Azure Machine Learning is Microsoft's cloud service for building, training, deploying and managing ML models. Collaboration in Azure Machine Learning is possible through workspaces, which share notebooks, compute, data, environments and run results with role-based access, and through registries, which share models and components across workspaces, subscriptions and regions.
Updated October 2026
Who this is for: Heads of data science, ML platform owners, cloud architects and IT leaders running ML on Microsoft Azure in financial services, healthcare, manufacturing, retail and the public sector.
The business problem: models stuck between teams and versions
Many Azure ML estates grew organically: a workspace per team, scripts on SDK v1, prompt flows built for a 2023 pilot, and models copied by hand between dev and production subscriptions. That creates three costs: slow releases, security risk from unsupported code, and models no one can trace. In 2026, several of those older pieces have reached or are near end of support, which turns technical debt into a deadline.
What changed in 2026
- SDK v1 support ended on 30 June 2026 and CLI v1 support ended on 30 September 2025, according to Microsoft's v1 to v2 migration guide. v1 workloads keep running, but without support or fixes.
- Prompt flow will retire on 20 April 2027 in both Microsoft Foundry and Azure Machine Learning; Microsoft recommends migrating to Microsoft Agent Framework (see the Azure ML overview).
- Generative AI and agent building have moved to Microsoft Foundry; Azure ML remains the home for custom model training, MLOps and managed endpoints.
- Registries are now the standard way to move models and pipeline components between dev, test and production.
Core Azure Machine Learning services
| Capability | What it does |
|---|---|
| Workspace and studio | Shared home for experiments, notebooks, data assets, jobs and models |
| Compute clusters and serverless compute | CPU and GPU training, including multi-node distributed training |
| Pipelines and components | Reusable, versioned steps for training and batch workflows |
| AutoML and designer | Low-code model building for tabular, vision and text tasks |
| Model catalog | Open and partner models to fine-tune or deploy |
| Managed online and batch endpoints | Real-time and batch inference with traffic splitting |
| Registries | Share models, environments, components and data across workspaces and regions |
| Responsible AI dashboard and monitoring | Error analysis, fairness, explanations and data drift |
How collaboration in Azure Machine Learning works
1. Workspaces
Microsoft describes the workspace as the place that "organizes a project and allows for collaboration for many users". Teams share notebooks, compute, serverless compute, data and environments, and see each other's runs and metrics in the studio. Azure RBAC controls who can read data, submit jobs or deploy.
2. Registries
A registry works "much like a Git repository" for ML assets. Train in a dev workspace, publish the model to a registry, and deploy it from there to an endpoint in a production workspace, even in another subscription or region.
3. Git and CI/CD
Code lives in Git; pipelines run from GitHub Actions or Azure DevOps with CLI v2 and SDK v2. See our guide to CI/CD for machine learning.
Reference MLOps flow on Azure
notebooks, experiments, shared compute
GitHub Actions or Azure DevOps, SDK v2
models, components, environments
managed endpoints, monitoring
Figure: Azure Machine Learning MLOps flow. Teams collaborate in a dev workspace, CI/CD tests and publishes to a registry, production deploys from the registry, and models that must run outside Azure are managed by an edge and private control plane.
Azure ML vs Microsoft Foundry vs Azure Databricks
| Choose | When |
|---|---|
| Azure Machine Learning | Custom model training, MLOps, managed endpoints, regulated workloads |
| Microsoft Foundry | Building generative AI apps and agents with hosted foundation models |
| Azure Databricks | Lakehouse-centric data engineering and ML on large datasets |
Microsoft's own AI and ML products guide compares the options. For tooling beyond Azure, see ML platform tools.
Still running SDK v1 or prompt flow? Book a free Azure ML modernisation assessment with a migration plan and risk list.
How to start in 30 days
- Week 1: Inventory workspaces, SDK v1 scripts, prompt flows and endpoints. Flag anything unsupported.
- Week 2: Migrate one training job to SDK v2 and pipelines; add it to CI.
- Week 3: Create a registry; promote one model from dev to production through it.
- Week 4: Add monitoring and a rollback plan; plan prompt flow migration to Microsoft Agent Framework before April 2027.
Where it's heading: 2020 → 2030
| Period | Azure ML |
|---|---|
| 2020 | SDK v1, designer and ML Studio (classic) era. |
| 2026 | SDK/CLI v2, registries, serverless compute; GenAI and agents move to Microsoft Foundry. |
| 2027 (scheduled) | Prompt flow retires on 20 April 2027. |
| 2030 (our view) | Classic ML, LLMs and agents share one governed lifecycle across cloud, private and edge. |
How NexaStack helps
Azure ML trains and manages models in Azure regions. NexaStack extends that lifecycle to where Azure regions do not reach: factories, stores, vehicles, private data centres and air-gapped sites, with a unified inference engine, observability and evaluation, and policy-aligned governance. Related: MLOps for edge AI and distributed machine learning.
FAQ
Collaboration in Azure Machine Learning is possible through what?
Workspaces, which let teams share notebooks, compute, data, environments and results under role-based access, plus registries for sharing assets across workspaces.
Is Azure Machine Learning being retired?
No. Azure ML is active. SDK v1 and CLI v1 support has ended, and prompt flow retires on 20 April 2027.
What are the latest Azure Machine Learning trends in 2026?
SDK v2 migration, registries for cross-environment MLOps, serverless GPU compute, and the move of GenAI work to Microsoft Foundry.
Is migration risky, and does our data stay in region?
Microsoft states Azure ML does not store or process data outside the region you deploy in. Migrate one job at a time with tests, so risk stays low and progress is visible.
Need Azure ML models running on-premises or at the edge too? Book a hybrid MLOps pilot with NexaStack.
