Local and open-source AI, decided before you buy the machine.
A desktop you can order from a shop now runs near-frontier open models. This independent guide looks at the questions that matter: what changed, what can open models do, and when does owning the machine beat renting the API.
Get the guide
Work address, one send, no sequence. The PDF arrives by return mail and nothing else follows.
Sent. The PDF is in your inbox, sources and all.
Machinegeist BV · BE 1042.323.495 · Privacy
What's inside
- Open and local are two different words — who controls the model file versus where your data runs. Most confusion about this subject comes from mixing the two.
- What open models handle today — documents, analysis, internal chat and bulk work, plus the agentic lane where the rented frontier still wins.
- The full cost of owning — why the token math rarely favors the machine, and what renting costs that never reaches an invoice.
- When to bother — buy when a named recurring workload exists over data you will not send out; otherwise, the preparation moves that cost nothing.
- For your CTO — deep dives on the models, the hardware, the serving stack, and security posture, each with dated sources.
The map.
Two axes matter: where inference runs, and who controls the model. Cost and capability are annotations inside the cells, because they change monthly; the axes do not. One structural fact does the organizing: the closed row has a single occupied cell.
| Model control | Public cloud API | Private cloud (rented dedicated infra) | On-premise | Edge device |
|---|---|---|---|---|
| Closed (Opus, GPT class) | The default. Data leaves, to the vendor under contract. Per-token pricing. Today's capability ceiling. No ops. | Not available. A closed model runs only where its maker or its maker's hyperscaler partner runs it; managed single-tenant offerings rent you a private slice of their cloud, never the model file. | ||
| Open weights | Open models behind someone's API. Data still leaves, to whoever hosts. Per-token, cheaper. The open frontier. No ops. | Open models on infrastructure you rent, in a jurisdiction you pick (EU possible). Monthly rental. Capability sized to what you rent. You run the stack, not the hardware. | Open models on hardware you own. Data never leaves. Owned capex, resale value. The strongest single-machine models fit a 256GB desktop today. All ops yours. | Small and mid-size models on laptops and workstations. Data stays on the device. Part of the fleet budget. Per-device configuration. |
What the hardware costs
| Tier | Price today | What it runs |
|---|---|---|
| Mac Studio M5 Max | from €3.029 | Small and mid-size open models |
| Nvidia DGX Spark | $6,950 US list since 2 October; €6.655 incl. VAT at a Belgian reseller | 128GB class; prototyping on the Nvidia stack |
| Mac Studio M5 Ultra, 96GB | €6.649 | Mid-size models; too small for the strongest single-machine models |
| Mac Studio M5 Ultra, 256GB | €11.599 incl. VAT | The strongest models that fit one machine at all |
| Mac Studio M5 Ultra, 512GB | late October per Apple's Belgian store; price unpublished | The larger open models the 256GB tier cannot hold |
| GPU workstation or Linux server | tracks the 2026 GPU market; single-GPU builds up to five-figure servers | Highest throughput; most maintenance |
| Rented EU GPU server (private cloud) | monthly rental, no purchase | Open models sized to whatever you rent |
Capability bands were read on 31 August 2026 and prices on 6 October 2026; both age monthly. The structure of the tables does not. The guide carries the numbers, with a source and a date on each.
Within about nine months, a €10k machine should run what the frontier sells today.
Get the guide
Work address, one send, no sequence. The PDF arrives by return mail and nothing else follows.
Sent. The PDF is in your inbox, sources and all.
Machinegeist BV · BE 1042.323.495 · Privacy