HP ZGX AI Stations — NVIDIA GB300 & GB10
NVIDIA GB300 & GB10 AI Stations — AI supercomputing, from palm-sized to departmental
The HP ZGX family puts NVIDIA Grace Blackwell on your desk: the palm-sized ZGX Nano (GB10, 128GB) for local AI development, and the ZGX Fury (GB300 Grace Blackwell Ultra, 748GB) for production inference at departmental scale. Prototype, fine-tune and run inference locally — then deploy to cloud or data centre when ready. No token fees, no queues, no data leaving your building.
The HP ZGX lineup
Two systems, one mission: bring AI compute on-premises. Start compact with the ZGX Nano for development, scale to the ZGX Fury for departmental production inference — the same NVIDIA software stack runs on both, so models move over cleanly.
- 20-core Arm CPU — 10x Cortex-X925 + 10x Cortex-A725
- Blackwell GPU with native FP4 — up to 1,000 TOPS
- 128GB LPDDR5x unified memory, 273 GB/s bandwidth
- 2TB or 4TB self-encrypted M.2 NVMe SSD
- 2x QSFP 200Gbps — link two units for models up to ~405B params
- 10GbE RJ-45, Wi-Fi 7, Bluetooth 5.4, 3x USB-C 20Gbps, HDMI 2.1a
- Runs NVIDIA DGX OS + HP ZGX Toolkit (MLflow, Ollama & more)
- Pairs with your Windows, Mac or Linux machine over the network
- Secure variant available — no Wi-Fi/Bluetooth, for regulated & air-gapped environments
- 72-core NVIDIA Grace CPU (Arm Neoverse V2) + Blackwell Ultra GPU
- 252GB HBM3e @ 7.1 TB/s + 496GB LPDDR5X @ 396 GB/s
- High-throughput local AI execution surpassing NVIDIA H100
- Serve multiple concurrent users with frontier-class models
- 2x PCIe Gen5 M.2 (OS) + 2x PCIe Gen6 M.2 (data)
- Convertible tower or 5U rack, liquid cooled, 1600W (20A)
- Ubuntu + NVIDIA AI Developer Tools + HP Z Runtime
- Air-gap capable — no Wi-Fi/Bluetooth pathways, wired only
- 400B→70B teacher–student distillation on a single system
A new category: local AI supercomputers
HP Z AI Stations are built for one thing — developing, fine-tuning and running AI on-premises, without cloud dependency.
Speed up model development cycles and build AI agents locally — no usage-based token charges, no cloud provisioning friction, no queue for GPUs.
High-throughput, low-latency local inference with predictable costs — eliminate usage-based pricing and reliance on cloud infrastructure.
Sensitive data stays local; AI runs on-device. Self-encrypted storage, secure variants without wireless, and air-gap capable options.
HP ZGX Toolkit and the NVIDIA AI software stack streamline workflows — device discovery, sync and model export, from any Windows, Mac or Linux client.
What is the NVIDIA GB300?
The NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip combines a 72-core Grace CPU and a Blackwell Ultra GPU on a single superchip, sharing 748GB of coherent memory. It's the silicon behind the new generation of deskside AI supercomputers — and the HP ZGX Fury G1n is HP's flagship GB300 system, built for AI development and departmental-scale production inference, on-prem.
Surpasses NVIDIA H100 throughput
High-throughput local AI execution surpassing the NVIDIA H100 GPU — in a liquid-cooled tower that converts to 5U rack mount.
Frontier-class models, locally
748GB coherent memory enables large-model workflows previously impractical on-prem — including 400B→70B knowledge-distillation (teacher + student on one system).
Private by design
Your data stays local. No Wi-Fi or Bluetooth pathways; connectivity is limited to customer-controlled wired interfaces. Air-gap capable for the most sensitive environments.
ZGX Nano vs ZGX Fury — full comparison
Same Grace Blackwell DNA, different scale. The Nano is your on-ramp to local AI; the Fury is a departmental inference server that happens to fit beside a desk.
