ASUS Ascent GX10 now comes in a 64GB memory configuration, placed beside the existing 128GB model in a compact line built to run AI agents on local hardware. The newcomer keeps the same processor and gives up only memory, which signals the kind of buyer ASUS is courting next: developers who want to start experimenting rather than load the biggest models.
ASUS says the ASUS Ascent GX10 64GB pairs 64GB of LPDDR5x memory with 273GB/s of bandwidth, shared coherently between CPU and GPU. The processor is the NVIDIA GB10 Grace Blackwell Superchip from the original, a Blackwell GPU alongside a 20-core Arm CPU, rated for up to 1 petaflop of AI performance. That rating is for FP4 with sparsity, as the company's own footnote spells out.
What 64GB buys is a defined range. ASUS sizes the ASUS Ascent GX10 64GB for open-source models of roughly 30 to 35 billion parameters and cites Qwen3.8-27B as one that fits comfortably. Anything bigger points to the 128GB version, which the company keeps in the lineup for the largest local models and the heaviest workloads. The announcement names no price, release date or regional availability, so the cost gap between the two remains unknown, even though ASUS describes the smaller system as a lower-cost way in.
Growth is handled by clustering. Two ASUS Ascent GX10 64GB units can be linked over the built-in NVIDIA ConnectX-7 networking for 128GB of pooled memory and up to 2 petaflops of compute, and ASUS lists the pair at up to 1.7 times a single 128GB machine. Those two headline figures do not match, so a buyer weighing a pair should plan around the lower one. Setup runs through NVIDIA Sync and its Cluster Assistant, and with NVFP4 support the pair can run compressed versions of larger models such as Qwen3.8-Flash-Next.
The box ships with NVIDIA DGX OS and the NVIDIA AI software stack, including PyTorch, Jupyter and Ollama, plus NVIDIA OpenShell and NemoClaw for policy-based agent security. SSH-based remote management lets IT teams provision fleets next to existing infrastructure. Cooling carries over from the 128GB design: seven fan-control levels, five heat pipes and two 140x80mm fans, aimed at long agent sessions that keep the chip busy.
The case ASUS makes for local hardware rests on two pressures. Agent workloads burn far more tokens than a chat prompt, which drives up cloud inference bills, and companies are wary of sending confidential data and prompts through public APIs. A 64GB ASUS Ascent GX10 tests that argument for teams that never planned to buy the top configuration.
