ASUS Ascent GX10 (NVIDIA GB10)

The ASUS Ascent GX10 is Asus’s take on Nvidia’s little GB10 supercomputer - the same Grace-Blackwell chip as the DGX Spark, 128GB of unified memory, the full CUDA stack, and a ConnectX-7 NIC for lashing two together into a 256GB pair. One correction up front: it runs Linux, not Windows. This is a developer’s CUDA box, and - until a 2026 memory shortage pushed its price above the Spark’s - it was the cheaper door into the platform. Now the reason to pick it is the clustering NIC, not the price.

The take

Buy it for the software and the memory, not the raw speed - the same line as every GB10 box. 128GB lets you load models a 24GB card can’t hold, and you get native CUDA and the DGX software stack to do it on, which is the real draw for anyone who lives in Nvidia’s tooling. But the memory runs at around 273 GB/s, so single-stream token speed on a large model is a walk, not a sprint. Where the GX10 earns its place over the DGX Spark is networking: the ConnectX-7 lets you pair two into one 256GB machine. Its old price advantage is gone, mind - the 2026 memory shortage pushed it to around $5,999, above the Spark’s ~$4,699 - so it’s the clustering, not the cost, that sells it now. Just don’t expect Windows either; this is Ubuntu and the DGX stack.

What it’ll actually run

The 128GB is the point. A 70B at Q4 needs roughly 40GB, so it drops in with masses of room for context, and you can push into 120B-class models or run a 70B at higher precision than any single consumer card allows. Clip two GX10s together over the ConnectX-7 link and you’ve 256GB to play with, enough for really large models at home.

The catch is the same as the Spark: bandwidth. All that memory feeds the chip at about 273 GB/s, so tokens land at a steady walking pace on big models rather than instantly. The headline 1,000 AI TOPS of FP4 compute is huge, but local inference is limited by bandwidth, not maths, so it’s capacity and software that sell this box, not speed. And that software is the treat - native CUDA, the DGX stack, every framework already at home. Just remember it’s Linux only; if you were hoping to run Windows on it, this isn’t that machine.

Who should buy it

Buy it if you’re a developer or researcher who wants Nvidia’s stack and 128GB on your desk without renting cloud time, and Linux is no barrier. For that person the GX10’s ConnectX-7 clustering is the reason to pick it over the DGX Spark - though the 2026 memory shortage has pushed it above the Spark on price, so it’s no longer the cheaper route it launched as.

Give it a miss if you wanted a Windows machine, or if single-stream speed matters more than capacity. A used 3090 runs anything that fits in 24GB far quicker, and a Strix Halo box loads the same 128GB for less if you don’t need CUDA. But for native Nvidia tooling, big models on your desk, and the option to cluster two into 256GB, the GX10 still makes a case - just no longer on price, and only on Linux.