Fury is the first GB300 station certified for Red Hat Enterprise Linux , the OS most commonly used in production data centers and training clusters. - A partnership with Infinity AI produced a custom quantization of Nvidia’s Nemotron Super model tuned for this specific hardware, roughly doubling concurrent users and token throughput without extra hardware. - In testing, the machine sustained over 3,000 tokens per second in aggregate across multiple concurrent streams and handled 64 simultaneous users while still returning strong per-user token speeds. - Two units can be linked over 400 Gbit/s networking ports for teams that outgrow a single box, mirroring how DGX Spark units can be clustered. Rapidly build enterprise-grade software applications. How does the hardware actually work? The Fury’s standout design choice is refusing to force a tradeoff between memory capacity and memory speed. Most local AI setups fall into one of two camps: unified memory systems like the DGX Spark that offer a lot of RAM but slower access, or discrete GPU cards like the RTX Pro 6000 that are fast but capped around 96GB per card, a serious limit for today’s largest open models. The ZGX Fury gets both. Of its 748GB total memory, 252GB is high-bandwidth memory HBM running at 7.1 TB/s, nearly four times the bandwidth of Nvidia’s fastest professional workstation card. The rest is unified memory shared between the Grace CPU and GB300 GPU, giving the system room to load very large models while keeping the active weights moving fast. Two 400 Gbit/s networking ports allow multiple units to be linked together, and the whole system is liquid cooled, built to run either as a standalone tower or mounted in a server rack with rack-optimized airflow. What software does HP add on top? Nvidia’s DGX Station is a reference design, and HP is, according to the video, the only partner allowed to customize what ships on top of it. The centerpiece is ZRT HPZ Runtime , a command-line tool meant to strip out the usual friction of getting a model running and tuned for serving. A few examples of what ZRT handles: - zrt setup and zrt serve <model pull a model directly from Hugging Face, configure it for vLLM, load it into memory, and expose an endpoint the whole team can hit. - zrt chat opens an interactive session to sanity-check that a model is configured and responding correctly. - Built-in benchmarking reports time-to-first-token, tokens per second, and how performance holds up under concurrent load. For developers, HP also offers an HPZ Toolkit extension that brings this functionality into VS Code directly. On the OS side, the Fury ships with Ubuntu by default but is the first GB300-based station certified for Red Hat Enterprise Linux RHEL , the OS commonly used in enterprise data centers and training clusters, which matters for teams trying to match their on-prem workstation environment to production infrastructure. Why does model-level optimization matter here? Hardware specs only tell part of the story. HP worked with Infinity AI to apply a custom quantization of Nvidia’s Nemotron Super model, tuned specifically for the GB300 chip inside the Fury using Infinity AI’s auto research system. The result, per the demonstration: the same hardware serving roughly double the concurrent users and double the aggregate token throughput, without slowing down any individual user’s responses. pass the audit. PASS PASS PASS PASS PASS PASS 0 That distinction matters because the Fury isn’t pitched as a single-user inference box. Doubling concurrent capacity on the same hardware is, in practical terms, close to getting a second machine without buying one. In testing shown in the video, the system sustained over 3,000 tokens per second in aggregate across multiple simultaneous streams, and in a RAG-style test with 16 concurrent users it held north of 1,300 tokens per second, with fan speed not even ramping up. Pushed to 64 simulated concurrent users on Qwen3’s 30B-class flash model, the machine still produced usable per-user token speeds and generated over 500,000 tokens in a few minutes. Who is the ZGX Fury actually for? The pitch is a shift away from “one box per person.” DGX Spark-class machines assume each team member gets their own desktop unit for running smaller models locally. The Fury flips that model: one shared machine serves a whole team concurrently. In practice that could mean several people running inference against a large model, a set of background AI agents running on a different model, someone fine-tuning in parallel, and another person testing a new model, all on the same hardware at the same time, with the Grace CPU’s high core count handling tool execution and sandboxed environments alongside the GPU-bound inference work. This target use case fits teams in regulated or sensitive industries, healthcare, government, manufacturing, cybersecurity, where sending data to a third-party API is a non-starter, or where control over model guardrails matters. A security team doing malware analysis or writing exploit proofs-of-concept, for instance, can’t rely on a closed model that might refuse the task or have its guardrails changed without notice. Running the model on owned hardware removes both the data exposure risk and the dependency on an external provider’s policy decisions. Is the ZGX Fury worth it? For an individual buyer, no, this is explicitly enterprise-tier hardware and priced and positioned accordingly. But for a team that wants serious on-prem inference capacity without a recurring per-token API bill, the value proposition is more straightforward: pool the hardware cost across everyone who uses it rather than buying a GPU card per person. The bigger architectural argument is consistency with production infrastructure. The GB300 chip inside the Fury is the same Blackwell Ultra silicon Nvidia deploys in its GB300 NVL72 data center racks, not a cut-down variant. Combined with HBM, high-speed networking, liquid cooling, and RHEL certification, a team prototyping agents or fine-tuning models on a Fury is working against the same software and hardware stack they’d scale to in a cloud or data center deployment later, reducing the rework typically needed when moving from a desktop prototype to production infrastructure. Frequently Asked Questions What chip does the HP ZGX Fury use? It uses Nvidia’s GB300, built on the Blackwell Ultra architecture, paired with a 72-core Nvidia Grace CPU, delivering 20 petaflops of AI compute at FP4 precision. How much memory does the ZGX Fury have? 748GB total, shared between the CPU and GPU. Of that, 252GB is high-bandwidth memory HBM running at 7.1 TB/s. How is the ZGX Fury different from a standard Nvidia DGX Station? HP customizes the software stack on top of Nvidia’s reference hardware design, adding its ZRT command-line toolkit, an HPZ Toolkit extension for VS Code, Red Hat Enterprise Linux certification, and a custom Infinity AI quantization of Nemotron Super tuned for the GB300 chip. Can multiple people use one ZGX Fury at the same time? 1. 1 Secure 2. 2 Regulation-compliant 3. 3 Auditable 4. 4 Scalable 1 SecureAdjusters can't export patient data✕ Export blocked 2 CompliantSSN •••-••-4821 ✓ HIPAA ✓ SOC 2 3 Auditable09:14 · M. Chen approved09:12 · J. Ruiz edited09:07 · A. Patel viewed 4 Scalable12,480users online Yes. It’s designed as a shared team resource rather than a personal workstation, and testing showed it sustaining strong throughput with dozens of concurrent users and multiple models or agents running simultaneously. Can you connect multiple ZGX Fury units together? Yes. Each unit has two 400 Gbit/s networking ports, allowing two or more Furys to be linked for teams that need more capacity than a single unit provides.