ASUS Agent Computer Guide 2026: NUC, Ascent GX10, Ascent QN10, ROG NUC & ExpertCenter Pro ET900N G3 Compared
Posted by Wei Fei on
ASUS now sells its AI hardware under one umbrella it calls the Agent Computer — a machine built to run AI agents locally, from a 0.7-litre mini PC on a desk to a 27 kg DGX Station under it. The range SourceIT stocks in Singapore spans five very different products that share that badge: the ASUS NUC 15 Pro and NUC 16 Pro, the Ascent QN10, the Ascent GX10, the ROG NUC, and the ExpertCenter Pro ET900N G3. They run different chips (Intel, Qualcomm, NVIDIA Grace Blackwell), different operating systems (Windows or DGX OS) and cost anywhere from under S$2,000 to well into six figures. This guide explains what "agent computer" actually means, lines the five up side by side, and tells you which one to buy for which job. Each model also has its own deep-dive linked below.
What is an ASUS Agent Computer?
ASUS defines an agent computer as a device built to run AI agents — software you delegate tasks to rather than operate step by step. The practical difference from a normal "AI PC" is where the model runs. A Copilot+ laptop uses a small NPU to accelerate features inside Windows; an agent computer is sized so the language model, the agent framework and your data all live on the machine, so the agent can work for hours without sending prompts to a cloud API or paying per token.
ASUS supports this with a software layer rather than just a badge. The ASUS Agent Computer programme lists the open-source agent frameworks each device is validated for — OpenClaw, NVIDIA NemoClaw (which wraps OpenClaw in policy-based security sandboxes), Hermes Agent and ASUS's own Zenni Claw — and lets you choose between cloud, local or hybrid deployment. ASUS's framing is three scenarios: a personal assistant on one PC, a private lab for a developer, and a shared AI service for a team of up to about 25 users on one workstation. The hardware below maps onto those three tiers.
The five ASUS Agent Computers at a glance
| Model | Silicon | Memory | AI compute | OS | Best for | SourceIT price (SGD) |
|---|---|---|---|---|---|---|
| NUC 16 Pro | Intel Core Ultra Series 3 (Panther Lake), up to Core Ultra X9 388H with Arc B390 | Up to 96 GB LPDDR5X-9600 (X models) / 128 GB DDR5 (356H) | NPU 50 TOPS + Arc Xe3 iGPU | Windows 11 Pro | Office desktops, Teams Rooms, small local models, signage | From S$1,955 |
| Ascent QN10 | Qualcomm Snapdragon X2 Elite, 18-core Oryon | 16 / 32 GB LPDDR5X | Hexagon NPU, 80 TOPS INT8 | Windows 11 Pro (Copilot+) | Always-on assistants, kiosks, edge inference, Windows on Arm dev | S$2,375 (32 GB / 1 TB) |
| ROG NUC 16 | Intel Core Ultra 9 290HX Plus + GeForce RTX 5080 Laptop (16 GB GDDR7) | Up to 64 GB DDR5 | RTX 5080 Tensor Cores (CUDA) | Windows 11 Home | Creators and gamers, CUDA dev on small models, Stable Diffusion | From S$3,467 |
| Ascent GX10 | NVIDIA GB10 Grace Blackwell Superchip (DGX Spark platform) | 128 GB coherent unified LPDDR5X, 273 GB/s | 1 PFLOP FP4 | NVIDIA DGX OS (Ubuntu) | Developers, researchers, 70B–200B LLMs, fine-tuning, agent labs | From S$7,200 |
| ExpertCenter Pro ET900N G3 | NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip (DGX Station) | 748 GB coherent (252 GB HBM3e + 496 GB LPDDR5X) | Up to 20 PFLOPS NVFP4 | Ubuntu + NVIDIA AI stack | Teams serving frontier-scale models on-premises, 400B+ parameter inference | Price on request |
Prices are at time of writing and move with stock; the product pages always show the live figure.
Tier 1 — the Windows mini PCs: NUC 16 Pro and Ascent QN10
Both are 0.7-litre boxes that sit behind a monitor, run Windows 11 Pro, and are the right answer when the "agent" is a personal assistant, a document summariser or a kiosk model rather than a 70-billion-parameter LLM. They differ in the silicon philosophy.
NUC 16 Pro is the x86 choice: Intel's Panther Lake Core Ultra Series 3, a 50 TOPS NPU, and on the X7 358H and X9 388H configurations the 12-core Arc B390 integrated GPU that makes 7B–14B models usable through Ollama or LM Studio. It keeps the NUC virtues businesses buy it for — dual 2.5GbE, dual Thunderbolt 4, four-display output and a toolless chassis — which is why it is also the standard compute for Microsoft Teams Rooms kits. Full breakdown in our ASUS NUC buyer's guide.
Ascent QN10 is the Arm choice: the first mini PC on Qualcomm's Snapdragon X2 Elite, with an 80 TOPS Hexagon NPU that comfortably clears the Copilot+ bar and is designed for inference that runs all day at low power. It is the pick for signage, retail computer vision, always-on desk-side assistants and anyone building on Qualcomm AI Hub's 175-plus pre-optimised models. It is not a CUDA machine. We compare the two AI-first Ascents in Ascent QN10 vs Ascent GX10.
