Read the 748GB on HP's new ZGX Fury G1n as two numbers, not one. According to TechRadar, it's 496GB of system memory plus 252GB on the GPU. The two are different kinds of RAM doing different jobs, and they sit alongside Nvidia's GB300 Grace Blackwell Ultra processor.
That split matters most for the machine's headline AI claims. Prices start at US$132,000, and TechRadar says that's before you add any options.
We couldn't check HP's own announcement directly. Everything below comes from TechRadar's report, which credits HotHardware.
The split behind the 748GB
TechRadar describes the ZGX Fury as bringing "computing hardware previously associated with large-scale AI infrastructure into an office-sized machine." Here's how it says the memory breaks down:
- 496GB of LPDDR5X system memory running at 6,400 MT/s, connected to a 72-core Arm Neoverse V2 Grace CPU, supporting "demanding processing workloads"
- 252GB of HBM3e built into Nvidia's Blackwell Ultra GPU, providing "dedicated high-bandwidth capacity for graphics and AI operations"
TechRadar says the two together make up "the advertised 748GB capacity." It also says buyers who add extra graphics hardware can raise available video memory to 276GB in selected configurations. That's 24GB above the 252GB baseline. Separately, TechRadar says HP offers configurations with discrete Nvidia RTX PRO workstation cards, including the RTX PRO 2000 and RTX PRO 4000. The report doesn't say which card or configuration gets you to 276GB.
One more thing to watch: TechRadar's copy calls it a "mobile workstation" at one point. Nothing else in the report fits that label. This is a liquid-cooled box with a 1,600W power supply that can go in a rack.
What it's for (hint: not your Steam library)
TechRadar says the memory isn't there for office apps or even demanding professional software. It's there for large language models. TechRadar reports HP saying the memory can handle local adjustment of models with up to 100 billion parameters. TechRadar adds, without crediting HP, that the system can run inference on models of up to one trillion parameters using FP4 quantisation, though actual performance depends on the workload.
The pitch, per TechRadar, is that developers and researchers can handle big workloads "without depending entirely on remote computing infrastructure." Mind the word "entirely". The pitch is about leaning less on remote hardware, not ditching it.
The tells that this isn't a normal desktop
- It ships with Ubuntu 24.04 LTS, not Windows. TechRadar says the Linux setup caters to developers and technical professionals.
- It's liquid-cooled and has a 1,600W Titanium-certified power supply. TechRadar says ordinary desktop cooling can't necessarily handle this much hardware in a compact case.
- The networking is built for linking machines. TechRadar says an Nvidia ConnectX-8 SuperNIC provides two QSFP112 connections rated at 400Gbps, so multiple machines can talk over high-speed networking.
- It fits in a rack. According to TechRadar, the enclosure can go in a 5U equipment rack. Organisations can run one under a desk or install several in racks.
The price ladder
Per TechRadar:
| Configuration | Price |
|---|---|
| Standard | US$132,000 |
| With RTX PRO 2000 | US$133,000 |
| With RTX PRO 4000 | US$136,000 |
TechRadar's headline teases a desktop for the price of a Ferrari. Its own copy then admits the base price buys a brand-new 2026 Cadillac Escalade IQL, "although not quite a Ferrari." TechRadar's report doesn't mention Australian pricing, availability or a ship date.
TechRadar says the ZGX Fury is "unlikely to attract ordinary desktop buyers." It says the likely buyers are organisations that need substantial local computing power and can justify spending well above normal workstation prices.
Jax's take: The advertised 748GB total adds up (496 + 252), but I think quoting it as one figure flatters the machine for AI shoppers. Run the numbers on the trillion-parameter claim. FP4 stores each parameter in four bits, or half a byte, so the weights alone come to roughly 500GB. That's about double the 252GB of HBM3e, so a model that size can't sit entirely in the fast GPU memory. TechRadar's report doesn't say how HP's 100-billion and one-trillion-parameter figures map across the two pools. Before I signed off on a US$132,000 purchase, I'd want HP to show exactly how much of each model sits in HBM3e and how much spills into system memory.
Reporting notes



