Anyscale has launched its managed Ray platform into public preview on Microsoft Azure, effective June 2, 2026. The move brings a fully managed, enterprise-ready distribution of the open-source Ray framework directly into Azure’s ecosystem, promising native integration, enhanced governance, and cost controls for organizations scaling AI workloads. This public preview marks a significant milestone for enterprises grappling with the complexities of distributed AI and eager to maintain sovereignty over their data and models while leveraging Azure’s global infrastructure.

Ray has become the de facto standard for scaling Python applications across clusters, particularly for machine learning training, hyperparameter tuning, reinforcement learning, and model serving. Anyscale, founded by the creators of Ray, offers a managed platform that abstracts away the complexities of deploying, operating, and scaling Ray clusters. By bringing Anyscale Managed Ray to Azure, the companies are targeting enterprises that demand the flexibility of Ray with the peace of mind of a fully managed service, coupled with Azure’s compliance, security, and identity frameworks.

What Anyscale Managed Ray Brings to Azure

Anyscale Managed Ray on Azure provides a turnkey experience for data scientists, ML engineers, and platform teams. Instead of manually provisioning virtual machines, configuring networking, and installing Ray, users can launch clusters with a few clicks or API calls. The platform handles auto-scaling, fault tolerance, and rolling updates, freeing teams to focus on their AI workloads rather than infrastructure.

Key capabilities of the public preview include:

  • Native Azure Integrations: Deep hooks into Azure Active Directory (AAD) for role-based access control, Azure Monitor for observability, and Azure Blob Storage and Data Lake Storage for seamless data access.
  • Sovereign AI Controls: Data residency guarantees, customer-managed encryption keys, and options to deploy in sovereign Azure regions, addressing the growing demand for sovereign AI among governments and regulated industries.
  • Cost Governance and Observability: Granular cost tracking per project, team, or user, along with budget alerts and auto-shutdown policies to prevent runaway spending. The platform provides detailed dashboards showing resource utilization and spend breakdowns.
  • Unified Workload Support: Workers can run Ray Data for ETL, Ray Train for distributed training, Ray Serve for online inference, and Ray RLlib for reinforcement learning—all on the same cluster, maximizing resource utilization.
  • Enterprise Security: Private networking via Azure Virtual Network (VNet) integration, private endpoints, and support for Azure Policy to enforce organizational standards.

Sovereign AI: Control Without Compromise

One of the standout themes of this partnership is sovereign AI. As governments and enterprises grapple with regulatory requirements like the EU AI Act and sector-specific data protection mandates, the ability to run AI workloads with full control over data locality and access becomes non-negotiable. Anyscale Managed Ray on Azure allows organizations to lock down clusters within their own VNets, use dedicated hosts, and encrypt data with keys they manage.

Moreover, the service supports deployment in Azure’s sovereign regions—such as Azure Government in the United States or Azure Germany—ensuring data stays within jurisdictional boundaries. This is a direct answer to public-sector entities and defense contractors that have been hesitant to adopt managed AI platforms due to sovereignty concerns.

Anyscale’s platform also integrates with Azure Confidential Computing, enabling processing of sensitive data at the hardware level. Combining this with Ray’s ability to scale Python apps securely opens doors for handling classified datasets or personally identifiable information across distributed nodes, all while meeting strict compliance standards.

Cost Control: Taming AI Infrastructure Sprawl

AI infrastructure costs can spiral quickly, especially when teams spin up GPU clusters for experimentation and forget to shut them down. Anyscale Managed Ray directly tackles this challenge with built-in cost governance tools. Administrators can set per-project budgets, define auto-shutdown policies based on idle time, and receive alerts when spending exceeds thresholds.

The platform also promotes efficient resource utilization through intelligent workload scheduling. Ray’s architecture already excels at sharing cluster resources across multiple tasks; Anyscale adds a management layer that optimizes bin packing and can preempt lower-priority jobs to free up GPUs for critical workloads. For enterprises running hybrid cloud or multi-cloud strategies, the unified dashboard helps compare costs across environments—a feature that becomes even more powerful with Azure’s Cost Management integration.

Early adopters during the private preview reported up to 40% reductions in cloud spend versus self-managed Ray clusters, primarily due to better resource consolidation and minimized idle time. Such numbers are likely to accelerate adoption as the public preview extends to a wider audience.

Public Preview Details and How to Get Started

The public preview is open to all Azure customers, though capacity may be limited in certain regions due to GPU demand. Interested users can sign up through the Anyscale website or directly via the Azure Marketplace. The initial preview supports a subset of Azure regions, including East US, West Europe, and Southeast Asia, with plans to expand rapidly.

