HP ZGX Nano G1n AI Station: The Complete Guide (Singapore)

The HP ZGX Nano G1n is HP's entry into a new category of machine: the desk-side AI supercomputer. Built around the NVIDIA GB10 Grace Blackwell superchip, it packs roughly 1,000 TOPS of FP4 AI performance and 128GB of unified memory into a 15cm square chassis — enough to develop, fine-tune and run large language models of up to ~200 billion parameters entirely on your desk, with no data leaving the building. This guide covers what makes the ZGX Nano different from other GB10 machines, exactly which models it supports, who it's for, how to set it up, and how to deploy it in Singapore.

What the ZGX Nano is

Think of it as a personal DGX. Inside the compact enclosure (150 × 150 × 51mm) sits the GB10 superchip: a Blackwell-generation GPU with 5th-gen Tensor Cores paired with a 20-core Arm CPU (10 Cortex-X925 performance cores plus 10 Cortex-A725), sharing 128GB of LPDDR5x coherent unified memory at 273GB/s. Because CPU and GPU share one memory pool, models that would overflow a conventional GPU's VRAM run comfortably. It ships with NVIDIA DGX OS (configured as HP ZGX OS, Ubuntu-based) and the full NVIDIA AI software stack, and is available with a 4TB NVMe SSD from SourceIT — see the HP ZGX Nano G1n 4TB product page for current pricing.

What makes it different from other GB10 machines

Every GB10 machine — NVIDIA's DGX Spark, the ASUS Ascent GX10, Lenovo's ThinkStation PGX and others — shares the same core silicon, so raw AI performance is essentially identical across brands. HP differentiates on three fronts that matter to organizations more than benchmarks.

1. Enterprise firmware security

The ZGX Nano carries HP's workstation security DNA, including HP Sure Start — self-healing firmware that validates the BIOS against a known-good copy on every boot. For regulated industries, that's assurance most consumer-grade AI boxes simply don't offer. HP also offers a Secure & Regulated Environments variant with no Wi-Fi or Bluetooth radios at all, designed for air-gapped deployments — highly relevant to Singapore's government, defence, healthcare and financial sectors, where sensitive data must never touch the cloud. If your AI use case exists precisely because the data cannot leave your premises, this is the GB10 machine built around that requirement.

2. The ZGX Toolkit developer experience

HP ships an exclusive ZGX Toolkit — a Visual Studio Code extension that lets developers configure, manage and deploy to the ZGX Nano from their everyday laptop, treating the unit as a desk-side AI server. The machine comes with a pre-configured environment — Jupyter, VS Code integration, Docker and MLflow work out of the box — and HP claims up to 45% faster time-to-results versus a manual DIY setup. In practice, your team writes code where they always have, and the Nano does the heavy lifting.

3. Sustainable, serviceable hardware

The chassis is built from up to 75% recycled aluminium and 20% recycled steel — useful for organizations with ESG procurement criteria — and the unit draws power over a single 240W USB-C connection. Rear I/O covers three USB-C ports (20Gbps), 10GbE RJ-45, and dual QSFP ports at up to 200Gbps via NVIDIA ConnectX-7.

Which models does it support?

The GB10 platform runs the mainstream open-weight ecosystem — Meta (Llama), OpenAI (gpt-oss), DeepSeek, Alibaba (Qwen), Mistral, Google (Gemma), NVIDIA (Nemotron) and more — through the tools your team already uses: Ollama, llama.cpp, vLLM, PyTorch and NVIDIA NIM microservices. What fits, and how it feels, depends on model size.

Inference: what runs on one unit

Fast and interactive (7B–32B): models like Llama 3.2, GPT-OSS 20B, DeepSeek-R1 distills (14B/32B) and Gemma run with snappy, chat-grade responsiveness — ideal for coding assistants, RAG over internal documents and agent prototypes.

Daily-driver class (≈70B): Llama 3.1/3.3 70B, DeepSeek-R1 70B and Qwen-class 70–80B models run comfortably in quantized form — this is the sweet spot where output quality meets usable speed, and the tier a conventional GPU workstation can't reach without multiple cards.

Prototyping ceiling (100–200B): large mixture-of-experts models such as GPT-OSS 120B and Nemotron Super 120B fit within the 128GB unified memory using FP4/quantized weights. Expect research-and-evaluation speeds rather than production throughput — the point is that they run at all, locally.

Dual-unit cluster (up to ~405B): linking two units over the QSFP ports pools 256GB of memory, extending reach to Llama 3.1 405B-class and Qwen3-235B-class models. We stock the matching HP ZGX QSFP112 DAC cable.

