Ninja
Open the console

Siting 04 · the boundary

You draw the line.

Most AI products decide for you that your data goes to their cloud. Ninja is shipped so that the boundary is a deployment choice, all the way to a network that has no route out.

Air gap capable
T+00:00Four ways to site it
SitingWho runs the machinesWhere your data sitsSuits
Managed serviceNinjaNinja's cloudGetting to a first working run this week.
Your virtual private cloudYou, in your AWS, Azure or Google Cloud accountInside your account, under your network policyA security review that ends in a data residency clause.
On premisesYou, on your own hardwareYour buildingRegulated work, or a cloud policy you cannot move.
Air gappedYou, with no route outA network with no internetDefence, critical infrastructure, anything where the answer is simply no.
T+00:40What crosses the boundary

In a private
deployment,
nothing does.

The orchestrator, the run's machine, the memory and the model can all sit inside the same perimeter. What leaves is what you decide leaves.

Your perimeter OrchestratorMemory and state Run machineModel weights plans, schedules, retriescheckpoints, logs, artefacts shell, files, browseropen weight, your GPUs Operator console only
Diagram 01 · private deployment · schematic
T+02:15On your own machines

Capacity you
own beats
tokens you rent.

When the model runs on hardware you already pay for, a long run stops being a cost decision. Nobody shortens a migration to save tokens. The team stops rationing the thing you bought it for, which is the difference between a pilot and a deployment.

Open-weight models on dedicated GPUs, frontier models through your own provider keys, or both in the same run.

A

Metered

Per token, per seat, per action. The finance team learns to ask people to use it less.

B

Owned

Per machine, per month. The only limit is how much work you can find for it.

T+06:00What the security review asks

The questions,
and the short
answers.

QuestionAnswer
Does our data leave our estate?Not in a private deployment. In the managed service, you set what is sent.
Is our work used to train a model?No, on the paid tiers you can opt out, and in your own deployment the question does not arise.
Single sign-on and provisioning?Enterprise sign-on is part of the enterprise tier.
Who can approve an action?Named people, per workspace, with the action types you choose.
Can we read what it did?Yes. Every step and input is logged and kept after the run.
What happens on the network we cannot open?Air-gapped deployment, with open-weight models on your hardware.

Send this to your security team