ASUS Ascent GX10 Guide: Specs, LLMs, Agents, Clustering and Price in Singapore

The ASUS Ascent GX10 is a desktop AI supercomputer the size of a coaster: 150 x 150 x 51 mm, 1.48 kg, and inside it the same NVIDIA GB10 Grace Blackwell Superchip, 128 GB of unified memory and 1 petaFLOP of FP4 compute as NVIDIA's own DGX Spark. ASUS calls it a Compact Agent Computer, and since launch it has become SourceIT's best-selling GB10 machine in Singapore — partly on price, partly because it is the only DGX Spark-class system available in three storage tiers. This guide covers what the GX10 is, what it runs, how it clusters, how the agent software stack works, and what each configuration costs. For how it compares against the Dell, HP, Lenovo, MSI, Gigabyte and Acer GB10 boxes, read our NVIDIA GB10 buyer's guide.

ASUS Ascent GX10 specifications

What the GX10 actually is: an ASUS DGX Spark

NVIDIA announced the GB10 platform at GTC 2025 as DGX Spark and named ASUS among the launch system builders. The GX10 shares the Superchip, the memory configuration, DGX OS and the NVIDIA AI software stack with every other GB10 system, so inference performance is the same — independent testing by StorageReview found the ASUS unit tracking the Founders Edition and the Dell, Acer and Gigabyte builds closely across vLLM serving sweeps, and holding its results after an hour of heat soak. What ASUS controls is the chassis, cooling, storage SKUs, port layout, price and the support relationship. The GX10 was announced on 18 March 2025, went on sale in October 2025, picked up three major design awards in 2026, and runs the latest DGX OS 7.x releases.

Which AI models run on the Ascent GX10?

Memory, not TOPS, sets the ceiling. With 128 GB of unified memory the GX10 runs a single model of up to roughly 200 billion parameters at FP4, which in practice means:

Interactive (7B–32B): Llama 3.2, Gemma, Qwen3 Coder 30B, GPT-OSS 20B, DeepSeek-R1 distills — chat-grade responsiveness for coding assistants, RAG and agents.

Daily driver (≈70B): Llama 3.3 70B, DeepSeek-R1 70B, Qwen 70–80B quantised — the quality-versus-speed sweet spot and the tier no single consumer GPU can load.

Ceiling (100B–200B): GPT-OSS 120B, Nemotron Super 120B at FP4 — research and evaluation speeds, entirely on the desk.

Fine-tuning: full fine-tuning to about 8B, LoRA at 8B-class speed, QLoRA for 70B-class, via LLaMA Factory, PyTorch or NVIDIA NeMo. Image and video generation (FLUX, Qwen-Image, LTX) also runs well.

Expect comfortable interactive token rates rather than datacenter throughput — 273 GB/s of bandwidth is the governing number for generation speed, and it is why a GX10 is the right tool for a developer or a small team rather than a 100-user service (for that, see the ET900N G3).

Running AI agents: NemoClaw, OpenClaw, Hermes and PAIR

This is where ASUS differentiates the GX10 from a plain DGX Spark in its marketing, and it is worth understanding. OpenClaw is the open-source agent framework that lets a local model take actions — read files, browse, call tools, run long jobs. NVIDIA NemoClaw wraps OpenClaw with OpenShell, which installs policy-based privacy and security sandboxes so you control what an autonomous agent can touch and where its data goes; ASUS ships the GX10 with an optimised NemoClaw setup. Hermes Agent (a self-improving agent with a browser admin panel and 200-plus supported models) is also validated. And NVIDIA's Personal AI Router (PAIR) lets the GX10 act as the anchor of a personal AI cluster, routing inference across every RTX GPU, Mac and Windows PC you own through a single endpoint with no cloud round-trip.

The economics follow from that. An agent running all day against a cloud API bills per token; the GX10 draws around 240 W, runs 24/7, and never sends a prompt off-premises unless you configure it to. For Singapore organisations handling PDPA-regulated data, client documents or source code, that is the entire argument — develop and run agents locally, burst to cloud only for frontier proprietary models. ASUS's Agent Computer programme frames the GX10 as the "private lab for a developer" tier.

