Asus has made a smaller version of its small AI computer, and the spec sheet needs a lot of footnotes to be read honestly. The Ascent GX10, built on Nvidia's DGX Spark platform, now comes with 64 GB of unified memory. Until now the only option was 128 GB. The new model, GX10-GG0062BN, is aimed at developers who want to run open models of roughly 30 to 35 billion parameters on their own desk. It can be pre-ordered in Germany, Austria and Switzerland, but Asus has not yet said what it costs.
What is actually new
The variant pairs the 64 GB of memory with a 1 TB SSD. Asus presents it as a configuration for less demanding workloads, and says it should make working with open AI models more affordable to get into.
The availability picture is thin. Some first partners in Germany, Austria and Switzerland are taking pre-orders. Asus has announced neither an official price nor a start date for shipments. The Geizhals catalogue lists the device at about 6,000 euros, which is a retailer's placeholder and not a price Asus has committed to.
The report does not give a price for the 128 GB model either, so there is no way to work out how much the halved memory saves anyone. Treat "more accessible" as an intention for now.
The hardware underneath
The core is Nvidia's GB10 Grace-Blackwell superchip. It combines a Blackwell GPU with fifth-generation tensor cores and a 20-core Grace CPU on the Arm architecture. The two are linked by NVLink-C2C, which Asus describes as giving them coherent access to a shared pool of memory, with five times the bandwidth of PCIe 5.0.
The 64 GB is LPDDR5x with a bandwidth of 273 GB/s. Both the CPU and the GPU can reach it at the same time. In practice that means an AI job does not have to copy data back and forth between separate system RAM and video memory, a chore that ordinary desktop GPUs make you pay for.
The cooling is identical to the 128 GB model:
- seven fan control modes
- wide cooling fins
- five heat pipes
- two fans measuring 140 x 80 mm
- air drawn in through openings in the bottom of the case
The published specifications contain no independent measurements of temperature, noise or power draw.
How big a model fits
Asus points this configuration at open models in the 30 to 35 billion parameter range and names Qwen3.8-27B as an example. The company is careful to add that the model size you can really run depends on numeric precision, the software stack and the length of the context. Those three variables cover most of what matters, so the 30 to 35 billion figure is a guide and not a promise.
The petaflop footnote
Asus quotes up to 1 petaflop of performance for the GB10. The figure is for FP4 calculations with sparsity.
FP4 is a low-precision number format. It cuts memory and compute needs for tasks that can tolerate it. Sparsity lets the hardware skip some operations that contribute little to the result. Put them together and you get a flattering peak number that does not describe how fast the machine will be on an arbitrary job. Real speed depends on the model, the software and the kind of load.
Two boxes instead of one
The most interesting claim concerns clustering. Asus suggests linking two GX10 units through the built-in Nvidia ConnectX-7 networking. The manufacturer's numbers for the pair:
| Metric | Two linked 64 GB units |
|---|---|
| Combined memory | up to 128 GB |
| AI compute | up to 2 petaflops |
| Aggregate memory bandwidth | up to 546 GB/s |
Asus also says such a pair can deliver up to 1.7 times the performance of a single 128 GB GX10. The company does not say which metric it used for the comparison or under what test conditions, so the figure is hard to check and harder to rely on.
Setup goes through the Nvidia Sync App, which includes a Cluster Assistant. Asus says joining two machines takes a few minutes and requires no change to the software environment or the usual workflow. It also states support for Nvidia NVFP4 and for running compressed versions of larger models, Qwen3.8-Flash-Next among them.
The suggested path is to start with one unit for development and testing, then add a second when the models grow, the context gets longer or more AI agents run at once. Models, prompts and data can stay on local hardware throughout.
One detail deserves a raised eyebrow. Two 64 GB machines add up to the same 128 GB that a single older GX10 already offered. Which route is cheaper, or whether the pair is simply the slower way to the same memory, cannot be settled without prices for both, and the report has none.
Software and the local-first argument
The software bundle includes Nvidia DGX OS, PyTorch, Jupyter and Ollama. Asus also claims support for Nvidia OpenShell and Nvidia NemoClaw, tools for setting security policies for AI agents. The machine can be managed remotely over SSH, so an IT team can deploy and monitor several units at once.
Asus frames the whole product line around three pressures: rising token consumption, growing cloud computing bills and the need to protect confidential data and intellectual property. Run a model locally and the data and the tokens it generates may never pass through a public API.
The word "may" is doing honest work there. Whether local hosting makes sense for a given team depends on the model it wants, the memory the machine has and the software it runs.
What is still missing
- An official price. The only number so far is a catalogue listing of about 6,000 euros.
- A shipping date. Pre-orders exist, a delivery schedule does not.
- Independent measurements. Temperature, noise and power consumption are all unreported.
- A defined benchmark. The 1.7x figure has no stated metric or test setup.
For the moment the 64 GB GX10 is a spec sheet with a retailer's price tag taped to it.