Fine-tuning: realistic expectations

Fine-tuning support follows NVIDIA's published guidance for the platform. Full fine-tuning (all parameters) is practical up to ~8B models — e.g. Llama 3.1 8B. LoRA handles 8B-class models at high speed, and QLoRA (4-bit) extends parameter-efficient fine-tuning to 70B-class models on a single unit. Supported frameworks include LLaMA Factory (CLI/WebUI for SFT, LoRA, QLoRA and RLHF), native PyTorch, and NVIDIA NeMo AutoModel for enterprise workflows. Larger-scale training belongs on DGX-class servers or cloud — the Nano's role is developing and fine-tuning on the same CUDA stack you'll deploy to production.

Beyond text

The platform also runs image and video generation models — FLUX.1/FLUX.2, Qwen-Image and LTX-class audio-video models — making it a capable local node for multimodal experimentation, not just LLMs.

Setting up your ZGX Nano: first boot in five steps

One of the Nano's underrated strengths is that setup is closer to a network appliance than a Linux build. No monitor or keyboard is required — it's designed for headless setup from your laptop.

Step 1 — Power on. Connect the power cord and switch the unit on. For the fastest experience later, also connect the 10GbE RJ-45 port to your network; otherwise Wi-Fi can be configured during setup.

Step 2 — Join the setup hotspot. From your laptop, connect to the ZGX's own Wi-Fi hotspot. The hotspot name and credentials are printed on the system information label on the unit itself.

Step 3 — Open the setup page. In a browser, go to the device URL in the format zgx-00abcd.local (the exact address is on the same label). Select Get Started.

Step 4 — Run the wizard. Choose your language, accept the terms, and create your Linux username and password. This is the account you'll use for SSH, Jupyter and the dashboard from here on.

Step 5 — Connect your dev environment. From the Dashboard UI, follow the quick start link to install the HP ZGX Toolkit — or go straight to VS Code on your laptop: open Extensions, search "zgx", install the HP ZGX Toolkit extension, and follow the prompts to add and connect your Nano. From there you can launch Jupyter, manage containers with Docker, track experiments with MLflow and deploy your first model — all from your own machine.

Budget 15–30 minutes from unboxing to running your first model. For fleet or secure-environment deployments (static IPs, proxy environments, air-gapped variants without the hotspot flow), our engineers handle commissioning as part of deployment — see below. HP's full documentation lives at hp.com/zgx-onboard and the HP support site's setup and user guides.

ZGX Nano vs NVIDIA DGX Spark

The DGX Spark is NVIDIA's reference design; the ZGX Nano is the same platform wrapped in HP's enterprise hardening. Choose the Spark for the canonical NVIDIA experience; choose the ZGX Nano when firmware security, an air-gapped variant, HP's support organization or the ZGX Toolkit workflow tip the balance. Comparing the wider field — ASUS, MSI, Gigabyte, Lenovo, Dell, Acer — see our complete NVIDIA GB10 buyer's guide.

Who should buy it

The ZGX Nano fits four buyer profiles particularly well in Singapore. Regulated organizations — agencies, hospitals, banks, law firms — that need LLM capability on data that cannot leave their premises. AI teams that want a local development node mirroring their production NVIDIA stack. Universities and polytechnics equipping research labs with per-team compute instead of contested shared clusters. And enterprises with HP-standardized procurement that want AI capability inside their existing support relationship.

Buying and deployment in Singapore

SourceIT supplies the HP ZGX Nano G1n with official Singapore warranty as an authorized HP business reseller. Our specialists advise on configuration, clustering, network integration and secure-environment deployment — including commissioning units so your team skips the setup entirely — and we support formal procurement and tender processes for institutional buyers. GB10-class hardware is subject to supply constraints — for project or volume orders, engage us early. Use Request for Quotation on the product page, call (65) 6978 3502, or email Sales@sourceit.com.sg.

Frequently asked questions

Can it run Llama 70B? Yes — 70B-class models are the platform's sweet spot for daily use in quantized form, and QLoRA fine-tuning of 70B models is supported on a single unit.

Can it run GPT-OSS 120B or larger? Yes — 100–200B-class models fit in 128GB unified memory with FP4/quantized weights at prototyping speeds. Two linked units extend reach to ~405B-class models.

Does the ZGX Nano run Windows? No — it runs NVIDIA DGX OS configured as HP ZGX OS (Ubuntu-based). You access it from any laptop — Windows or Mac — over the network, including via the ZGX Toolkit in VS Code.

Do I need a monitor and keyboard to set it up? No — setup is headless: power on, join the unit's Wi-Fi hotspot from your laptop, open the zgx-*.local setup page and follow the wizard. See the five-step setup section above.

How is it different from a gaming PC with a big GPU? Unified memory. A discrete GPU tops out at its VRAM; the ZGX Nano gives models access to 128GB of coherent memory, letting far larger models run locally — plus enterprise firmware security no gaming build offers.

Is there a fully offline variant? Yes — HP offers a Secure & Regulated Environments configuration with no wireless radios for air-gapped deployments. Ask our team about availability in Singapore.

What warranty applies? Official HP local warranty applies to units supplied by SourceIT, with enterprise support options available — details on request.