AI computer for the business: four concrete purchases for local AI

In Run AI locally you will find the table of which model sizes require how much memory. Here is the next step: what you can actually buy in Sweden, in four paths by need and budget. The rule of thumb from the local AI world applies all the way: the amount of memory decides which models you can run; performance only decides how fast it goes.

Transparency first: the links to retailers below are advertisement links via /ut/. The selection is our own and is based on memory per krona, not commission. Prices as of July 2026; always check current price with the retailer.

Path 1: Pre-built top class (NVIDIA)

Webhallen Config Ultra with RTX 5090: Ryzen 7 9800X3D, RTX 5090 with 32 GB VRAM, 64 GB RAM, 2 TB SSD, 3-year warranty. 32 GB VRAM runs 27–32 billion-class models entirely on the graphics card; that is where local AI begins to compete with cloud quality. Webhallen’s Config series has been built since 2012, stress-tested before delivery, and comes without pre-installed bloatware.

For whom: the agency or consultancy that wants the strongest local AI machine available pre-built, and that will also train smaller models or run image generation.

Path 2: Good-value AMD build

Komplett-PC Epic Gaming a255 RGB: Ryzen 7 7800X3D, Radeon RX 9070 XT with 16 GB VRAM, 32 GB RAM, 2 TB SSD. AMD cards now work fine for local AI: Ollama and LM Studio run them via Vulkan/ROCm without fuss. 16 GB VRAM covers the whole mid-range (12–14 billion) with good margin and also handles quantized larger models partially offloaded to system memory.

For whom: the small business that wants a solid all-round step into local AI at about half the top-class price.

Path 3: Portable workstation (128 GB unified memory)

HP ZBook Ultra G1a at Dustin: AMD Ryzen AI Max+ PRO 395 with 128 GB unified memory (43,999 kr for the 128 GB/4 TB configuration at the time of writing). The unified memory is shared between processor and graphics part; up to 96 GB can be allocated to the graphics part according to HP. In practice, it is Windows’ answer to the Mac shortcut: a laptop that runs 70 billion-class models, something no portable graphics card can do. On the same Dustin page there are cheaper alternatives with the same architecture, such as ASUS ProArt with 128 GB for around 32,000 kr.

For whom: the lawyer, healthcare provider, or consultant who handles sensitive material and wants maximum local AI capacity in the bag.

Path 4: Custom build

Shark Gaming’s Mighty Shark series: built to order with 1–3 days build time. At the top, Almighty Shark (Ryzen 9 9950X3D, RTX 5090 32 GB, from 64,999 kr); in the mid-range RX 9070 XT machines from around 26,600 kr. The point of the builders is customisation: ask them to put in 96 or 128 GB RAM, it is cheap and makes a big difference when a model does not fit in VRAM and has to spill over into system memory.

For whom: those who want to choose exact component level without screwing themselves.

Just the graphics card? Three cards that go far

If you already have a decent desktop computer, swapping the graphics card is the cheapest path to local AI. Counted in VRAM per krona:

The used tip from the previous guide still stands: an older RTX 3090 with 24 GB beats a newer card with 12 GB for this purpose. And remember to check that the power supply can handle it; an RTX 5090 wants 1,000 W with margin.

Match memory to the model

This is the whole point of the memory hunt: this is how much model you get per memory level. Sizes apply to Q4-quantised versions (the default format in Ollama and LM Studio), approximate values.

Two practical notes. A model needs headroom beyond the file itself: context memory grows with long documents, so aim for models that leave a few GB of margin. And if the model does not fit in VRAM, Ollama/LM Studio automatically spills over to system memory; it works, but speed collapses, which is why 64 GB+ RAM is a cheap safety net in path 1 and 4. For Swedish, Mistral and Gemma families remain the first-choice tip.

Choose in 30 seconds

• Most capacity, desktop: path 1 or 4 (32 GB VRAM).
• Best price/performance: path 2 (16 GB VRAM).
• Portable with sensitive data: path 3 (128 GB unified memory).
• Cheapest path: keep the computer, swap the graphics card.

Software next? Back to Run AI locally: LM Studio or Ollama, and you are up and running the same evening. If the machine is to be shared by several people in the company, operations tools are in the same guide.