Rack it, connect it, govern it.
A private AI appliance is hardware delivered with Data.co Control, model serving, retrieval, and your selected models already installed and validated. It shortens the path from purchase order to a working, governed environment.
Most delays are integration, not hardware.
Standing up private AI from parts means matching drivers to runtimes, runtimes to models, and models to configurations, then debugging the combination on site. Appliances move that work before shipment, onto a configuration we have validated together.
Appliances suit teams with limited platform engineering capacity, remote or branch sites, and air-gapped environments where on-site troubleshooting is expensive.
From order to first governed request.
- 1
Specify
Assessment results set the hardware configuration and the models to preload.
- 2
Build and validate
We install the stack, load models, and run validation tests before shipping.
- 3
Install
Rack, power, and connect to your network using the provided plan.
- 4
Integrate
Connect your identity provider and first data sources, then apply policies.
Ready on arrival.
Data.co Control
Gateways, policy engine, RAG with pgvector, model registry, and audit log.
Model serving
vLLM or Ollama, matched to the hardware and validated with your selected models.
Licensing
Online or offline signed license, including for air-gapped sites.
As-built record
Configuration, versions, and validation results documented for your records.
Appliances are a practical path to air-gapped AI.
With models, software, and an offline license loaded before delivery, the first install does not depend on media transfers.
When each approach makes sense.
| Factor | Appliance | Build from components |
|---|---|---|
| Platform engineering capacity | Limited, or focused elsewhere | Strong team that wants control of every layer |
| Site | Remote, branch, or disconnected | Central data center with on-site staff |
| Existing hardware standards | Flexible | Strict vendor or chassis standards to follow |
| Time to first governed request | Integration done before delivery | Integration done on site |
| Customization | Within validated configurations | Anything your team can support |
What is typically checked before an appliance leaves staging.
The exact test plan is agreed in the order and its results are recorded in the as-built documentation.
- Hardware burn-in under sustained GPU, memory, and storage load
- Firmware, driver, and runtime versions match the recorded baseline
- Each preloaded model loads and its weights match the pinned digest
- Gateway health and readiness checks pass
- A sample request completes end to end through policy and audit
- Audit log chain verifies cleanly
- License verifies offline against the deployment it was issued for
An appliance is still a server you will own for years.
After installation, an appliance is operated like any other Control deployment. Software and model updates come through your registry mirror or signed offline media, policies come through the admin API or signed bundles, and hardware health goes to your monitoring. Treat the as-built record as the baseline: every later change should be recorded against it so the next node can be built to match.
Plan expansion and retirement from the start. Adding units extends capacity under the same control plane. At end of life, drives that held indexes, audit logs, or model weights should be sanitized or destroyed under your media handling procedure, and the audit segments archived before the hardware leaves your custody.
Questions buyers ask
Can we choose the models that come preloaded?
Yes. Models are selected during the assessment from open-weight options that fit the hardware, registered with pinned digests, and validated before shipment.
How is an appliance updated?
The same way as any Control deployment: from your registry mirror, or from signed offline media in disconnected environments.
Can appliances be clustered?
Yes. Multiple appliances can form one deployment, with capacity added by adding units.
What support comes with an appliance?
Hardware support terms and software support are defined in the order. Managed services can be added if you want Data.co to operate it.
What do we need ready on site before an appliance arrives?
Rack space, power and cooling for the unit, network addresses and VLANs from the network plan, a reachable identity provider, and a person who can approve the first policies. A local registry mirror is needed later for updates.
Can we swap the preloaded models later?
Yes. New models are registered, evaluated, and promoted through the model registry like on any deployment, subject to the memory the hardware provides.
Is an appliance locked to Data.co hardware support?
Hardware warranty comes from the manufacturer and is documented in the order. The software runs the same way it would on hardware you sourced yourself.
Size it before you buy it.
An infrastructure assessment maps your workloads, data constraints, and existing hardware to a deployment model and a bill of materials you can review.