Microsoft Open-Sources TauGrid for GPU AI on Kubernetes

Microsoft has open-sourced TauGrid, a Kubernetes-native platform designed to simplify the operation of GPU-based artificial intelligence workloads. Developed by the Azure Kubernetes Service engineering team, the platform combines workload submission, queueing, distributed execution, GPU health checks, and observability in a stack that can be installed with a single Helm command, according to Marktechpost’s report.

TauGrid is available under the MIT license. Its container images and Helm charts are published as public OCI artifacts through Microsoft Container Registry. The platform can run on Kubernetes 1.30 or later with GPU nodes, kubectl, and Helm 3.0 or later, allowing deployment beyond Azure environments.

A unified stack for shared GPU clusters

Teams running AI jobs on Kubernetes commonly assemble several separate systems. TauGrid brings those responsibilities together through the tau command-line interface, Kueue for workload queueing and admission, KubeRay for Ray cluster orchestration, GPU node health monitoring, and cluster and workload observability.

The platform separates infrastructure administration from researcher workflows. Platform teams configure workspaces, queues, compute profiles, storage, identity, and observability. Researchers can then work from a code repository and submit workloads through the CLI without directly configuring Kubernetes resources.

TauGrid’s components are primarily written in Go. Microsoft open-sourced the project on August 28, 2026, in the Azure/taugrid GitHub repository.

From a workload file to a scheduled GPU job

A workload begins with a tau.yaml configuration file that specifies the entrypoint, workload type, resource requirements, and runtime image. Microsoft’s published GPU training example describes a PyTorch workload using one A100 GPU, one worker, 16 CPUs, and 64 GiB of memory. The example uses a Ray runtime image and installs PyTorch as a dependency.

When a user runs tau run, TauGrid applies platform policy, creates either a Kubernetes Job or a KubeRay RayJob, and submits it through Kueue. In a shared cluster, multiple teams can place workloads into a common Kueue ClusterQueue. Kueue uses quota and priority to determine admission, while Kubernetes schedules admitted jobs onto available healthy GPUs.

Microsoft describes the workload lifecycle in six stages: submission, queueing, execution, monitoring, recovery, and evidence. Recovery features include retrying a workload, resuming from a checkpoint, and diagnosing failures.

Evidence records retain workload metadata, configuration, logs, metrics, checkpoints, and execution history. These records support later review of how a run was executed and provide platform operators and researchers with a way to reproduce and audit workloads using the captured configuration and execution data. The available reporting does not specify how broadly those records can be exported or integrated outside the platform’s documented observability options.

Deployment and operational controls

Operators can install TauGrid from Microsoft Container Registry with a versioned Helm chart. Microsoft publishes first-party images for Tau, the TauGrid Portal, and the tau core controller under the mcr.microsoft.com/aks/ai-runtime/ path. The project recommends using versioned tags or immutable image digests instead of the latest tag.

The Tau CLI is available through GitHub Releases for Linux and macOS, with a PowerShell installer for Windows amd64. The installer verifies the release checksum and does not modify the system PATH.

TauGrid also sends no telemetry to Microsoft by default. Remote export remains disabled unless an operator configures a destination, giving deployments control over whether workload information is sent elsewhere.

Cloud-neutral goal, Azure-specific integration

The platform’s stated goal includes supporting both cloud and on-premises Kubernetes without requiring an Azure dependency. Some integrations remain Azure-specific, however. In particular, the documented observability integration through Azure Data Explorer currently ties that capability to Microsoft’s cloud ecosystem.

TauGrid offers a packaged starting point for organizations that already operate GPU-enabled Kubernetes clusters and want a unified workflow for submitting and managing AI jobs. Its practical scope will depend on how the project’s cloud-neutral integrations develop and how operators adapt its queueing, evidence, and observability features to their own environments.