--- title: "HP ZGX Fury Review: Can One GB300 Box Serve a Whole Team?" url: "https://goremy.ai/blog/articles/hp-zgx-fury-gb300-local-ai-workstation-review" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-31T21:20:09.244Z" updated: "2026-10-31T21:20:09.244Z" topic: Local AI Hardware Guides description: "HP's ZGX Fury puts Nvidia's GB300 under a desk with 748GB of memory and 7.1 TB/s HBM. Here are the specs, ZRT software and measured multi-user throughput." --- # HP ZGX Fury Review: Can One GB300 Box Serve a Whole Team? > HP's ZGX Fury puts Nvidia's GB300 under a desk with 748GB of memory and 7.1 TB/s HBM. Here are the specs, ZRT software and measured multi-user throughput. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-10-31 ## What does the HP ZGX Fury offer over a DGX Spark? The HP ZGX Fury (officially the ZGX Fury G1 AI Station) is HP's version of Nvidia's DGX Station. It pairs a Blackwell GB300 with a 72-core Grace CPU, 748GB of shared memory and 20 petaflops of FP4 compute. A DGX Spark is a desk-top box for one person running smaller models. The Fury is built to sit beside the desk, or in a rack, and serve models to a whole team at once. The figures below come from a hands-on test by a creator who used a loaner unit HP supplied and sponsored. Treat the throughput numbers as one person's measurements, not a lab benchmark. ## TL;DR - The Fury combines **748GB of shared CPU and GPU memory**, of which 252GB is HBM running at 7.1 TB/s. - In the creator's tests, aggregate output passed **3,000 tokens per second** across many parallel streams. - With **64 simulated concurrent users**, each stream still produced roughly high 80s to 90 tokens per second once generating. - A quantization of Nvidia's Nemotron 3 Super tuned by Infinity AI reportedly **doubles concurrent users and total throughput** on the same hardware. - HP ships its own **ZRT command line runtime** and claims the first GB300 station certified for Red Hat Enterprise Linux. - Pricing was not given, and the source itself says the box is out of reach for most individuals. ## What are the ZGX Fury's core specs? | Spec | ZGX Fury | |---|---| | Accelerator | Nvidia Blackwell GB300 (Blackwell Ultra) | | CPU | 72-core Grace | | AI compute | 20 petaflops at FP4 | | Total memory | 748GB shared between CPU and GPU | | HBM portion | 252GB at 7.1 TB/s | | Networking | Two 400 Gbit/s ports | | Cooling | Liquid, tower or rack mount | | OS | Ubuntu by default, RHEL certified | The two 400 Gbit/s ports let you link two units when one is not enough, the same idea as chaining DGX Sparks. ## Why does the memory design matter? Local AI hardware usually forces a choice. Unified-memory machines like the DGX Spark give you a lot of capacity but slower bandwidth. A discrete card like the RTX Pro 6000 is faster but caps out around 96GB per card, which rules out many large open models. The Fury avoids that trade. The 748GB pool is big enough for very large open-weight models, and the 252GB HBM tier runs at 7.1 TB/s. The creator's comparison point is the RTX Pro 6000 at 1.8 TB/s, which Nvidia sells as its fastest professional workstation card. That puts the Fury closer to multi-GPU server territory than to a typical workstation. Memory bandwidth matters most for serving. Token generation is largely bound by how fast weights move, and many users hitting one model at once multiplies that pressure. ## How does it perform serving a whole team? The tests ran many streams in parallel against a Qwen model (named in the demo as Qwen 3.8 Flash) and measured what came out. - **Aggregate throughput:** over 3,000 tokens per second across concurrent streams. - **RAG test, 16 users:** well over 1,300 tokens per second, with the machine's fans not yet spinning up. - **64 simulated users:** the box kept going, producing more than 500,000 tokens in a few minutes. - **Per-user speed at 64 users:** around the high 80s to 90 tokens per second once generation started. That last number is the useful one for a team. Aggregate throughput can look great while each person waits. Around 90 tokens per second per stream is comfortably faster than reading speed, even with 64 people sharing the model. Caveats apply. The prompts, context lengths and output lengths of the load tests were not fully specified. "Simulated users" are scripts, not people who pause, rewrite and send long contexts. Run your own workload before sizing a purchase. ## What is ZRT and why does HP ship it? HP says the