Clustering: two GX10s for 405B models, four for 512 GB

Every GX10 carries NVIDIA ConnectX-7 networking. Two units connected directly with a QSFP direct-attach cable pool memory to 256 GB, enough for Llama 3.1 405B-class and Qwen3-235B-class models. NVIDIA's Sync Cluster Assistant now supports up to four units — 512 GB and up to 4 petaFLOPS — through an external switch. SourceIT stocks the ASUS QSFP cable for the GX10 (40 cm, S$120) and the NVIDIA QSFP112 400G DAC (S$155). Place clustered units side by side, not stacked, for airflow — the magnetic footpads are designed for modular alignment.

Ascent GX10 price in Singapore: 1 TB vs 2 TB vs 4 TB

The three configurations are identical in compute and memory; only the NVMe capacity differs. GX10 128 GB / 1 TB at S$7,200 is the lowest-cost entry into the entire GB10 ecosystem in Singapore. GX10 128 GB / 2 TB at S$8,050 is the sensible default if you will keep several 70B checkpoints on the box. GX10 128 GB / 4 TB at S$9,850 suits fine-tuning datasets and a library of models. Prices are at time of writing and move with supply; the live GX10 collection always shows current pricing. Higher-education institutions and NVIDIA Inception members may qualify for ASUS special-programme pricing — ask us.

Setup: 20 minutes to your first model

The GX10 is closer to an appliance than a Linux build. Position it with clear airflow, connect wired 10 GbE, plug in the 240 W adapter, and either attach a monitor or complete the DGX OS wizard headlessly. Run updates first — the platform gains performance with each DGX OS release. Then ollama run gpt-oss:20b gives you a local chat endpoint; for production-style serving deploy an NVIDIA NIM container, and for maximum control use vLLM. Work from your laptop over SSH, VS Code Remote or JupyterLab. SourceIT commissions units on request — delivered updated, networked and serving your chosen models. Our DGX Spark setup and enterprise deployment guide applies to the GX10 unchanged.

Who should buy the Ascent GX10 — and who shouldn't

Buy it if you are a developer, data scientist or research group that needs 70B-class models privately, wants to build agents against the production CUDA stack, or is replacing a monthly cloud-API bill with a one-off purchase. Buy two if 405B-class models are on the roadmap. Consider the Ascent QN10 instead if you need Windows and a low-power NPU for always-on small-model inference — see QN10 vs GX10. Consider a ROG NUC if the models fit in 16 GB of VRAM and you also want Windows gaming. Step up to the ET900N G3 if a whole team will hit the endpoint or the models are 200B–500B parameters.

Buying the Ascent GX10 in Singapore

SourceIT is an authorised ASUS reseller with all three GX10 configurations in stock, official Singapore warranty, QSFP cables for clustering and commissioning as a service. GB10 units are supply-constrained worldwide, so for institutional 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

Is the ASUS Ascent GX10 the same as the NVIDIA DGX Spark? Same GB10 Superchip, memory, DGX OS and software stack, so the same AI performance. ASUS designs the chassis, cooling, storage options and pricing, and provides the warranty.

What is the biggest model a GX10 can run? About 200B parameters at FP4 on one 128 GB unit; two linked units reach roughly 405B, four units 512 GB of pooled memory.

Does it run Windows? No. It runs NVIDIA DGX OS (Ubuntu-based). You use it from any laptop over the network, or attach a display for a local desktop.

Can it run OpenClaw? Yes — ASUS validates OpenClaw and Hermes Agent, and ships an optimised NVIDIA NemoClaw setup that adds security sandboxes around OpenClaw.

Is it faster than an RTX 5090 PC? For models that fit in 32 GB of VRAM the 5090 generates tokens faster. The GX10 wins on capacity — it loads 70B–200B models the 5090 cannot — and on power (about 240 W total).

What is the warranty in Singapore? Official ASUS Singapore warranty on all units supplied by SourceIT; confirm the term for your SKU when requesting a quote.

Which storage size should I choose? 1 TB if you will run one or two models at a time; 2 TB as a general default; 4 TB for fine-tuning datasets and a model library. Compute is identical across all three.