Why some companies run artificial intelligence on their own hardware, what they need and when it isn't worth it
When a company uses ChatGPT, Claude or Gemini, what it types travels to that provider's servers. For many tasks that's not a problem. For others it is: contracts, customer data, medical records, financial information. That's where an option that was impractical a couple of years ago comes in: running the model on a computer of your own.
Why some companies do it
- Privacy. The data never leaves the company. For regulated sectors, or for clients who require it by contract, it can be the difference between using AI or not.
- Predictable cost. With high, steady use, paying for the hardware once can work out cheaper than paying per request.
- Independence. If the outside service changes prices, terms or goes down, your own model keeps running.
Which model
There are open models you can download and use without paying for a licence. One of the most widely used is Qwen, from Alibaba, which comes in different sizes. The versions with around 27 billion parameters, compressed, fit on a single graphics card and are enough to summarise documents, sort emails or answer questions about internal information.
Which machine
- A PC with a powerful graphics card, such as a 32 GB RTX 5090. It's the fastest option for a model that size.
- A Mac with 48 GB of memory or more. Quieter and simpler to maintain, somewhat slower.
- A dedicated machine such as NVIDIA's DGX Spark, with 128 GB, for larger models or for several people to use at once.
To run them, programs like Ollama or LM Studio are enough; they install like any other application.
When it isn't worth it
With light use, cloud services are still cheaper and more capable. You also have to count the time of someone who installs, updates and maintains the machine: it's the most underestimated cost. And the models that fit on one machine are good, but they don't yet match the largest ones on complex tasks.
For many companies the answer is a mix: sensitive work on your own model, the rest in the cloud.
Does your company handle information it would rather not send outside? Get in touch and we'll see whether a model of your own makes sense in your case.




