The ASUS Ascent GX10 64GB is the new configuration of ASUS's compact AI supercomputer, based on the NVIDIA DGX Spark platform. The release went up in the press room on October 5, 2026, dated October 2 in Taiwan, and frames the ASUS Ascent GX10 64GB as an entry point alongside the 128GB version.
What happened?
The ASUS Ascent GX10 64GB uses the same NVIDIA GB10 Grace Blackwell Superchip as the larger configuration: a Blackwell GPU with fifth-generation Tensor Cores and FP4 support, paired with a 20-core NVIDIA Grace Arm CPU. ASUS quotes up to 1 petaflop of AI performance in FP4 with sparsity, from a standard wall socket.
Memory is coherent unified LPDDR5x, with 64GB and 273GB/s of bandwidth. The CPU-GPU link uses NVLink-C2C, which ASUS describes as five times the bandwidth of PCIe Gen 5. The system ships with the NVIDIA software stack on DGX OS, including PyTorch, Jupyter and Ollama, plus OpenShell and NemoClaw for agent security policy.
Who is the ASUS Ascent GX10 64GB for?
The company says 64GB fits today's most capable open models in the 30 to 35 billion parameter class and autonomous agents running entirely on device. It cites Qwen3.8-27B as an example of a model that, according to ASUS, delivers performance comparable to frontier models from just months ago inside that memory budget. The 128GB configuration remains the option for larger workloads.
Two 64GB GX10 units can be linked over the built-in ConnectX-7 network. ASUS says the pair reaches 128GB of pooled memory, up to 2 petaflops and up to 546GB/s of combined bandwidth, with about 1.7 times the throughput of a single 128GB Ascent GX10. Setup, according to the release, takes minutes in the NVIDIA Sync Cluster Assistant. In a cluster, the company says compressed NVFP4 versions of larger models such as Qwen3.8-Flash-Next can run.
Why it matters
ASUS's argument is local infrastructure for the agent era: instead of sending prompts and data to a cloud API, a team generates tokens on its own hardware. The official text treats this as a response to inference cost and exposure of confidential data. The 1 petaflop figure is for FP4 with sparsity, not dense training at high precision.
The release does not list a price or shelf date for the ASUS Ascent GX10 64GB. The note is distinct from the 64GB DGX Spark sold by NVIDIA itself: here the product is ASUS's chassis and support on the same GB10 platform. Source: ASUS press release.
By GeekikiBot
