Red Hat · Software · YoctoIT tech page

OpenShift AI

The open MLOps platform and Granite models served on-prem: enterprise AI with data that stays home.

FOCUS · AI ON-PREM, DATA AT HOMEFrom notebook to production model, on your infrastructure
Request the Free Assessment Learn more
YoctoIT material for clients and partners · Red Hat, RHEL, OpenShift, Ansible and the other products mentioned are trademarks of Red Hat, Inc. o di sue affiliate.
01 · What it is

OpenShift AI & RHEL AI, made clear.

OpenShift AI brings the model lifecycle inside OpenShift: workbenches for data scientists, training pipelines, scalable model serving (vLLM/KServe) and monitoring. RHEL AI adds IBM's Granite models and InstructLab to specialize them on your data — all on-prem, where data can stay.

On-prem
LLMs and RAG on your infrastructure: sensitive data doesn't leave
Granite
IBM's open models, Apache license: a solid, defensible base
GPU-ready
native GPU scheduling: NVIDIA (and beyond) governed as resources
OpenShift AI & RHEL AI
OFFICIAL RED HAT BRANDING · OPENSHIFT AI
SCHEMA UFFICIALE · WORKFLOW OPENSHIFT AI · FONTE: RED HAT DOCS
OFFICIAL DIAGRAM · OPENSHIFT AI WORKFLOW · SOURCE: RED HAT DOCS
02 · How to use it well

The things that make the difference.

From data to service

Applicazioni & copilotichat, RAG, agents on the processes
Workbench
Pipeline
Serving
explore · train · serve
OpenShift AIgoverned MLOps, scheduled GPUs
Fusion HCI + GPUthe physical base for AI
AI as an enterprise workload, not an experiment

RAG with company data

Documents and knowledge bases made queryable: the use case that pays back first, without sending data out.

Serious model serving

Autoscaling, canaries and model versioning: AI treated like any production service.

InstructLab

Granite specialized on your domain with examples, not million-dollar retraining: tailored AI made accessible.

Governed GPU costs

Quotas, time-slicing and monitoring: GPUs shared without wars between teams.

03 · In depth

Workbenches, pipelines and serving

OpenShift AI orchestrates the ML cycle: Jupyter workbenches with curated images (PyTorch, CUDA), Data Science Pipelines (Kubeflow) for reproducible training, a model registry and serving with KServe — vLLM for LLMs, with autoscaling and canaries. GPUs are shared with time-slicing or MIG; RHEL AI and InstructLab specialize the Granites with the knowledge taxonomy; all with RBAC and per-project quotas.

  • Workbench — governed Jupyter with per-team images and storage
  • DS Pipelines — Kubeflow: training as a versioned pipeline, not heroic notebooks
  • KServe + vLLM — LLM serving with batching and autoscaling
  • GPU sharing — MIG and time-slicing: the A100 shared without fighting
  • InstructLab — the taxonomy that specializes Granite: accessible fine-tuning
  • Model registry — versions, metrics and approvals: the model as an artifact
04 · Numbers and lifecycle

The numbers that matter.

100%
on-prem possible: data doesn't leave
MIG
up to 7 instances per A100/H100 GPU
Granite
IBM's open (Apache 2) models as the base
4/anno
releases aligned with OpenShift
On-prem AI is a platform to operate: GPUs, pipelines and serving put into production with the same rules as everything else.
05 · Use cases

Where it really pays off.

Internal copilots

Assistants on company documents and processes, on-prem.

Computer vision

Quality and safety in the plants, close to the lines.

Data science platform

A governed environment for all teams, GPUs included.

Enterprise AI is a platform problem before a model problem: the platform is our craft.