| Specification | HP ZGX Nano G1n | HP ZGX Fury G1n |
|---|---|---|
| Superchip | NVIDIA GB10 Grace Blackwell | NVIDIA GB300 Grace Blackwell Ultra |
| Best for | Local experimentation and model development | Departmental-scale production, multi-user inference & advanced fine-tuning |
| CPU | 20-core Arm (10x Cortex-X925 + 10x Cortex-A725) | 72-core NVIDIA Grace (Neoverse V2) |
| GPU | Blackwell, native FP4 — up to 1,000 TOPS (1 PFLOP) | Blackwell Ultra — surpasses NVIDIA H100 throughput |
| Memory | 128GB LPDDR5x unified, 273 GB/s | 748GB coherent — 252GB HBM3e @ 7.1 TB/s + 496GB LPDDR5X @ 396 GB/s |
| Model capacity | Inference to ~200B params; fine-tune to ~70B; link 2 units for ~405B | Frontier-class models; 400B→70B teacher–student distillation on one box |
| Storage | 2TB / 4TB self-encrypted NVMe M.2 | 2x Gen5 M.2 (OS) + 2x Gen6 M.2 (data) |
| Networking | 10GbE + 2x QSFP 200Gbps + Wi-Fi 7 (secure variant: wired only) | ConnectX-8 SuperNIC, 800Gbps total (wired only — air-gap capable) |
| Ports | 3x USB-C 20Gbps, HDMI 2.1a, RJ-45 | Enterprise I/O, rack-ready (5U) |
| Form factor | 150 × 150 × 51 mm, 1.2kg, 240W USB-C | Convertible tower / 5U rack, liquid cooled, 1600W (20A) |
| OS & software | NVIDIA DGX OS + HP ZGX Toolkit | Ubuntu + NVIDIA AI Dev Tools + ZGX Toolkit + HP Z Runtime |
| Availability (SG) | Available to order now | Available to order now |
| Indicative price | From S$7,800* | From S$125,000* |
What can you actually run?
Unified, coherent memory is the superpower of Grace Blackwell — the whole model lives next to the GPU. A practical guide to model sizes (FP4 quantised):
Local LLMs & copilots
Serve OpenAI-compatible API endpoints on your LAN with HP Z Runtime — chatbots, coding copilots and RAG over internal documents, with zero per-token cost.
Fine-tuning on private data
Train and fine-tune on proprietary datasets that can't go to the cloud — clinical data, classified material, IP-sensitive design and simulation data.
Edge AI & computer vision
Build real-time computer-vision and domain-specific models for edge deployment — develop locally, validate, then export to production infrastructure.
HP ZGX Toolkit: measurably faster workflows
The free, open-source ZGX Toolkit (with VS Code extension, MLflow, Ollama and pre-configured frameworks) delivers up to 45% faster time-to-results on average versus a DIY stack.* HP's published workflow timings:
Discover & pair
Find ZGX systems on your network and pair them with your laptop or workstation — Windows, Mac or Linux — straight from VS Code.
Develop & fine-tune
Quick-start templates and pre-configured packages (MLflow, Ollama, open-source frameworks) so you build and iterate immediately.
Export & deploy
Sync and export models when validated — deploy to your production infrastructure or cloud when ready. Local first, cloud when it counts.
Why teams are bringing AI in-house
Cloud AI costs are unpredictable and escalating — some AI-native teams report cloud AI development spend exceeding ~US$75K/month. Queue times slow development velocity, and compliance requirements can block cloud processing of sensitive data altogether. Here's how to choose:
When ZGX is the right fit
- Data can't leave your facility
- Designed to work within HIPAA, ITAR, or IP-compliance frameworks
- Cloud costs have become a concern
- Queue times are slowing your AI team's development
- You need deterministic latency for production inference
- You want predictable CapEx, not variable OpEx
When cloud is still the right fit
- You're experimenting and need elastic scale
- Workloads are bursty and unpredictable
- No compliance constraints on data processing location
- You only need access to frontier models via API
- You're not yet hitting cost or queue friction
Most organisations use both. ZGX systems fill the gaps — compliance-sensitive development, cost-sensitive inference, and workflows where data gravity matters. Build locally, deploy to cloud when ready.
Use cases by industry
AI development and production inference across regulated and data-sensitive sectors.
Healthcare
Fine-tune medical imaging models on proprietary clinical datasets, designed to work within HIPAA-class compliance. Real-time radiology AI assist with deterministic latency.
Government & defence
Air-gap capable and SCIF-deployable. Fine-tune LLMs on classified intelligence and run analyst copilots with no external connectivity. Federal/TAA-compliant options.
Product design & manufacturing
Train generative design on simulation data with full IP protection. Real-time design optimisation without cloud round-trips or queueing delays.
Research labs
Knowledge distillation (400B→70B) on a single system thanks to 748GB memory. Multi-modal, long-context experimentation on internal data.
Financial services
Run models on confidential trading, risk, and customer data entirely on-premises with predictable costs — no per-token fees.
