Reference architecture
A documented design sized for your workloads, GPUs and availability needs.
A pre-engineered blueprint to stand up Red Hat OpenShift and OpenShift AI with GPU enablement, model serving, agent runtime, security baselines and GitOps. It runs on-prem, in a sovereign cloud or in the public cloud.
Building an AI platform on Kubernetes from scratch takes months: GPU drivers, model serving, networking, security baselines, GitOps and monitoring all have to be designed and tested. Many teams get a cluster running but stall before it is ready for production agents.
This accelerator uses a tested blueprint to stand up Red Hat OpenShift and OpenShift AI with GPUs, model serving, an agent runtime, security baselines and GitOps, on premises, in a sovereign cloud or in the public cloud, and then hands it over to your team.
A well-built platform is what makes later agent projects fast. Once GPUs, model serving, identity, monitoring and security baselines are in place and understood by your team, each new agent becomes an application deployment rather than an infrastructure project. Organisations with strict data residency rules, such as banks and government entities in the Gulf and India, often start here so that every later use case can stay on infrastructure they control.
A documented design sized for your workloads, GPUs and availability needs.
GitOps-driven install of OpenShift and OpenShift AI, repeatable for each environment.
GPU operator configuration and model serving for open and commercial models.
A runtime for agents with guardrails, tracing and observability.
Hardening, policies, image scanning and secure base images.
Operations runbooks and hands-on knowledge transfer for your platform team.
| Phase | When |
|---|---|
| Design and sizing | Weeks 1 to 2 |
| Install and configure | Weeks 3 to 5 |
| Model serving and agent runtime | Weeks 6 to 7 |
| Testing and handover | Week 8 |
A platform owner, two engineers who will run the platform afterwards (about half their time during build), and data centre or cloud teams to provide hardware, network and identity integration.
Timelines are typical for a defined scope and depend on data access, integration and internal approvals.
| Clusters | Production and non-production, three control-plane nodes each |
|---|---|
| GPU nodes | Four nodes recognised, drivers and operator healthy (illustrative) |
| Model serving | Two open models served; latency and throughput tests passed against agreed targets |
| Security | Baseline policies applied; image scanning enabled; no critical findings open |
| GitOps | All configuration in Git; a test change rolled out and rolled back successfully |
| Handover | Runbooks reviewed with the platform team; on-call rota agreed |
Illustrative example with made-up details, to show the format and level of evidence. It is not from a client.
See it as a worked example: published case study.
Yes. The blueprint supports disconnected and air-gapped installations, which many banks and government entities need.
Open models such as Llama, Mistral or Granite, and commercial models where licences allow, using OpenShift AI model serving.
Open hybrid cloud and AI platforms, from data centre to edge
Capabilities & use casesAccelerated computing and the software stack for AI agents
Capabilities & use casesPrivate cloud AI infrastructure, turnkey and on-prem
Capabilities & use casesSecure-by-default containers and libraries for the software supply chain
Capabilities & use casesGet a free consultation from our AI, data and security specialists. Honest advice, clear options and a reply within 6 hours.
Free consultation. Our team replies within 6 hours.
Free consultation. Our team replies within 6 hours.