The shift
For a decade the sensible default was to rent. Compute was elastic, the workload was spiky, and owning hardware meant owning a problem. That logic holds for experiments.
It holds less well for a workload that runs every hour of every working day, reads the most sensitive records a company has, and gets more valuable the longer it accumulates. At that point three things change. Capacity you rent forever costs more than capacity you buy once. Data residency stops being a procurement checkbox and becomes something a client or a regulator asks you to demonstrate. And an operation that depends on a link to someone else’s building stops when that link does.
None of this is an argument against public cloud. It is an argument that the part of your business that agents operate belongs somewhere you control.
The junk data problem
The quiet reason agents disappoint is rarely the model. It is what the model was allowed to read.
An agent that reaches out to a general corpus brings back material that is stale, irrelevant or simply not yours, and then reasons confidently over it. Worse, nobody can tell afterwards which part of the answer came from your business and which came from the open internet.
When agents live on your infrastructure, the question answers itself. They read one organised record of the business and nothing else. Information carries its age, and an agent refuses to commit against a feed that has gone stale rather than quietly using it, because in a system whose whole claim is that it knew before you did, a rotting inventory feed is not a monitoring problem. It is a failure that looks exactly like success.
What a boundary actually is
A dedicated server is a foundation for control. It is not, on its own, a guarantee, and we would rather say so than sell the comfortable version.
How well a system is protected depends on its whole configuration: network restrictions, identity and permissions, which tools an agent may call, how policy is enforced, what is logged and what is monitored. Those are decisions made per deployment, written down, and checked.
Ground is configured so that the systems holding your data have no outbound network route at all. We call that No-Egress. It is written into the configuration agreed with you, not left to the hardware, and it has a real cost: a model that cannot reach out also cannot be updated from outside, so the update path and the support path have to be designed deliberately rather than assumed.
Inside your building
Ground is installed inside your enterprise. Servers on your premises hold the platform, the agents and every record, so all your data sits in one place, none of it leaves the building, and the agents work with the full context of the business.
It is sized to the business rather than sold as a specification: a site office that must keep working when the connection does not, a single office, a headquarters with sites reporting into it, a group large enough to train on its own data alongside live work, or a self-contained room delivered to your premises when there is no suitable space.
How it arrives
- Scoping sizes it: A call sizes it from your people, your sites and the work it has to carry.
- Set it up for you: Your records, permissions, approval rules and network boundaries are configured before it leaves us.
- Deliver and install: It arrives ready to plug in, with the agreed boundaries already in place.
- Run and support: Monitoring, updates and capacity planning against the rules documented at scoping.
Everything is confirmed at scoping before anything is built. We would rather size it conservatively and revise than deliver something wrong for the work.
What we believe
Most companies will not want to think about any of this, and they should not have to. What they want is for the system to work, for the data to stay where they said it would, and for the answer to “where is our data?” to be a room rather than a region name.
Our view is that the infrastructure question is not a technical preference but the thing that decides whether agents are safe to give real authority to. An agent that lives inside the business, on its own data, with no general route out, can be trusted with more than one that reaches across the internet to do its job. The infrastructure is not an upsell attached to the software. It is the reason the software can be trusted.