GB300 Grace Blackwell Ultra
One NVIDIA Grace CPU with 72 Arm Neoverse V2 cores and one NVIDIA Blackwell Ultra GPU, joined as a single desktop superchip.
Home / Workstations / ASUS ExpertCenter Pro ET900N G3
ASUS ExpertCenter Pro ET900N G3
An AI supercomputer that fits under your desk
Built on NVIDIA DGX Station architecture around the GB300 Grace Blackwell Ultra superchip, with 748 GB of coherent memory in one standard tower.
Rp 3.499.000.000
Price per unit. Contact us for details.
748 GBcoherent memory
20 PFLOPSAI performance, up to (ASUS preliminary figure)
2 × 400GConnectX-8 network ports
72 coresArm Neoverse V2 CPU
A new class of deskside AI
Data-center AI, in an office-ready tower
The ET900N G3 is ASUS's version of the NVIDIA DGX Station. It is designed to build and run AI locally: training, fine-tuning and inference on very large models, plus autonomous agents, without sending your data to someone else's cloud. It is compatible with NVIDIA NemoClaw and NVIDIA AI Enterprise, and ASUS plans to add Windows-based AI development later.
The superchip
One GB300 superchip, one pool of memory
A Grace CPU and a Blackwell Ultra GPU sit side by side and talk over NVLink-C2C, so large models are not squeezed through the slow path between a normal CPU and GPU.
NVLink-C2C interconnect
A chip-to-chip link that gives the CPU and GPU fast, coherent access to each other's memory.
Blackwell Tensor Cores
Support for 4-bit floating point (NVFP4), which lets memory hold larger next-generation models while keeping accuracy high.
Large coherent memory
Train and run big models inside one 748 GB memory pool instead of splitting them across separate CPU and GPU memory.
Workload-aware power
The system shifts power toward whatever is running at the moment, to keep performance and efficiency high.
Add an RTX PRO card
One extra NVIDIA RTX PRO Blackwell GPU can go in beside the superchip, pairing AI compute with ray-traced visualization and simulation.
Memory and networking
748 GB for the models that will not fit elsewhere
ASUS says the machine can host trillion-parameter models and autonomous agents at the desk.
GPU memory at 7.1 TB/s, CPU memory at 396 GB/s.
- ConnectX-8 SuperNICUp to 800 Gb/s of network bandwidth, and ASUS says it can link up to two DGX Stations to scale model size and speed further. The ET900N G3 has two 400G QSFP112 ports.
- Room to growThree PCIe 5.0 slots and up to four M.2 NVMe drives (8 TB in total). Two 2 TB drives for the operating system come installed.
Software
Ready for AI from the first boot
The machine runs Ubuntu with NVIDIA AI Developer Tools, with system-specific settings and the CUDA-X libraries already in place. On top of that, ASUS adds Zenni AI Hub.
- ASUS Zenni AI HubDeploy language models and route each request to the best-fit one. Monitor system resources, manage MIG and benchmark performance, then connect deployed models to your applications through APIs.
- Agents and knowledgeMulti-agent collaboration, plus Knowledge+ so staff can chat with documents, search them and automate tasks from one platform.
- Share the GPUNVIDIA Multi-Instance GPU (MIG) splits the GPU into up to seven isolated instances, so several people or services can work at once.
Agentic AI
Agents you can keep under control
With the NVIDIA NemoClaw open-source stack, the ET900N G3 adds privacy and security controls to OpenClaw agents. NemoClaw installs NVIDIA OpenShell, which enforces policy-based guardrails on how agents behave and what data they touch.
The machine also has the compute to run open models such as NVIDIA Nemotron locally, which keeps data private and can lower running cost.
Applications
Where teams put it to work
ASUS points to five areas where deskside AI of this size makes a difference.
Wondering which models fit which hardware? See Which hardware runs which AI?
Thermals and acoustics
Cooled like a data center, quiet like an office
A closed-loop all-in-one liquid cooler keeps performance steady under long AI jobs. The machine is designed to run in rooms from 10 to 35 °C, and in Auto mode at full load without an extra graphics card it stays near 50 dB, which ASUS compares to a quiet office.
10–35 °Coperating temperature
~50 dBAuto mode, full load, no extra GPU
Connectivity and size
Ports on the front and back, in a standard tower
Front
- Power button and message LEDs
- Audio: headphone and microphone
- USB 10Gbps: 2 × Type-C, 2 × Type-A
- USB 2.0 × 1
Back
- 400G QSFP (ConnectX-8) × 2
- 10GbE and 1GbE (BMC) Ethernet
- USB 10Gbps Type-A × 4
- mini DisplayPort and micro-USB COM (both for the BMC), 3 audio jacks
Size
- 584 × 232 × 565 mm
- 27 kg net, 32 kg boxed
Buy it from EMD Lab
The machine is ASUS's. The rest is ours.
When you order the ET900N G3 from us, it is not just a box on a pallet.
Delivered and set up in Indonesia
We bring the machine to you and set it up where it will live.
AI software and models installed
The AI software and the models you need are installed, so you can start work on day one.
Tested before handover
We test the machine before it leaves us, so you receive one that is known to work.
Local support
You have an Indonesian team to call. ASUS offers onsite hardware support after remote diagnosis; availability varies by country, so ask us about support in Indonesia.
Before you buy
A serious machine needs a little planning
Check these points first; we will go through them with you.
