Private AI on hardware you control
Private AI means running language models, document search and agents on machines your organisation controls, so prompts and documents are processed on your own hardware instead of a cloud AI provider's. It gives you control over where data goes. It does not make a system secure or compliant by itself; that still depends on how it is set up and run.
What stays local, and what may not
The boundary is a design decision, not a default. We agree it with you before anything is built, and record it at handover.
Stays on your hardware
- The model weights, once downloaded.
- Prompts, responses and conversation history.
- Your documents and the search index built from them.
- Logs and usage records, when configured to stay local.
May still leave your network
- Software updates and model downloads, unless they arrive by offline transfer.
- Calls to outside tools or APIs that an agent is allowed to make.
- Remote support sessions, only if you grant access.
- Backups, if they are sent to a cloud service.
Three ways to deploy it
Start with who will use it and where the machine can live. Hardware sizing follows from the models and users.
| Deployment | Suits | Where it lives | Start with |
|---|---|---|---|
| Single workstation or tower | A person or a small team, one main use case | An office or desk-side; quiet enough for a shared room | Local LLM servers |
| Departmental server | A department sharing models and documents | A comms room or a quiet corner, managed by IT | AI agent servers |
| Rack server | Many users or applications across the organisation | A server room, data centre or colocation rack | AI servers |
Private AI or a cloud service: the honest tradeoffs
Many organisations use both: private AI for sensitive documents and internal tools, a cloud service for everything else.
| On your own hardware | Cloud AI service | |
|---|---|---|
| Where data is processed | On machines you control | On the provider's infrastructure, under their terms |
| Models available | Open-weight models you can download and run | The provider's models, including closed ones that cannot run locally |
| Upfront cost | Hardware purchase and setup | None |
| Running cost | Power, space and someone's time; no per-use fee | Billed per use, which grows with usage |
| Scaling up | Add or replace hardware, with lead time | On demand |
| Maintenance | You, or us under a scoped support arrangement | The provider |
| Without internet | Keeps working on the local network | Stops working |
How a private AI rollout works
Small first, then expand. A pilot with one use case tells you more than any specification sheet.
Pick the first use case
One workload, one team, and a result you can check, such as answering questions over a policy library.
Size the hardware
Models, context length and simultaneous users decide the GPU memory and the form factor.
Set the boundary
Which data goes in, who can reach the system, and what, if anything, may leave the network.
Build and test
Burn-in under sustained load, then the software stack installed and checked on your own documents.
Pilot with a small group
Real users, real questions. Adjust models, prompts and access before widening it.
Expand or adjust
Add users, use cases or hardware once the pilot shows what is needed.
Common questions
Is private AI automatically secure?
No. Running models locally keeps processing on your hardware, but security still depends on access control, network design, updates and who can reach the system. We set it up with your IT team and document how it is configured.
Does running AI on our own hardware make us PDPA compliant?
No system makes an organisation compliant by itself. Keeping data on your own hardware can help you meet your obligations, but compliance depends on your policies and processes. Take legal advice for your situation.
Can open models do what cloud models do?
For many tasks, such as answering questions over documents, summarising, drafting and coding help, open-weight models are capable. The largest cloud models still lead on some tasks. Try your real task on an open-weight model before you decide.
Can it run without an internet connection?
Yes. Once installed, models run over your local network. Software updates and new models then arrive by offline transfer. Fully isolated (air-gapped) setups are scoped per project.
What does a private AI system cost?
It depends on the models, the number of users and where the machine will live, from a single workstation to a rack server. Prices are quoted per project; the advisor is the quickest way to an indicative recommendation.
Plan your private AI system
Tell us the use case, the users and where the machine can live. We recommend a deployment model and the hardware for it, and explain why.