Asus Ascent GX10 a perfect local AI computer

Imagine an AI powerhouse that fits on your desk, small and light yet delivering performance that previously required an entire server room. The Asus Ascent GX10 is not only impressive, it is downright ridiculously powerful, and it’s time to talk about why this little box can change the way you work with AI locally. I’ve been testing it for a while since we have one at work, and all that can be said is that it delivers.

Petaflops on the desktop

Not so long ago, a petaflop belonged to powerful data centers, cooled by industrial fans and surrounded by security doors. Today, you can put it on the desk next to your coffee mug. The Asus Ascent GX10 is powered by Nvidia’s GB10 Grace Blackwell Superchip, the same chip used in Nvidia’s own DGX Spark, delivering up to 1 petaFLOP AI performance in FP4. That’s a number so large it almost feels a bit reckless to have at home.

The machine measures a modest 150 x 150 x 51 mm and weighs 1.48 kg. So it is about the size of a thick sandwich bread, but significantly more useful if you work with large AI models.

One important thing to know is that no, it doesn’t beat a dedicated graphics card like the Nvidia 5090 in speed, but an Nvidia 5090 cannot load large models like this AI computer can.

Jensen Huang about local AI

Nvidia’s founder and CEO Jensen Huang is not known for understatement, and when it comes to local AI infrastructure he has expressed himself clearly. At the World Economic Forum in Davos 2026 he said:

“AI is all about infrastructure. You should have local AI as part of your infrastructure.”

His message is crystal clear: AI is no longer an experiment or a luxury tool for tech giants, it is a fundamental resource that individuals, companies and countries need to own and control themselves.

It is precisely from this perspective that the Asus Ascent GX10 starts to make real sense. If AI is to be your infrastructure, why rent it from the cloud when you can own it on your desk?

It is small and really good looking

AI-class tech gadgets usually look like they were built by someone with no sense of aesthetics who got the entire IKEA catalog wrong. Not here. Asus has designed the GX10 with matte side panels and a cross-hatched mirrored lid on top, and IT Pro reviewers say that the GX10 is by far the best-looking option among all GB10 variants on the market.

One detail that actually matters is that the power switch is located on the front. That sounds trivial until you realize that neither Nvidia’s own DGX Spark nor Dell’s version has that, and if you plan to stack multiple units in a cluster, you won’t have to dig around in the back every time.

This little beast is also very quiet

Powerful computers and deafening fan noise usually go together like a bad chemistry couple. Not here either. The GX10 is remarkably quiet, even under heavy load. The fans operate at a level that can almost be described as a loud whisper, which is an achievement for a system delivering petaflop-level AI calculations.

The cooling system consists of a dual-fan arrangement with seven adjustable levels for airflow control, and Asus claims that the thermal coverage area is 1.6 times more efficient compared to similar compact systems. Air is drawn in from below and the front through the wavy ventilation slots and expelled through the large air outlet at the back. The chassis can get somewhat warm in spots under extremely heavy workloads, roughly like a laptop under strain, but never in a way that causes concern. It’s impressive cooling technology packed into a form factor that doesn’t sound like a jet at takeoff.

The operating system is Linux of course

GX10 runs Nvidia’s DGX OS, which is essentially Ubuntu 24.04 LTS customized for ARM architecture. This may sound intimidating for those used to Windows, but it is actually one of the machine’s secret strengths.

Why does the operating system matter? Because the entire AI ecosystem is built on Linux. CUDA, PyTorch, TensorFlow, vLLM, and practically all serious AI frameworks are optimized for this environment. Running these tools on Windows is entirely possible, but it’s like driving a Formula 1 car on a gravel road. It works, but it’s not what it was built for.

Nvidia’s DGX OS comes with a complete dashboard to manage AI projects, the Nvidia Sync tool for SSH keys and tunneling, as well as all relevant drivers preinstalled. Additionally, the platform is directly connected to Nvidia’s GitHub repository for DGX Spark Playbooks, a collection of ready-made recipes to quickly get started with models like ComfyUI, PyTorch, and vLLM. Support is also available for CUDA, TensorFlow, Jupyter, and DeepSeek R1 inference, making the entire AI stack ready the moment you plug in the power cord. For those new to the AI world, this is a shortcut that truly delivers on its promise.

Advantages of Asus Ascent GX10

• 1 petaFLOP AI performance in a format that fits in a suitcase
• 128 GB LPDDR5X RAM enabling running large language models locally without the cloud
• Two 200 Gbps ConnectX-7 network ports enabling clustering and scalability
• Power switch on the front, a small detail that makes a big difference in cluster environments
• The best design of all GB10 variants on the market
• Extremely quiet cooling with dual fans and seven adjustable levels, barely audible even under heavy load
• DGX OS on Linux provides a fully optimized AI environment with all software pre-installed
• Access to Nvidia’s DGX AI playbooks for a smoother startup
• The SSD storage is upgradeable, making cheaper base configurations attractive

Disadvantages of Asus Ascent GX10

• The price is painful, starting from just over 3,000 dollars and up can sting for a private individual… you know income tax, VAT as the cherry on top… but for a company this is a real bargain.
• Runs Ubuntu Linux on ARM, not Windows, which may discourage some
• Everything important is soldered, the memory cannot be upgraded
• Only supports 2242-size SSDs, not the more common 2280 size
• Nvidia’s platform limits the options, all GB10 variants are very similar

Asus Ascent GX10 compared with Nvidia DGX Spark

The conclusion from the comparison is clear. The technology is essentially identical since Nvidia strictly controls the platform specifications, but Asus offers a lower starting price on the base model, a sleeker design, quieter and more efficient cooling, as well as the practical detail of the power button on the front. That is enough for many developers to prefer the GX10 over the original.

Performance that makes ordinary computers sad

Benchmark tests in GeekBench 6 show 3,104 points in single-core and 20,048 in multi-core. A Mac Mini M4 Pro actually beats that in these tests, but that is not really what the GX10 is designed for. The really exciting part happens when running AI workloads. A fine-tuning of Pytorch LoRA Llama3 8B reached up to 53,000 tokens per second as a peak score. Image generation in ComfyUI was completed in 216 seconds, while a comparable machine with 32 GB RAM needed over 400 seconds for the same task.

200 gigabits per second and scalability smiling towards the future

What really puts the GX10 in a league of its own is the network port. Two 200 Gbps ConnectX-7 QSFP112 ports are located on the back and the network card alone costs around 15,000 to 20,000 SEK to purchase separately. These ports make it possible to connect two GX10 units and run models that otherwise wouldn’t even fit on a single machine, like Llama 3.1 with 405 billion parameters.

Who is the target audience for this computer?

Asus Ascent GX10 is built for AI developers, data scientists, and tech enthusiasts who want local computing power without turning their living room into a server room. It is not a machine for those who just want to browse the internet and watch series. But for those seriously working with local AI models, training, inference, and experimentation without relying on the cloud, the GX10 is currently one of the best solutions on the market.

Asus Ascent GX10 is the winner of the Taiwan Excellence Award 2026, the most stylish among its competitors and packed with technology that would cost three times as much if you assembled it yourself. It’s not cheap, but as Jensen Huang would say, AI is infrastructure, and good infrastructure has never been free.