Back to blog
DeploymentNovember 2026

How to Bring a New Site Online in an Edge AI Fleet

Deplero Team

Deplero

A customer signs, and the first question after the kickoff is always the same: we have a new hospital, plant or warehouse coming online next month, so what does it take to get it running the same models as everyone else? If the answer is "a week of an engineer on site and a spreadsheet of what got installed", every new site costs more than the last one. It should be closer to an afternoon, and the site should show up in the console with the current release already serving.

Why adding a site is harder than it looks

A new site is somebody else's building. The network belongs to the customer's IT team, the firewall does not accept inbound connections, the hardware may be a GPU server in a closet or a small box bolted to a robot, and the people on site are not your engineers. Any step that needs remote access, a hand-edited config file or a call to your team to "push the model" turns into a delay measured in weeks.

The steps, in order

  1. Declare the site before anyone travels. The customer's deployment lead describes the new site once: its name, what kind of hardware it has, where its data may go (metrics only, de-identified clips to the vendor's cloud, or nothing leaves the building) and who approves rollouts there. The site now exists in the system, with no hardware behind it yet.
  2. Prepare one install package for that site. It holds the software, the site's own identity, and the list of signatures the site will accept. It only works for that site, and only once, so a lost copy cannot be used to join a different building.
  3. Install the site server. Someone at the site runs the install on the local server. A preflight checks disk, memory, clock and GPU driver first, so problems surface in the first minute instead of the fortieth. The site then connects outward to the control plane and reports ready. Nothing needs to reach into the site.
  4. Add each device. Every camera box or robot gets its own install, tied to that unit's hardware. The device proves who it is, reports what it has (GPU memory, board version, storage) and appears in the console under its site.
  5. Let the first release arrive on its own. The new devices receive whatever release the rest of the fleet is running, check its signature, and start serving. Nobody pushes anything by hand.

What it should cost

  • Under two hours for the site server, including the preflight, and under an hour when the site reuses a GPU box it already has.
  • Under twenty minutes per device. A plant with three lines is done before lunch.
  • Only outbound connections from the site, so the customer's IT team has nothing to open.
  • The same steps for an air-gapped site, carried in on an install archive instead of over the network.
  • No visit after the first day. A device that is powered off for a weekend or shipped from a staging area rejoins on its own.

Who does what

The customer's own people run every step on their own hardware. The vendor's team writes the site description and watches the site appear; the person on site follows a short, checked procedure and does not need to know how the models work. During the first installs we walk both through it, and after that the customer adds sites without us. That is the point: the platform runs in their environment, and adding a site should never depend on a call to us.

Questions to ask about your current process

  • How long did the last new site take, from contract to first result served?
  • Does anyone need remote access into the customer's network to finish it?
  • When the site came online, did it run the same model version as every other site?
  • Is where that site's data may go written down and enforced, or remembered?
  • Could the customer add the next site without your engineers?

If the honest answers are "weeks", "yes", "not sure", "remembered" and "no", the fleet grows only as fast as your team can travel. Fix the onboarding path before the tenth site, not after it.

INFRASTRUCTURE FOR PHYSICAL AI

Building this is the hard part. That's our part.

Deplero is the platform for running your AI at every site: vision and custom models at the edge, on-prem GPU clusters, and your cloud accounts, in one system. Installed in your environment, run by your team.

See the platform