Datalumo is light. The AI models do the heavy lifting, and those run on the provider's servers, not yours. Your server mostly stores content and routes requests.
For most sites, on the default SQLite stack:
- 2 vCPU, 2 GB RAM, 20 GB disk. A $12 to $20 per month VPS.
- Docker and Docker Compose.
- A domain with a DNS A record, and ports 80 and 443 open.
This comfortably handles a knowledge base of tens of thousands of documents and steady widget traffic.
- Trying it out: 1 vCPU and 1 GB RAM is enough. Your laptop works too.
- Large or busy: 4 vCPU and 8 GB RAM, on Postgres, once you have hundreds of thousands of documents or several busy instances on one box. Add disk to match your content.
Those numbers assume your AI provider runs the models. Datalumo does not ship with a model. If you run one locally instead (the custom provider with Ollama, vLLM, or LM Studio), the model becomes the dominant cost, not Datalumo, and you size the box for it:
- RAM: several GB for a small model, much more for larger ones. The model has to fit in memory.
- Speed: a GPU or Apple Silicon for usable latency. CPU-only works but is slow.
- Disk: a few GB per model.
Check the model's own requirements and add them on top of Datalumo's modest footprint. A small chat model plus an embedding model is comfortable on a 16 GB machine; larger models need more.
Any provider that runs Docker works: Hetzner, DigitalOcean, Linode, Vultr, AWS Lightsail, a Hetzner or OVH dedicated box, or your own hardware. Pick one near your users to keep responses snappy.
For running several customers or sites on one server, see Reverse proxies and multiple instances.