Education & smart nation
Give research and AI teams a shared departmental AI supercomputer — serving multiple concurrent users with frontier-class models.
Frequently asked questions
What Singapore buyers ask us about the NVIDIA GB300, GB10 and the HP ZGX series.
What is the NVIDIA GB300?
The NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip pairs a 72-core Grace CPU with a Blackwell Ultra GPU, sharing 748GB of coherent memory (252GB HBM3e + 496GB LPDDR5X). It brings datacenter-class AI performance to deskside systems like the HP ZGX Fury G1n.
How does the GB300 compare with the NVIDIA H100?
HP positions the ZGX Fury's GB300 as delivering high-throughput local AI execution surpassing the NVIDIA H100 GPU for supported workloads — with far more accessible memory in a deskside form factor, and no data centre buildout required.
Is the HP ZGX Nano the same as the NVIDIA DGX Spark?
The ZGX Nano G1n is built on the same NVIDIA GB10 Grace Blackwell platform as the NVIDIA DGX Spark — same 20-core Arm CPU, Blackwell GPU and 128GB unified memory, running NVIDIA DGX OS. HP adds the free, open-source ZGX Toolkit (with VS Code integration) and HP enterprise support options, at a competitive Singapore price through SourceIT.
Which one should I buy — Nano or Fury?
Choose the Nano if you're a developer, researcher or small team experimenting, prototyping and fine-tuning models up to ~70B parameters. Choose the Fury if you need production inference for multiple concurrent users, frontier-class model sizes, deterministic latency, or air-gapped/compliance-bound deployment. Many customers start with Nanos for development and add a Fury for production — both share the same software stack, so models move over cleanly.
Can I link multiple ZGX Nano units?
Yes — two Nano units can be connected over the built-in 200Gbps QSFP interconnect to run inference on models up to roughly 405B parameters (FP4).
Is there a secure option for regulated or air-gapped environments?
Yes. HP offers a ZGX Nano variant for secure and regulated environments — no Wi-Fi or Bluetooth, self-encrypted NVMe storage, and customer-controlled wired connectivity only. The ZGX Fury is wired-only and air-gap capable as standard, and SCIF-deployable for government use.
Do I need to change my laptop or OS to use a ZGX?
No. The ZGX sits on your network as a dedicated AI system; you develop from your existing Windows, Mac or Linux machine via the ZGX Toolkit's VS Code integration — discover the device, pair, and start building.
Is the GB300-powered ZGX Fury available in Singapore?
Yes — both the ZGX Nano and the GB300-powered ZGX Fury are available to order in Singapore now through SourceIT. There's no obligation at enquiry stage; we'll confirm configuration, lead time and final pricing before you commit.
How much do the ZGX AI Stations cost in Singapore?
The GB10-powered ZGX Nano starts from an indicative S$7,800* and the GB300-powered ZGX Fury from an indicative S$125,000*, depending on configuration. Email sales@sourceit.com.sg or call +65 6978 3502 for a formal quotation, including enterprise support and volume pricing.
Can SourceIT help with deployment and integration?
Yes. SourceIT is a Singapore-based IT consultation and system-integration provider serving government, education, healthcare, and corporate sectors — from pre-sales sizing to installation, networking, and after-sales support.
Talk to a Singapore AI infrastructure specialist
Get a formal quotation for the ZGX Nano or the GB300-powered ZGX Fury, or ask us to size a hybrid cloud + on-prem setup for your workloads.
Email sales@sourceit.com.sg Call +65 6978 3502SourceIT Pte Ltd · 178 Paya Lebar Road, #03-11, Singapore 409030 · Mon–Fri 9am–6pm
* Indicative pricing based on prevailing Singapore market rates; final pricing subject to configuration, availability and formal quotation. Time-savings figures per HP's published workflow scenarios (prototyping 45→25 min, fine-tuning 50→30 min, inference 43→23 min; ~45% average) and may vary. Model-size guidance assumes FP4 quantisation and varies by model architecture and context length. "Designed to work within" compliance frameworks does not constitute certification; customers are responsible for their own compliance validation. Product specifications per HP/NVIDIA announcements and subject to change. Third-party quotes are the property of their respective publications. NVIDIA, GB10, GB300, Grace, Blackwell, DGX, DGX Spark and ConnectX are trademarks of NVIDIA Corporation. HP and ZGX are trademarks of HP Inc. OpenAI-compatible API references do not imply sponsorship, endorsement, or affiliation with OpenAI. All other trademarks are property of their respective owners. © 2026 SourceIT Pte Ltd. All rights reserved.