A dedicated AI Factory Landing Zone for the full ML lifecycle: data engineering, feature engineering, model training, registry, deployment pipelines, and monitoring — all within a purpose-built subscription.
The AI Factory pattern creates a purpose-built landing zone for the complete ML lifecycle. Azure Databricks handles data engineering and distributed training, Azure Machine Learning manages experiments, model registry, and pipelines, while Azure OpenAI provides LLM inference endpoints.
This is the go-to pattern for enterprise ML teams that need CI/CD for models, feature engineering at scale, and automated retraining pipelines.
| Tier | Premium |
| VNet Injection | Enabled |
| Public Subnet | /22 delegated |
| Private Subnet | /22 delegated |
graph TB
subgraph Factory["AI Factory VNet 10.20.0.0/16"]
subgraph DBX["Databricks VNet-Injected"]
DBPUB["Public Subnet /22"]
DBPRV["Private Subnet /22"]
end
subgraph Compute["ML Compute"]
AML["Azure ML Workspace<br/>Pipelines & Registry"]
MLCOMP["ML Compute /22"]
end
subgraph Data["Data Layer"]
FEAT["ADLS Gen2<br/>Feature Store"]
MODELS["ADLS Gen2<br/>Model Storage"]
end
subgraph AI["AI Services"]
AOAI["Azure OpenAI<br/>LLM Inference"]
ACR["Container Registry<br/>Premium"]
end
subgraph Sec["Security & Monitoring"]
KV["Key Vault"]
MON["Log Analytics<br/>120-day retention"]
AI2["App Insights"]
end
end
subgraph CICD["CI/CD"]
GH["GitHub Actions /<br/>Azure DevOps"]
end
DBX -->|"ETL / Training"| FEAT
DBX -->|"Feature Eng"| MODELS
AML --> FEAT
AML --> MODELS
AML --> ACR
AML --> KV
AML --> AOAI
GH -->|"Pipeline Trigger"| AML
AOAI --> MON
AML --> MON
DBX --> MON
| # | Service | Resource Name | SKU / Tier | Purpose | Monthly Cost |
|---|---|---|---|---|---|
| 1 | Azure Machine Learning | {base}-{env}-aml | Workspace + ACR | Pipelines, experiments, model registry | $0* |
| 2 | Azure Databricks | {base}-{env}-dbw | Premium, VNet-injected | Spark training, ETL, feature engineering | ~$2,500 |
| 3 | Azure OpenAI | {base}-{env}-openai | S0 | LLM inference endpoints | ~$1,800 |
| 4 | ADLS Gen2 (Features) | {base}{env}features | Standard LRS, HNS | Feature store, training data | ~$60 |
| 5 | ADLS Gen2 (Models) | {base}{env}models | Standard LRS, HNS | Registered models, packages, serving artifacts | ~$40 |
| 6 | Container Registry | {base}{env}acr | Premium, 30-day retention | ML model container images | $167 |
| 7 | Key Vault | {base}-{env}-kv | Standard | Keys, connection strings | ~$5 |
| 8 | Log Analytics | {base}-{env}-law | PerGB2018, 120-day | Audit logging, experiment tracking | ~$35 |
| 9 | Application Insights | {base}-{env}-ai | Workspace-based | Pipeline and model telemetry | Incl. |
| 10 | VNet | {base}-{env}-vnet | /16 with 5 subnets | Network backbone for factory | Free |
| 11 | Private DNS Zones | 6 zones | Global | Private name resolution | ~$3 |
| 12 | Private Endpoints | 4 endpoints | — | OpenAI, Features-Blob, KV, ACR | ~$29 |
| Estimated Total (Moderate Production) | ~$6,120/mo | ||||
* AML workspace free; compute instances billed separately. Databricks cost based on Premium DBU consumption.
VNet-injected Databricks workspace for Spark-based distributed training, ETL pipelines, and feature engineering. Public and private subnets each /22. No public Databricks infrastructure exposed.
Full MLOps workspace linked to ACR, ADLS, Key Vault, and App Insights. Pipeline orchestration, experiment tracking, model registry with versioning and staged deployments.
LLM inference endpoints for the AI Factory. Used for generating embeddings, fine-tuning evaluation, and production inference. Private endpoint access only.
Two separate accounts: (1) Feature Store for computed features, training data, and experiment outputs; (2) Model Storage for registered models, packages, and serving artifacts.
Premium ACR for ML model container images. Admin user disabled (managed identity only). 30-day image retention policy. Geo-replicated for HA.
Log Analytics with 120-day retention for audit compliance. Tracks all Databricks jobs, AML experiments, model deployments, and OpenAI usage.
Databricks for distributed training, AML for orchestration. Handles petabyte-scale data processing and model training.
Automated pipelines triggered by data drift detection. Retrain, validate, and deploy models without manual intervention.
Full MLOps lifecycle in one landing zone. From data engineering to production model serving.
Spark-based feature computation on Databricks with ADLS Gen2 feature store. Delta Lake for ACID transactions.
| Constraint | Mitigation |
|---|---|
| Databricks Premium cost is significant | Use job clusters (auto-terminate), spot instances for training |
| VNet-injected Databricks consumes large IP ranges | Plan /22 subnets; ensure VNet has capacity |
| Complex to operate (many moving parts) | Use AML pipelines for orchestration; GitOps for config |
| ACR Premium is expensive for low usage | Consider Standard tier if geo-replication not needed |
Generate a complete deployment spec sheet with GitHub Actions workflow, Bicep file structure, and prerequisite checklist.
Generate Deployment Spec Sheet