HP sells ZGX Fury G1n AI workstation with 748 GB of memory for $132,000
HP has started selling the ZGX Fury G1n, a workstation for AI. It looks like an ordinary desktop tower, but inside is an NVIDIA GB300 platform with 748 GB of memory (496 GB attached to the CPU and 252 GB attached to the accelerator). HP says this is enough to run models of up to 1 trillion parameters in your own office without sending data to the cloud. The base version costs $132,000, so it pays off only under near-constant load.
AI-processed from Habr AI; edited by Hamidun News
HP has started selling the ZGX Fury G1n AI Station, a workstation that can run and fine-tune large AI models locally. It looks like an ordinary desktop tower, but inside is a server platform, the NVIDIA GB300 Grace Blackwell Ultra. The headline cites 748 GB of memory.
That is the sum of 496 GB of standard LPDDR5X memory attached to the CPU and 252 GB of fast HBM3e memory used by the GPU accelerator. For comparison, a top gaming graphics card typically has 24–32 GB, so this machine has more than twenty times as much memory. The base version costs $132,000 in the US.
Picture a clinic or a bank that needs an internal assistant working on confidential documents: they cannot send those documents to the cloud, and a large model does not fit on an ordinary computer. The station is built for exactly that case. But buying it makes sense only if it will be busy almost all the time.
What's inside
The core of the machine is a 72-core Grace CPU based on the Arm Neoverse V2 architecture and a Blackwell Ultra accelerator. FP4 compute performance reaches 20 PFLOPS. FP4 is a format in which each model parameter takes up just four bits. It is what makes it possible to fit huge networks into limited memory.
The rest of the hardware is also server-grade: up to 32 TB of total storage, three PCIe 5.0 slots, two 400 Gbit/s QSFP112 network interfaces, a 1.6 kW power supply and liquid cooling. The stated software is Ubuntu 24.04 LTS.
What you can do with it
According to HP, a single station can run models of up to 1 trillion parameters locally, and it also supports local fine-tuning for models of up to 100 billion parameters. A back-of-the-envelope calculation shows the first claim is plausible only in compressed form: 1 trillion parameters in a 4-bit format take about 500 GB and fit into 748 GB with room left for working data. Fine-tuning a 100 billion-parameter model in a higher-precision format takes on the order of 200 GB, with the rest going to training overhead. This is our estimate, not HP's figure, and the actual headroom depends on the fine-tuning method and context length.
You can optionally add an NVIDIA RTX PRO graphics card to the built-in accelerator: the RTX PRO 2000 Blackwell with 16 GB of GDDR7 or the RTX PRO 4000 Blackwell SFF with 24 GB. Two stations can be connected by a high-speed link and used together. A rack-mount version also exists: in a server rack, the ZGX Fury takes 5U.
Prices are as follows: the base model is $132,000, with an RTX PRO 2000 it is $133,000, and with an RTX PRO 4000 it is $136,000. The graphics cards add only $1,000–4,000, and most of the money goes to the GB300 platform.
Who needs it and where the catch is
The station makes sense when three conditions are met: the data must not leave the company, the model is needed constantly rather than occasionally, and the team needs to experiment quickly without queuing for cloud GPUs. Typical examples are a bank with a closed document archive, a studio fine-tuning a model on its own material, or a startup with a local assistant for developers.
There are plenty of catches too. First, $132,000 is a one-time payment that pays off only under high utilization. If the station sits idle, the cloud will be cheaper. Second, 20 PFLOPS and "up to 1 trillion parameters" are manufacturer claims, and there are no independent measurements of generation speed yet. Running a giant model and getting comfortable response speeds are not the same thing: there is enough memory, but per-token speed depends on the specific model and quantization. Third, the platform is built on Arm, so you need to check that your stack and libraries build for this architecture.
The practical takeaway is simple. Before buying, decide which model you will actually run, in what format and for how many hours a day. If the answer is "a large open model, almost around the clock, on confidential data", a station of this class becomes a reasonable alternative to renting in the cloud. If not, first calculate the cost of cloud hours.
- Nov 21, 2026Skein Agent Erases Finished Stages From Context: 97% Success Rate, But More Expensive Than OpenCode
- Nov 21, 2026SGLang with Humming Kernels Cut Time to First Token by 7–23% and Output Throughput by 17–34%
- Nov 21, 2026Instead of Cursor: Six Open-Source Coding Agents, OpenCode Has 228K Stars on GitHub
- Nov 21, 2026ClawDOS Writes and Runs Code on an IBM PC 5160: Primes Under 100 in 64 Seconds
Key Questions Answered
What's inside?
The core of the machine is a 72-core Grace CPU based on the Arm Neoverse V2 architecture and a Blackwell Ultra accelerator. FP4 compute performance reaches 20 PFLOPS. FP4 is a format in which each model parameter takes up just four bits. It is what makes it possible to fit huge networks into limited memory. The rest of the hardware is also server-grade: up to 32 TB of total storage, three PCIe 5.0 slots, two 400 Gbit/s QSFP112 network interfaces, a 1.6 kW power supply and liquid cooling.
What you can do with it?
According to HP, a single station can run models of up to 1 trillion parameters locally, and it also supports local fine-tuning for models of up to 100 billion parameters. A back-of-the-envelope calculation shows the first claim is plausible only in compressed form: 1 trillion parameters in a 4-bit format take about 500 GB and fit into 748 GB with room left for working data.
Who needs it and where the catch is?
The station makes sense when three conditions are met: the data must not leave the company, the model is needed constantly rather than occasionally, and the team needs to experiment quickly without queuing for cloud GPUs. Typical examples are a bank with a closed document archive, a studio fine-tuning a model on its own material, or a startup with a local assistant for developers. There are plenty of catches too. First, $132,000 is a one-time payment that pays off only under high utilization.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.