Fury is the only DGX Station allowed to ship with customized software, and ZRT (HP ZGX Runtime) is the main piece. It is a command line tool that handles the tedious parts of getting a model online. ``` zrt setup zrt serve zrt chat ``` According to the demo, `zrt serve` pulls the model from Hugging Face, configures it for vLLM, loads it into memory and gives you an endpoint your team can use immediately. Other functions include: - Checking what is loaded on the GPU. - Monitoring usage while a model runs. - Benchmarking against a dataset you choose, reporting time to first token, tokens per second and behavior under concurrent load. - Opening an interactive chat to confirm the model responds sensibly. HP also offers a toolkit extension for VS Code, so developers can work against the box from their editor. ## Does the Infinity AI Nemotron quantization really double capacity? HP partnered with Infinity AI on a quantization of Nvidia's Nemotron 3 Super tuned for the GB300 inside the Fury. Infinity AI used its automated research system to do the tuning. The claim: same hardware, but double the number of concurrent users and double the total token throughput, without slowing individual responses. If that holds in your workload, it matters more than most spec-sheet items. Twice the users on one box is effectively a second machine. The source does not publish the underlying quality evaluations, so check output quality on your own tasks before trusting any aggressive quantization. ## Why does Red Hat Enterprise Linux certification matter? The Fury ships with Ubuntu, like the DGX Spark and most AI machines. HP also says it is the first GB300 station certified for Red Hat Enterprise Linux. The creator points out that many data-center training clusters run RHEL. For a team, this means the workstation can match the production OS. Prototype an agent or fine-tune a model on the Fury, and moving to a data center or cloud needs less rewriting. The same logic applies to the hardware: GB300 is the architecture Nvidia uses in its GB300 NVL72 racks. ## Who is the Fury for, and is it worth it? The pitch is a shift from one box per person to one box per team. A single Fury can handle several workloads together: - Multiple people running inference against a large model. - Background agents on a separate model. - Someone fine-tuning. - Embedding, image and video models alongside the language models. - Tools and sandboxed jobs on the Grace CPU's many cores. An orchestrator model with sub-agent models and sandboxed tooling on one machine is the scenario the creator highlights. Two other arguments sit beside raw speed. There is no per-token API bill, and data stays on premises, which matters in healthcare, government and manufacturing. Control matters too: a security team doing malware analysis or red teaming can pick its own model and guardrails instead of relying on a hosted provider's policy, which can change without notice. Is it worth it? No price appeared in the source, so there is no cost-per-token math to run. The creator says it is out of budget for most individuals. It makes sense for a team with steady, high-volume usage, privacy requirements, or a need to prototype on data-center-class hardware. For one developer running small models, a Spark-class machine is the better fit. ## Frequently Asked Questions ### How much memory does the HP ZGX Fury have? It has 748GB shared between the CPU and GPU. Of that, 252GB is HBM with 7.1 TB/s of bandwidth. The rest is unified memory that extends capacity for very large models. ### Can the ZGX Fury be rack mounted? Yes. It is liquid cooled and works as a tower beside a desk, but HP also built it for rack mounting with airflow optimized for that setup. The source says this separates it from other DGX Stations. ### How many users can it serve at once? In the creator's tests, it handled 64 simulated concurrent users on a Qwen model, with per-stream speeds around the high 80s to 90 tokens per second. Aggregate output exceeded 3,000 tokens per second. Real capacity depends on model, context length and workload. ### Can I link multiple Fury units? Yes. It has two 400 Gbit/s network ports, and you can connect two units together when one box does not provide enough capacity, similar to linking DGX Sparks. ### What is ZRT? ZRT is HP's command line runtime for the Fury. It sets up the machine, downloads and serves models from Hugging Face, runs benchmarks, shows usage and opens test chats.