Nebius with Saturn Cloud: Setting up an AI development platform
Nebius with Saturn Cloud: Setting up an AI development platform
Saturn Cloud is an AI development platform that is deployed from your Nebius console. It enables AI/ML teams to build and deploy models with their favorite tools, without reliance on engineering support. You can use it for free and watch your AI/ML team onboard and start shipping right away.
Why Saturn Cloud matters
When working with raw Kubernetes, infrastructure engineers still have to do a lot of heavy lifting to get AI teams productive: setting up GPU node pools and scaling policies, managing Python and package environments, configuring secure data access and adding the monitoring, logging and cost controls that keep clusters reliable and efficient. These tasks can take weeks to implement before the first model can even run. Saturn Cloud manages this entire layer automatically, providing ready-to-use environments, autoscaling compute, built-in observability and team-level controls out-of-the-box. This allows infrastructure engineers to focus on higher-value initiatives, while their AI teams can be productive on day one: opening notebooks, connecting to data and training models immediately.
How it works
Saturn Cloud installs into an Managed Kubernetes cluster hosted in Nebius AI Cloud. Most deployments are managed by Saturn Cloud, where customers provide Saturn Cloud with access keys to their Nebius account. Saturn Cloud then provisions an MK8s cluster, installs itself and provides ongoing maintenance, support and upgrades. You can also self-manage the deployment if you prefer.
This section covers the installation procedure.
Installation procedure
If you are provisioning your own MK8s cluster, Saturn Cloud’s team recommends managing MK8s with Terraform. Here’s a sample Terraform recipe you can customize, or you can write it yourself.
The Terraform recipe creates:
-
A system node pool (where the Saturn Cloud control plane and supporting microservices run)
-
A few CPU node pools if you have CPU workloads
-
A few GPU node pools
Once you have an MK8s cluster, Saturn Cloud can be installed via a Helm chart. To begin, please register your company with Saturn Cloud. The registration process will present you with a short-lived (4-hour) bootstrap API token that you will pass into the Helm chart for the saturn-helm-operator which is publicly hosted on GitHub at oci://ghcr.io/saturncloud/charts/saturn-helm-operator-nebius.
The operator will:
-
Exchange the 4-hour bootstrap token for a long-lived API token, and keep it refreshed
-
Set up the necessary secrets for SSL termination at
*.your-company.saturnenterprise.io -
Install Saturn Cloud and all supporting microservices
Kubernetes services
The supporting microservices include:
-
Fluent-bit and Elasticsearch for log management
-
Traefik, along with the Saturn Cloud auth-server, because Saturn Cloud gates network access to deployed resources using RBAC
-
An SSH proxy so that customers can SSH into your containers if you choose to do so
-
Prometheus for monitoring and metrics
The Saturn Cloud installation also relies on solutions in MK8s, including:
- The cluster-autoscaler
- Cilium for network policies and encryption
- GPU Operator for GPU management
- Network Operator for interconnect support
Post-installation
Once Saturn Cloud is installed, you will receive an email to reset your password and log in to the admin account. After that, you can invite your team to the installation, and optionally configure SSO for access.
Saturn Cloud resources
The primary purpose of Saturn Cloud is to manage AI/ML resources. All Saturn Cloud resources utilize public container images that Saturn Cloud has published, which include the most commonly used AI/ML libraries. Most customers end up building their own images either using the Saturn Cloud image builder or with Docker directly.
Saturn Cloud resources also load code from your Git repositories. Because Saturn Cloud resources just run container images with code from your Git repositories, the workloads are entirely portable. They can be easily exported from Saturn Cloud if you choose to do so in the future.
There are three primary resource types:
-
Development workspaces are K8s deployments that run either JupyterLab or RStudio. They also support SSH, allowing customers to use VS Code, Cursor, PyCharm or any other IDE that supports SSH with Saturn Cloud. Development instances are backed by a Nebius persistent disk mounted in your home directory.
-
Jobs are K8s jobs that can also be scheduled as K8s cronjobs. Jobs can also be used in multi-node training. Saturn Cloud jobs propagate environment variables so that your training code (for example, torchrun) can coordinate the multiple nodes within the job.
-
Deployments are K8s deployments that serve HTTP traffic. This is primarily used to deploy LLMs, ML models and dashboards.
All three can also attach Nebius shared filesystems.
All Saturn Cloud resources can be exported or created from a YAML recipe. This facilitates CI/CD patterns where you want to deploy straight to Saturn Cloud from your CI/CD system. For more information, check out the Saturn Cloud docs here.
Ask your Nebius solutions architect or account executive about using Saturn Cloud in your Nebius account. It’s free of charge for accounts approved by the Nebius team.