Tier 2 — the CUDA desktop: ROG NUC
The ROG NUC is sold as a gaming mini PC, but in a 3-litre chassis it carries a laptop-class GeForce RTX 5080 with 16 GB of GDDR7, which makes it the cheapest ASUS box that runs the real CUDA stack on Windows. For a creator or a developer who wants Stable Diffusion, local coding assistants on 7B–20B models and gaming on the same machine, it is the sweet spot between a NUC and a GX10. The 2026 model moves to Intel's Core Ultra 200HX Plus series; see ROG NUC 16 vs ROG NUC 15 for what changed and whether the 2025 unit is the better buy.
Tier 3 — the developer lab: Ascent GX10
The Ascent GX10 is ASUS's build of the NVIDIA DGX Spark platform: one GB10 Grace Blackwell Superchip, 128 GB of unified memory the GPU can address directly, 1 petaFLOP of FP4 compute, and ConnectX-7 networking to cluster two to four units. It runs DGX OS, not Windows, and it is the machine ASUS puts at the centre of its agent-developer story — NemoClaw ships optimised for it, OpenClaw and Hermes Agent are validated, and NVIDIA's Personal AI Router lets it anchor every other GPU on your network behind one endpoint. If you need to run a 70B model privately, fine-tune with LoRA, or give a research team a lab that mirrors production CUDA, this is the entry point. Our Ascent GX10 guide covers specs, supported LLMs, clustering and Singapore pricing; our GB10 buyer's guide compares it with the other DGX Spark OEMs.
Tier 4 — the deskside supercomputer: ExpertCenter Pro ET900N G3
At the top sits the ExpertCenter Pro ET900N G3, ASUS's NVIDIA DGX Station: a 72-core Grace CPU and a Blackwell Ultra GPU on one GB300 superchip, 748 GB of coherent memory, up to 20 PFLOPS NVFP4 and dual 400GbE ConnectX-8. It is a tower that needs a 20 A circuit and a 1,600 W Titanium power supply, and in independent testing it served a 433 GB GLM-5.2 checkpoint that four RTX PRO 6000 cards could not keep up with. This is the "AI service for a department" tier — one box replacing a GPU server for a team's inference, fine-tuning and evaluation. Details, benchmarks and the DGX Spark-versus-DGX Station decision are in our ET900N G3 guide.
Which ASUS Agent Computer should you buy?
"I want Copilot-style agents on a quiet office desktop." NUC 16 Pro. Pick the X9 388H if local models matter, the Core Ultra 7 356H if upgradable RAM matters more.
"It will run 24/7 in a kiosk, lobby or retail floor." Ascent QN10 — the 80 TOPS NPU at a fraction of the power of a dGPU, with Windows on Arm and Qualcomm AI Hub.
"I need CUDA on Windows, and I also game or edit video." ROG NUC with RTX 5080 or 5070 Ti.
"I need to run 70B-class models privately, or build agents against a production NVIDIA stack." Ascent GX10 — and buy two if 405B-class models are on the roadmap.
"A whole team will hit it, and the models are 200B–500B parameters." ExpertCenter Pro ET900N G3.
Unsure about the model-size-to-memory maths? Our AI hardware calculator estimates which tier a given parameter count and quantisation needs.
Local vs cloud: why Singapore buyers are moving to agent computers
Three things drive the local-AI conversation with our customers. Cost predictability: an agent that runs all day against a cloud API bills per token forever; a GX10 is a one-off purchase with unlimited tokens, and for a team hammering a 70B model the hardware can pay for itself within a year. Data residency: for PDPA-regulated data, client files or source code, "the prompt never leaves the office" settles the compliance question. Latency and availability: no rate limits, no queue, no dependency on a provider's uptime. The pragmatic pattern is hybrid — develop and run day-to-day on local hardware, burst to cloud only for frontier proprietary models — which is exactly how ASUS positions the Agent Computer programme.
Buying ASUS Agent Computers in Singapore
SourceIT is an authorised ASUS reseller and stocks the complete range — NUC, Ascent QN10, Ascent GX10 in 1 TB, 2 TB and 4 TB, ROG NUC 15 and 16, and the ExpertCenter Pro ET900N G3 — with official Singapore warranty. We advise on sizing, clustering and commissioning, and deliver GX10 and ET900N G3 units already updated and serving your chosen models. GB10 and GB300 units are supply-constrained, so engage us early for institutional orders. Use Request for Quotation on any product page, call (65) 6978 3502, or email Sales@sourceit.com.sg.
Frequently asked questions
What does "agent computer" mean? It is ASUS's term for a PC, mini PC, workstation or deskside supercomputer sized to run AI agents locally — the model, the agent framework and your data on the device — rather than relying on a cloud API.
Which ASUS products carry the Agent Computer badge? In Singapore, the Ascent GX10, Ascent QN10 and ExpertCenter Pro ET900N G3 are the headline Agent Computer products; ASUS extends the programme to NUC, ROG NUC, ExpertCenter desktops and Powered-by-ASUS custom builds.
Does an agent computer need a GPU? Not necessarily. The QN10 and NUC 16 Pro run small models on an NPU and integrated graphics; the GX10 and ET900N G3 use NVIDIA Grace Blackwell silicon for large models. Match the tier to the model size.
Which runs Windows? NUC 16 Pro, Ascent QN10 and ROG NUC run Windows 11. Ascent GX10 runs NVIDIA DGX OS and the ET900N G3 runs Ubuntu with the NVIDIA AI stack — you work on them from any laptop over the network.
What is the cheapest ASUS machine for a 70B LLM? The Ascent GX10 128 GB. No Windows mini PC in the range fits a 70B model at a quality quantisation.
Can I cluster them? Ascent GX10 units link over ConnectX-7 (two direct, up to four via a switch) and ET900N G3 units over ConnectX-8. The Windows mini PCs are standalone.