Clusters can be provisioned using an Azure-native experience, meaning users authenticate via their existing Azure accounts and can launch clusters directly from the Azure portal. Anyscale also provides a Terraform provider and a CLI for infrastructure-as-code workflows. Once a cluster is up, users interact with Ray as they normally would—submitting jobs via the Ray client or using Anyscale’s SDK, which extends Ray with enterprise features like job scheduling and artifact management.

During the public preview, Anyscale and Microsoft are offering a limited free tier, including credits for compute up to a certain threshold, along with joint support through Azure’s support channels. This co-support model is crucial for enterprises that demand a single throat to choke when things go wrong.

Impact on Azure’s AI Ecosystem

Anyscale Managed Ray fills a gap in Azure’s AI platform. While Azure offers managed services like Azure Machine Learning for training and deployment, and Azure Kubernetes Service (AKS) for container orchestration, neither was purpose-built for the dynamic, heterogeneous workloads that Ray excels at. AKS can run Ray, but it requires significant expertise to configure autoscaling and GPU sharing correctly. Anyscale’s offering abstracts that away and adds enterprise features that many users would otherwise have to build themselves.

The preview also reinforces Microsoft’s commitment to embracing open-source AI frameworks as first-class citizens on Azure. Just as Azure Databricks provides an optimized Spark experience, Anyscale Managed Ray aims to deliver a similarly polished environment for the Ray ecosystem. This could accelerate the adoption of Ray among Azure-native enterprises that previously viewed it as too complex to manage.

Partnerships with other parts of the Microsoft ecosystem are already taking shape. Anyscale and Microsoft are collaborating on integration with Microsoft Fabric, enabling Spark and Ray workloads to interoperate seamlessly. Additionally, tight coupling with Azure OpenAI Service allows models served via Ray to leverage Azure’s managed inference endpoints with minimal overhead.

What Early Adopters Are Saying

While the public preview is just opening, feedback from the private preview has been positive, especially around ease of use and cost visibility. One large financial services company reported reducing their ML training orchestration time from weeks to days, all while meeting strict regulatory requirements through sovereign controls. Another e-commerce player scaled their recommendation engine tuning from 500 to over 5,000 parallel trials, dramatically improving model accuracy without ballooning costs.

A common thread among early users is the reduction in operational overhead. Infrastructure teams no longer need to become Ray experts to support their data science counterparts; instead, they rely on Anyscale’s platform to handle upgrades, security patches, and performance tuning. This operational simplicity is a key differentiator in the managed platform market.

Anyscale Managed Ray enters a competitive but fragmented market. Databricks offers a managed Spark and MLflow environment; Amazon SageMaker provides managed training and inference; Google’s Vertex AI integrates with Ray on GKE. However, none currently offer a fully managed, cloud-native Ray platform with such deep integration into a single cloud provider’s identity, networking, and cost management stack—at least not with Anyscale’s pedigree.

The partnership is exclusive for now, but Anyscale has not ruled out similar offerings on other clouds. For Azure, having first-mover advantage could attract Ray-committed enterprises moving to the cloud or looking to consolidate their AI infrastructure.

Forward-Looking: General Availability and Beyond

Anyscale and Microsoft have not announced a timeline for general availability (GA), but typical preview periods range from three to nine months. Based on the feature set already present in the preview, GA is expected to include additional regions, enhanced SLA-backed support, and tighter integrations with Azure Machine Learning v2.

Longer-term, the roadmap includes support for Azure Arc-enabled hybrid deployments, allowing enterprises to run Managed Ray clusters on-premises or at the edge and manage them from Azure. This would further strengthen the sovereign AI narrative, enabling organizations to keep sensitive data on-prem while bursting to the cloud for additional capacity.

For Windows and .NET developers, the preview also nods toward supporting Ray’s experimental .NET bindings. While Ray’s core is Python-centric, Anyscale is committing to improving the developer experience for C# and F# workloads, recognizing the large number of .NET shops within the Azure ecosystem. This could open the door for enterprise applications that currently rely on .NET-based data processing pipelines to tap into Ray’s distributed computing power.

Getting Ready for the Shift

Organizations interested in the public preview should start by inventorying their existing Python-based AI workloads. Identify training jobs that take too long, inference endpoints that struggle under load, or data processing pipelines that could benefit from distributed execution. Ray shines in these scenarios, and Anyscale Managed Ray on Azure removes the operational barriers that often prevent teams from adopting it.

Microsoft and Anyscale are hosting a series of webinars and workshops starting this month, with hands-on labs showing how to migrate existing Ray scripts to the managed platform. Documentation is available on both the Anyscale docs site and Microsoft Learn, and a dedicated community channel has been set up for preview participants to share feedback.

The convergence of Azure’s enterprise muscle and Anyscale’s Ray expertise creates a compelling proposition for any organization serious about scaling AI. As the public preview unfolds, it will be well worth watching how quickly enterprises move from experimentation to production—and what new AI workloads this partnership ultimately enables.