The architecture
The four boundaries a private deployment has to hold
Confidentiality is not a setting you switch on at the end; it is four separate boundaries, each of which a public API crosses by design. A private deployment exists to hold all four.
01
Where the data sits
Documents and prompts stay inside your tenancy and region, so residency is a fact about storage rather than a term in a contract.
- Your tenancy
- Region
02
Where the model runs
The weights are hosted on infrastructure you control, so inference happens where the data already is instead of being shipped to it.
- Self-hosted weights
- GPU
03
Who can reach it
Network isolation and access control mean the model answers only callers you allow, with no path out to a public endpoint.
- Isolation
- Access control
04
What is written down
Every prompt and answer is logged where you can review it, so an audit or an incident has a record rather than a black box.
- Audit log
- Retention
The public-API shortcut
Calling a public model is fast to build and crosses all four boundaries at once: the data leaves, the weights are someone else's, the caller is on the open internet, and the log lives outside your reach. It suits public information; it does not suit the documents that made you read this page.
A model is confidential only when all four boundaries hold; crossing any one of them is a data transfer.
What an API call really is
The document that leaves the building
The phrase "we use AI" hides a physical fact: when a public API is called, the text of the prompt — the contract clause, the customer record, the internal memo — travels out of your network to a machine you do not own, in a jurisdiction you did not choose. For most content that is harmless. For the material a regulated organisation holds under residency and confidentiality obligations, it is the exact transfer those obligations exist to prevent, and no amount of the provider's good intentions changes where the data went.
A private deployment changes the physics rather than the promise. The model runs where the data already lives, so answering a question no longer requires exporting the question. The trade is real: you own the infrastructure, the capacity planning and the upgrade path that a public API absorbs for you. That is a cost worth naming, and for a great deal of general work the public API remains the right answer. The private path is for the material where the boundary is not negotiable.
Before it holds your data
What a deployment must satisfy before a regulator's data touches it
These are the conditions that separate a model kept private from one merely installed on a nearby server. Each is a boundary a reviewer will ask about.
Inference inside your boundary
The model answers on infrastructure you control, so no prompt is transmitted to a third party to be served.
What it protects: The confidentiality of every question asked of it.
Data kept in-region
Storage and processing stay in a location residency rules permit, verifiable rather than assumed.
What it protects: Compliance with where the data is required to live.
Access scoped per user
A person sees only what their role already permits; the model does not become a way around existing permissions.
What it protects: The access model you already enforce elsewhere.
A retained audit trail
Prompts and answers are logged and kept for the period review requires, so an incident is reconstructable.
What it protects: The ability to answer "what did it say, and to whom" after the fact.
Our role
What we stand up and what stays under your governance
- 01
The right model for the boundary
We select an open-weight model whose capability fits the task and whose licence and size let it run inside your infrastructure, rather than defaulting to the largest name.
- 02
The deployment, hardened
We stand up the serving, the isolation and the logging so the four boundaries hold, and document them for the reviewer who will ask how.
- 03
A system your team operates
We leave the runbook, the upgrade path and the capacity plan with your people, because a private model only they can run is a dependency dressed as independence.
Which data is confidential enough to warrant a private deployment, and which is fine for a public API, is your governance call; we build the boundary that makes the first choice real and honest about the trade it carries.
Common questions
What buyers ask
Can we run a large language model on our own servers, or does it have to be a cloud API?+
You can run one on your own infrastructure. Open-weight models — ones whose parameters are published for you to host — have reached a capability that covers most enterprise tasks, and they run on hardware you can put inside your own tenancy or region. It is more work than calling a public API: you own the serving, the capacity and the upgrades. The reason to accept that work is when the data cannot be allowed to leave your boundary, at which point a cloud API is not a lighter version of the same thing but a different decision about where your documents go.
Which open-weight models are good enough to replace a public API?+
It depends on the task, and the honest answer is that we choose it against your workload rather than from a leaderboard. For grounded question-answering over your own documents, retrieval carries much of the weight and a mid-sized open model performs well. For open-ended reasoning across long, ambiguous problems, the gap to a frontier public model is real and worth measuring before you commit. We test candidate models on a sample of your actual questions, so the choice is made on evidence from your workload, not on a benchmark someone else designed.
Does keeping the model private mean we give up the quality of the frontier models?+
On some tasks there is a gap, and we will not pretend otherwise — that is one of the two claims this page refuses to make. But the gap is narrower than the headlines suggest for the work most enterprises actually need, especially where answers are grounded in your own content rather than generated from the model's general knowledge. The practical route is a split: keep the confidential workloads on a private model where the boundary matters more than the last increment of quality, and use a public API for general, non-sensitive tasks where it does not. The decision is per workload, not once for the whole organisation.
In this cluster
Next step
Scope it with us
A deployment plan that names where the data sits, where the weights run and who can reach them, before any model is stood up.

