On 2 October 2026, ASUS announced from Taiwan a new 64GB configuration of the ASUS Ascent GX10, expanding the lineup alongside the existing 128GB model. Press coverage on 5 October placed the product in the wave of personal “AI supercomputers” for local agent development based on the NVIDIA DGX Spark platform.
The system shares the NVIDIA GB10 Grace Blackwell Superchip of the original Ascent GX10: a Blackwell GPU with fifth-generation Tensor Cores and FP4 support plus a 20-core NVIDIA Grace Arm CPU. NVLink-C2C delivers a coherent CPU+GPU memory model with 5× the bandwidth of PCIe Gen 5, enabling up to 1 petaFLOP of AI performance in a compact chassis that runs from a standard wall socket — ideal for always-on agents.
Sixty-four gigabytes of coherent unified LPDDR5x memory with 273 GB/s of bandwidth are sized for today’s efficient open-source models: ASUS cites the 30–35-billion-parameter class and fully on-device autonomous agents. It names Qwen3.8-27B as an example that fits comfortably in this memory footprint.
When workloads grow, two units can be clustered with the NVIDIA Sync Cluster Assistant and built-in ConnectX-7 networking for 128GB of pooled memory, up to 2 petaFLOPs and up to 546 GB/s of combined memory bandwidth — delivering around 1.7× the performance of a single 128GB Ascent GX10. With NVIDIA’s NVFP4 format, users can run compressed versions of larger models, including Qwen3.8-Flash-Next.
ASUS frames the 64GB SKU as a more accessible entry point for the “early majority” of developers and researchers, while the 128GB variant continues to serve teams pushing the largest local models. The Ubiquitous AI. Incredible Possibilities. brand strategy aims to move builders from curiosity to serious agentic development without forcing the higher-ticket SKU.
The business case is token and data sovereignty: as agents move from prototype to production, token consumption drives up cloud costs and raises IP risks on public APIs. Running open-source models on local hardware keeps token generation and data in-house without giving up the performance multi-step workflows demand.
Factory software includes the full NVIDIA AI stack on DGX OS — PyTorch, Jupyter and Ollama — plus NVIDIA OpenShell and NVIDIA NemoClaw for policy-based agent security, and SSH-based remote management for fleets. The thermal design matches the 128GB model: 7-level fan control, ultrawide fins, five heat pipes and two 140×80 mm fans.
ASUS places the GX10 in a broader portfolio: ProArt laptops and mini-PCs with RTX Spark as personal front ends; Ascent GX10 with DGX Spark for prototyping; and ExpertCenter Pro ET900 with DGX Station for larger enterprise workloads.
Parallel market coverage placed the 64GB DGX Spark class around $4,999 versus about $6,950 for 128GB through Acer, Dell, ASUS and HP, amid a DRAM crunch. Neural Pulse treats those channel prices as indicative and subject to region and availability.
For startups and university labs, the 64GB SKU lowers the barrier to enter the DGX Spark ecosystem without buying double the memory “just in case.” Clustering offers an upgrade path without rewriting the software environment.
Limits: 64GB will not host the largest dense models without aggressive quantisation; the 1 petaFLOP figure is the superchip’s marketing rating in low-precision formats; and OpenShell/NemoClaw still need to prove low friction in real agent workflows.
Competitors include traditional RTX workstations, Mac Studio with unified memory and regional clouds. ASUS bets that wall-socket form factor plus the full NVIDIA stack will differentiate against generic assembled PCs.
Neural Pulse recommends buyers verify real HBM/DRAM availability in their region, compare 12-month TCO against cloud inference for their agent token volume, and trial OpenShell with egress policies before deploying sensitive data.
The announcement reinforces a 2026 thesis: local agent infrastructure is leaving the niche of rich labs and becoming a professional-consumer product line tiered by memory.
In short, ASUS cuts the entry ticket to DGX Spark with 64GB coherent memory, clustering to 128GB and NVIDIA software ready for agents — a hardware answer to the on-prem agentic inference boom.
