Artificial intelligence no longer lives exclusively in the cloud. Today, it’s possible to build a workstation capable of running advanced large language models locally for less than €2,000. And for many use cases, it makes more sense than you might think.
Until recently, talking about artificial intelligence meant talking about the cloud.
Every request went through an API. Every document was processed on a remote infrastructure. Every new project came with costs directly tied to usage.
That was the natural model, and for many scenarios it still is.
Over the past few months, though, I’ve noticed that the conversation is changing.
The question is no longer just Which model should I use? or Which API is the cheapest?
Increasingly, it’s this:
Does it still make sense to rely on the cloud for everything?
From where I stand, the answer is no.
Not because the cloud is becoming less important. Frontier models will continue to require massive infrastructure, and cloud services will remain essential.
But a growing number of everyday tasks can now be handled directly on your own machine, keeping documents, source code, and sensitive data under your control.
That is, in my opinion, one of the most interesting shifts happening in AI today.
Why now?
Three factors are driving this change.
The first is the models themselves.
Families such as Llama, Qwen, Mistral, Gemma, and DeepSeek have shown that open-weight models are now capable enough for real-world tasks, including software development, document analysis, retrieval-augmented generation (RAG), personal assistants, and workflow automation.
The second factor is the software.
Tools like Ollama, LM Studio, and llama.cpp have dramatically lowered the barrier to entry. Running a local language model now takes minutes instead of hours.
The third factor is hardware.
The second-hand market has made high-end graphics cards much more affordable. As a result, the balance between cost and performance has shifted significantly.
The most important resource isn’t the CPU
When people build a computer, they usually start by looking at the processor.
For large language models, however, the critical component is usually GPU memory.
Every model must be loaded into VRAM before it can run. If it doesn’t fit, part of the workload moves to system memory, causing a noticeable drop in performance.
That’s why a GPU with 24 GB of VRAM represents one of today’s best sweet spots.
Thanks to quantization, it’s possible to run much larger models than their original size would suggest.
In practical terms:
- 7B–8B models run comfortably.
- 14B–22B models are currently the most interesting range for professional work.
- 70B models require more advanced setups, often involving multiple GPUs.
Model size, however, isn’t everything.
A smaller model optimized for a specific task can outperform a much larger general-purpose model.
Desktop or laptop?
For local AI workloads, a desktop is still the most practical choice.
The reason is simple: VRAM.
Laptop GPUs often include less memory than their desktop counterparts. For example, an RTX 4090 laptop GPU typically offers 16 GB of VRAM, while the desktop version provides 24 GB.
That difference directly affects the size of the models you can run.
A desktop also offers better cooling, easier upgrades, and a far more attractive second-hand market.
The trade-off is portability and higher power consumption.
Three AI workstation builds under €2,000
RTX 3090: the sweet spot
If I had to recommend a single configuration today, this would be it.
A used RTX 3090 with 24 GB of VRAM still delivers one of the best price-to-performance ratios available.
With a budget between €1,400 and €1,700, you can typically build a system with:
- Nvidia RTX 3090 (24 GB)
- AMD Ryzen 9 or Intel Core i7
- 64 GB RAM
- 2 TB NVMe SSD
It’s an excellent setup for developers, consultants, and professionals who want to run local models on a daily basis.
Refurbished professional workstation
If you’d rather avoid building a PC from scratch, refurbished enterprise workstations are worth considering.
Systems such as the Dell Precision, HP Z, or Lenovo ThinkStation are designed for continuous workloads and offer excellent reliability.
Equipped with an RTX A5000 (24 GB), 64–128 GB of RAM, and fast NVMe storage, they often fit within the €2,000 budget.
They’re especially well suited for small technical teams or professional environments.
Dual RTX 3090
If your goal is to experiment with larger models, a dual RTX 3090 setup provides 48 GB of total VRAM.
That opens the door to significantly larger quantized models.
The downside is greater complexity: higher power consumption, better cooling requirements, a larger power supply, and a more advanced software configuration.
It’s a powerful option, but not one I’d recommend as a first build.
Buying used hardware
The second-hand market can dramatically reduce costs, but it requires some attention.
Many GPUs have previously been used for cryptocurrency mining.
That doesn’t automatically make them a bad purchase, but it’s worth checking temperatures under load, fan condition, and the overall physical state of the card.
Whenever possible, buying from a professional seller with a warranty provides additional peace of mind.
Privacy isn’t the only advantage
Local AI is often discussed in terms of privacy.
That’s certainly important.
But I think the biggest advantage lies elsewhere.
AI agents are changing how we interact with language models.
They don’t perform a single request.
They plan, read documents, write code, verify results, fix mistakes, and repeat the process.
When every step goes through a paid API, costs can grow quickly.
Running part of that workload locally changes the economics.
It doesn’t replace the cloud.
It simply allows you to use the cloud where it adds value and keep everything else under your own control.
Final thoughts
For years, the cloud was the only realistic option.
Today, that’s no longer true.
With a budget below €2,000, it’s possible to build a workstation capable of running advanced language models locally and handling many everyday AI tasks without depending entirely on external services.
It won’t be the right solution for everyone.
But I have the feeling we’re witnessing the beginning of a shift similar to what happened with personal computers.
At first, they were tools for enthusiasts.
Eventually, they became something everyone expected to have.
I wonder whether, a few years from now, having a personal AI model running on your own workstation will feel just as ordinary.
