In short: Since June 2026, my office has been running an ASUS Ascent GX10, a compact AI computer with an NVIDIA GB10 chip and 128 GB of memory shared by CPU and GPU. On it I run local language models with Ollama and use them via Open WebUI or directly through the API from my own tools. Nobody needs this for occasional chat questions. It gets interesting when you want to test many models, keep data in-house and build AI into your own workflows.
Why your own hardware at all?
In my article Local AI or Cloud I described the criteria I use to decide where an AI task should run. Three reasons tipped the balance towards my own hardware:
- Experimenting: trying different models and variants without paying for every request
- Data in-house: code, documents and product data don't have to go to a third-party provider for processing
- Integration: a local API that my own scripts and tools can use at any time
The device: ASUS Ascent GX10
The GX10 is based on the NVIDIA GB10 platform. Its special feature is the memory: 128 GB shared by processor and graphics unit. So there is no classic graphics card with separate video memory.
That is exactly what makes it interesting for language models. Large models need a lot of memory, and on normal hardware video memory is usually the limit. With 128 GB, models fit that hardly run on an ordinary PC.
What many underestimate: lots of memory does not automatically mean high speed. How fast a model responds depends on its size, quantisation, context length and answer length. Models where only part of the network computes for each word, so-called mixture-of-experts models, are noticeably faster on this device than models of the same size where everything always computes.
On price: I paid around €3,500 for the device. Prices for this kind of hardware have risen considerably since. For comparison, AI servers for enterprise use play in a completely different league and quickly cost many times as much.
The software behind it
The core is Ollama. It manages the models, loads them into memory when needed and provides a local API. On top of that, Open WebUI offers a browser interface. That way I can try models in a chat and call them from other programs at the same time.
The setup is deliberately simple:
Browser / IDE / own scripts
|
Open WebUI / API client
|
Ollama
|
local language model
|
ASUS Ascent GX10Very important: the API is not openly reachable from the internet. When I'm away, I only access my network via VPN. An unprotected model API would be an open door, even if the model runs locally.
Which models interest me
I test models from different families, including Qwen, GPT-OSS, Llama and DeepSeek. The name matters less to me than whether a model delivers useful results on real tasks:
- understanding and explaining PHP code
- answering technical questions in German and English
- following instructions precisely without making things up
A practical test of my own is more useful for that than any leaderboard. I rate answer quality, speed, memory needs and above all whether I want to work with it day to day.
What your own AI server doesn't solve
- Models still make mistakes. A local model can make false statements just like a cloud model.
- Operating it takes time. Updates, power, backups and network maintenance are part of it.
- Memory is finite. Loading two large models at once pushes even 128 GB to the limit. If you don't keep an eye on it, you bring the machine to a halt.
- The top models of the big providers don't run locally. For the most demanding tasks, the cloud often stays ahead.
If you only occasionally have a question for the AI, a cloud service is much simpler.
Conclusion
For me, the GX10 is mainly a server for experimenting and working. It enables independent tests, a local API and my own AI workflows. It gets really interesting as soon as not just one chat window but several tools access it at the same time.
Frequently asked questions
Do I need a machine like the GX10 for local AI?
No. Smaller models also run on a good PC or laptop. For large models with many billions of parameters, however, you need a lot of fast memory, and that is where such a device shows its strength.
What is the difference between Ollama and Open WebUI?
Ollama runs the models and provides the API. Open WebUI is a browser interface for chatting with the models. Other programs can use Ollama without Open WebUI as well.
Can the GX10 run several models at once?
Yes, as long as they fit into memory together. A very large model quickly takes up most of the 128 GB, though. Then another model has to be unloaded before the next one can start.