Asus launches Ascent GX10 AI supercomputer at $2,999, undercutting Nvidia's DGX Spark

Asus is stepping into the desktop AI supercomputer market with the Ascent GX10, a machine that costs $2,999 and promises to put serious AI processing power within reach of individual developers, researchers, and data scientists who have grown tired of relying on cloud services or distant data centers.

The device runs on Nvidia's Grace Blackwell GB10 Superchip—the same processor that powers Nvidia's own DGX Spark—paired with 128 gigabytes of unified memory and a 20-core Grace Arm CPU. That combination delivers up to 1,000 TOPS (tera operations per second) of AI processing capacity, enough to handle machine learning models with as many as 200 billion parameters without hitting major performance walls. The Blackwell GPU includes fifth-generation Tensor Cores and FP4 precision support, technologies designed specifically for the kind of mathematical operations that power modern AI inference and model fine-tuning.

The price difference between the Ascent GX10 and Nvidia's competing DGX Spark is substantial. Nvidia's machine costs $3,999 and arrives with 4 terabytes of storage. Asus's version undercuts it by a thousand dollars but includes only 1 terabyte of storage—a meaningful trade-off for users who plan to keep large model weights or datasets locally. The Ascent GX10 is scheduled to enter pre-order in the second quarter of 2025.

What makes the Ascent GX10 technically interesting is its use of NVLink-C2C, a high-bandwidth connection between the CPU and GPU that offers more than five times the throughput of PCIe 5.0. This architecture allows the processor and graphics chip to share memory directly, which matters because it reduces the bottlenecks that typically slow down AI workloads when data has to shuttle back and forth between separate memory pools. Asus also equipped the machine with an integrated ConnectX network interface, meaning two Ascent GX10 units can be daisy-chained together to tackle even larger models—Llama 3.1, for instance, which contains 405 billion parameters.

KuoWei Chao, the general manager of Asus's IoT and NUC Business Group, framed the machine as a democratization play. "AI is transforming every industry," he said, "and the Asus Ascent GX10 is designed to bring this transformative power to every developer's fingertips." The pitch is straightforward: instead of spinning up expensive cloud instances or waiting for access to shared lab hardware, a researcher can now own a machine capable of running serious AI work from their own desk.

The move reflects a broader shift in how AI development is happening. As models have grown larger and more capable, the infrastructure required to work with them has become more expensive and more centralized. Asus's entry into this space, using Nvidia's own silicon, suggests that vendors see real demand from developers who want local control over their AI experiments—and who are willing to pay a few thousand dollars to get it. The real test will come when pre-orders open and the market reveals whether the price-to-storage trade-off actually appeals to the people Asus is trying to reach.