Private LLM hosting in your own cloud account, with nothing leaving it
Open-weight models on a dedicated inference endpoint in your account, your region or your hardware, for data that may not leave. Sized, locked down and monitored, with the bill in your name.
No list price. It starts with a diagnostic at a fixed fee, and the build is quoted in writing after it.
When you need private inference
And when you do not.
A dedicated endpoint for the models your systems use: in your account, in your region, or on your own hardware. We size it, run it and report on it. Nothing a system sends it leaves the boundary you set.
You need it when
- Data that may not leave your account, your region or your building.
- Volume where a dedicated endpoint costs less than paying for each call.
- A fine-tuned model of your own that has to be served somewhere.
- Latency that a shared service cannot promise.
You do not when
- Low volume and no residency rule: a hosted model through your own provider account is simpler and cheaper.
- A task that needs a frontier model no open-weight model matches yet. We say so.
What private inference includes
Part by part.
- 01
The boundary is drawn
Your cloud account, your region or your hardware, and what may cross the line, written on one page.
- 02
The model is chosen
An open-weight model that fits the task, the licence and the hardware, with a smaller fallback.
- 03
Capacity is sized
From your measured volume and latency target, with headroom and a plan for peaks.
- 04
The endpoint is built
Serving stack, batching, caching and rate limits, behind an interface your systems already speak.
- 05
Access is locked
Private networking, dedicated credentials with the minimum permissions, every call logged.
- 06
Cost is visible
Usage, utilisation and spend reported each month, against the figure the quote set.
How private inference is proved
A locked test set and a pass mark, agreed before the build.
- Residency: a test that traces where every call is processed and stored, run before launch and after every change.
- Latency and throughput under your measured load, against the target set in the quote.
- Quality: the endpoint scored on the same test set as the system that uses it.
How private inference is run
After launch: monitoring, updates and what you own.
- Monitoring: errors, saturation and latency raise an alert; a failed instance is replaced without a call from you.
- Model and security updates applied on a schedule, re-scored before release.
- Handover: the infrastructure as code, the runbook and the dashboards, in your account from day one.
What private inference will not do
Limits we set on purpose.
- It does not send anything across the boundary you drew, and the data map says so in writing.
- It does not claim your data stays in Canada unless it does.
- It does not resell compute at a hidden margin: usage is billed in your account, or at cost plus the published markup.
- It does not run a model whose licence does not allow your use.
Questions about private inference
Whose account is it in?
Yours. The endpoint, its credentials and its bill are in your name from day one. If you stop the run fee, it keeps running without us.
Can it stay in Canada?
Yes, when the model and the capacity are available there. We say which option each part uses before we build, and we never claim your data stays in Canada unless it does.
What does it cost to run?
The compute is billed in your account at the provider’s price, or through us at cost plus the published markup. The run fee covers monitoring, updates and the monthly report. The figures are on the pricing page.
Can our own fine-tuned model run on it?
Yes. Serving a model of your own is the usual reason to want a private endpoint. Custom models and private inference are quoted together when both are needed.
Talk to us about private inference
Fifteen minutes on a call is enough to tell whether it fits your work. If it does not, we say so.