- Power1600 W power supply, IEC C19 power cord (ASUS lists the cord as optional). Outside North America it needs a 200–240 V supply, so plan a dedicated circuit.
- Size and weight584 × 232 × 565 mm and 27 kg (32 kg boxed). Put it on a solid desk or on the floor.
- Room temperatureOperating range is 10–35 °C. With the optional RTX PRO 6000 under heavy load, ASUS recommends an ambient temperature below 30 °C.
- Arm softwareThe CPU is Arm (aarch64), so your software needs a Linux Arm64 build. Most AI tools do, including PyTorch, vLLM, llama.cpp and the NVIDIA stack.
- 400G network portsThe two QSFP112 ports need QSFP transceivers, which are optional and not included.
- Optional graphics cardOne card can go in the first PCIe slot: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, RTX PRO 4000 Blackwell SFF or RTX PRO 2000 Blackwell.
Tech specs
Specifications
System
- Model
- ExpertCenter Pro ET900N G3
- Operating system
- Ubuntu with NVIDIA AI Developer Tools
- Processor / system bus
- Single NVIDIA Grace CPU Superchip with 72 Arm Neoverse V2 cores; single NVIDIA Blackwell Ultra GPU
- Memory
- 748 GB of coherent memory: 496 GB LPDDR5X CPU memory (396 GB/s) and 252 GB HBM3e GPU memory (7.1 TB/s)
- AI performance
- Up to 20 petaFLOPS (ASUS preliminary specification, subject to change)
Storage, slots and graphics
- Storage
- 4 M.2 slots in total (optional). 2 × M.2 Key M, type 2280, PCIe 5.0 x4, pre-configured for OS RAID 1; 2 × M.2 Key M, type 2280, PCIe 6.0 x4, for training data. Includes 2 × 2 TB NVMe M.2 SSDs, pre-installed for the OS.
- Expansion slots
- 3 in total: location 1 PCIe x16 slot (Gen 5.0 x16); locations 2 and 3 PCIe x16 slots (Gen 5.0 x8)
- Graphics (optional)
- Up to one NVIDIA graphics card in the first PCIe slot: RTX PRO 6000 Blackwell Max-Q Workstation Edition (ambient under 30 °C recommended during heavy workloads), RTX PRO 4000 Blackwell SFF, or RTX PRO 2000 Blackwell
Networking and management
- Networking
- 2 × NVIDIA ConnectX-8 SuperNIC QSFP112 ports (QSFP transceiver required); 1 × Marvell 10G Ethernet; 1 × Realtek 1Gb Ethernet (dedicated management LAN for the BMC); 1 × M.2 slot (Key E)
- Management
- ASPEED AST2600 with AMI MegaRAC based firmware, supporting IPMI 2.0 and DMTF Redfish API; on-board TPM 2.0
Ports and indicators
- Front I/O ports
- 2 × USB 10Gbps Type-C, 2 × USB 10Gbps, 1 × USB 2.0, 1 × headphone, 1 × microphone
- Rear I/O ports
- 4 × USB 10Gbps (Type-A), 1 × micro-USB COM (for BMC), 2 × CX-8 NIC 400G QSFP, 1 × Realtek 1Gb Ethernet (for BMC), 1 × Marvell 10Gb Ethernet, 1 × mini DisplayPort (for BMC), 3 × audio jacks
- Switch / LED
- Front: power button, storage activity LED, system message LED. Rear: 2 × QSFP LAN activity LEDs, 2 × 10G LAN message LEDs (activity and speed), 2 × 1G LAN message LEDs (activity and speed)
Power, size and environment
- Power supply
- 1600 W Titanium ATX power supply (115–240 Vac, 15–8 A, 50–60 Hz). Power cord IEC C19 type (optional); maximum output depends on regional input voltage and cord. North America: 125 V AC (IEC C19 to NEMA 5-20P). Global: 200–240 V AC (IEC C19 to regional plug).
- Dimensions
- 584 mm × 232 mm × 565 mm (23.0″ × 9.1″ × 22.3″)
- Weight
- Net 27 kg, gross 32 kg
- Environment
- Operation temperature 10–35 °C. Non-operation temperature 5–40 °C. Non-operation humidity 20–90% (non-condensing).
- Regulatory compliance
- CE, FCC (Class A), BSMI, cTÜV SÜD us, KCC
- Optional items
- QSFP 400G transceivers; Wi-Fi module and antenna
From ASUS's product page
- ASUS software
- ASUS Zenni AI Hub: deploy and route LLMs, monitor resources, manage MIG, benchmark, multi-agent collaboration, Knowledge+ document chat
- GPU sharing
- NVIDIA Multi-Instance GPU (MIG), up to seven isolated instances
- AI stack
- NVIDIA AI Software Stack; compatible with NVIDIA NemoClaw and NVIDIA AI Enterprise. ASUS plans future support for Windows-based AI development.
- Cooling and noise
- Closed-loop AIO liquid cooling. About 50 dB in Auto mode at maximum load without a discrete graphics card.
- Support
- Onsite hardware service after remote diagnosis; availability and coverage vary by country or region
Source: ASUS's Tech Specs page and product page, read 25 October 2026. All specifications are subject to change without notice. The 20 petaFLOPS figure is ASUS's preliminary specification. Product pictures are ASUS's.
Talk to us about the ET900N G3
Rp 3.499.000.000 per unit. Tell us what you want to run and how many people will use it. We answer with a written quote and go through power, network and support with you.