On-prem
- What it is
- Omni runs on hardware you own, inside your network.
- Typical fit
- Teams that cannot move the data, or already run the rack.
- What stays inside
- Knowledge, models, logs and backups stay in the building.
Omni Enterprise
Your infrastructure. Your rules.
Run Omni on your hardware, in your private cloud, or with no internet connection at all.
01 · Deployment
Four ways to keep it inside.
The same platform in each mode. What changes is who holds the hardware, the network and the keys.
Private cloud
- What it is
- Omni runs in a cloud account you control, in the region you choose.
- Typical fit
- Teams that want elastic capacity without shared tenancy.
- What stays inside
- The account, the network and the keys stay yours.
Hybrid
- What it is
- Core and knowledge on-prem; selected model capacity or channels in your private cloud.
- Typical fit
- Teams with mixed latency, capacity or residency needs.
- What stays inside
- The boundary is drawn per service. Knowledge stays where you put it.
Air-gapped
- What it is
- Omni runs with no internet connection at all.
- Typical fit
- Regulated or remote sites, and anyone who cannot allow an outbound link.
- What stays inside
- Everything stays inside. Updates arrive as a signed offline bundle.
The boundary in each mode
01On-prem
Runs on your hardware, inside your network.
02Private cloud
Runs in a private cloud account you control.
03Air-gapped
Runs with no internet connection.
02 · Hardware
Reference configurations, not promises.
Three shapes we have seen work. They are starting points for a quote, not a minimum or a guarantee.
Reference 01
Small starter
NVIDIA DGX Spark class
arm64 · single box
- Typical fit
- A small team, a pilot, or a first production box with no rack.
- Runs
- Omni Core, Knowledge, Conversations and a local model.
- Notes
- One power and cooling envelope. No redundant node, so back up to your own store.
Reference 02
Workstation
RTX Pro 6000 class
x86_64 · single node
- Typical fit
- A department, studio or site that wants local model capacity at the desk side.
- Runs
- Omni Core, Knowledge, Conversations and a local model for the workgroup.
- Notes
- Deskside power and cooling. Single node, so plan backup and restore.
Reference 03
Server
Multi-GPU rack
x86_64 · redundant data stores
- Typical fit
- A production rollout with more than one team and a support window.
- Runs
- Core, Knowledge, Conversations, Workflows and more than one model route.
- Notes
- Rack power, cooling and network separation. HA data stores for Enterprise.
03 · Models
Bring your own model.
Cloud, a private endpoint or a local GPU. The route is a deployment decision, not a platform rewrite.
Cloud
A hosted model under your account and your terms.
Use whenElastic capacity for low-sensitivity work.
Private endpoint
A model endpoint inside your private network or cloud account.
Use whenWhen the data can leave the desk but not the boundary.
Local GPU
A model running on the same hardware as Omni.
Use whenAir-gapped sites and the most sensitive work.
Per-agent choice
Each agent gets its own route, so a customer-facing room and an internal one do not have to share a model.
Allow-list and cost caps
Only approved routes are available. Caps limit what a route can spend.
Fallback
If a route is unavailable, the agent falls back to the next approved route or stops and says so.
04 · Operations
Keep it running, on your terms.
The operating work is part of the deployment. None of it depends on a connection back to us.
Signed updates and rollback
Updates arrive signed. A failed update rolls back to the previous release.
Backup and restore
Knowledge, configuration and audit records have a documented backup and restore path.
Monitoring
Health, queue depth and route availability are visible to your operators.
Audit export
Every answer and every action can be exported to your own store or SIEM.
Offline bundle
Air-gapped sites receive releases, models and connectors as a signed bundle.
05 · Procurement
What the buyer gets.
The technical review and the commercial review run together. The pilot is scoped in man-days, with a named owner on both sides.
For the buyer
A pilot with the paperwork in place.
Start with the security summary and the deployment checklist. Bring your platform team, your legal reviewer and the workflow that cannot leave the building.
Book a scoping call- Security summary — The deployment boundary, identity model and audit path in one document.
- Deployment checklist — Hardware, network, identity and backup prerequisites, written for your platform team.
- DPA and SLA on request — Reviewed with your legal and service owners before the order.
- Pilot scoped in man-days — A fixed pilot with a named business owner and a short list of workflows.
Continue
The rest of the suite.
Procurement
Scope the first deployment.
Bring the boundary, the hardware and the first workflow. We will scope the pilot in weeks, not quarters and leave the security pack with your reviewer.