Choose your path
Choose a path based on what you need to do. Create and manage clusters covers cluster and Node operations. Configure cluster access covers authentication to existing clusters. Deploy workloads covers scheduling, Charts, and frameworks. Secure your cluster covers optional private access, workload identity, and Object Storage federation.Create and manage clusters
Create clusters and Node Pools, and manage Nodes.
Configure cluster access
Authenticate users to an existing cluster with Managed Auth or an external identity provider.
Deploy workloads
Schedule workloads, use CoreWeave Charts and third-party frameworks, and follow tutorials.
Secure your cluster
Private API access, workload identity, and Object Storage federation.
Create and manage clusters
If you’re new to CoreWeave, start with Get started with CoreWeave for account activation, Identity and Access Management (IAM), and tool setup. Complete these steps to create a cluster and make it usable: To maintain an existing cluster, manage Node Pools, check Node Pool status, cordon Nodes, reboot Nodes, or connect to a Node for troubleshooting. For how CoreWeave provisions and repairs Nodes, see Node Lifecycle under Nodes and Node Pools.Configure cluster access
Use Configure cluster access after a cluster exists to choose how users authenticate to its Kubernetes API. Platform identity, Managed Auth, SAML, API access tokens, and IAM roles live in the Security docs. CKS pages cover Managed Auth kubeconfig and unmanaged authentication. Private API access, workload identity, and Object Storage federation live under Secure your cluster.Deploy workloads
When you already have a cluster, use these resources to run workloads:- Deploy an open source LLM for an end-to-end first workload.
- Read Workload scheduling to learn how CKS places Pods, including interruptible workloads and Internode Memory Exchange (IMEX).
- Install CoreWeave Charts or use a third-party framework.
- Follow a tutorial such as Deploy vLLM for Inference.
Secure your cluster
Use these guides for optional or advanced hardening after you have a cluster:- Access a CKS cluster through Tailscale VPN for private API access.
- Workload Identity when Pods need short-lived credentials for external services.
- CKS Workload Federation for AI Object Storage to federate access to Object Storage.
How CKS is different
The following sections describe how CKS differs from other managed Kubernetes services.High-performance Kubernetes on high-performance compute
CKS is engineered to orchestrate and serve highly computationally intensive workloads, designed for model training, inference, and HPC tasks.- CKS harnesses CoreWeave’s extensive fleet of high-performance GPU and CPU servers and advanced HPC networking infrastructure to maximize throughput and minimize latency.
- CKS runs Kubernetes directly on bare metal Nodes, without a hypervisor. Customer clusters don’t run Virtual Machines.
- CKS clusters use NVIDIA® BlueField® Data Processing Units () attached to each Node to offload processing tasks. This allows the Node to focus primarily on executing application workloads.
- CKS Nodes are stateless. At each boot, they load a clean Operating System image. This allows Nodes to be rapidly scaled and re-provisioned, and ensures that all Nodes use the correct software versions.
- Clusters integrate with an InfiniBand fabric that features a non-blocking, Fat-Tree architecture with NVIDIA® Scalable Hierarchical Aggregation and Reduction Protocol (SHARP)™ optimizations. These optimizations are capable of supporting the demand of tasks like training Large Language Models (LLMs) across clusters comprised of thousands of GPU instances. CoreWeave doesn’t turn on SHARP by default. See NCCL network plugin guidance to learn how to enable SHARP.
Do not install the NVIDIA GPU Operator on CKS clusters
Hyper-secure infrastructure
CKS is engineered with a strong emphasis on security.- The DPU-based architecture used by CKS enables advanced security features including custom network and security policies, dedicated Virtual Private Clouds (VPCs), and privileged network access controls.
- CoreWeave houses physical infrastructure in high-surveillance data centers, providing comprehensive security. Within the CKS platform, your data is operationally isolated, ensuring complete separation and confidentiality.
Data Plane flexibility
CKS provides a managed Control Plane alongside configurable Data Plane elements. This strikes a balance between operational simplicity and the flexibility to tailor the environment to specific workload requirements.- CKS clusters include a pre-installed Container Storage Interface (CSI) and Container Network Interface (CNI) to standardize storage and networking across container environments.
- Unlike other managed Kubernetes services, CKS extends your control beyond the Control Plane by allowing direct management of Data Plane components. This approach minimizes your infrastructure management burden while still offering extensive customization possibilities.
Additional advantages
The following sections describe additional advantages that CKS offers.Privilege and access management
CKS offers comprehensive privilege management, balancing managed solutions with the flexibility of self-managed ones. This gives you the freedom to choose the most suitable option for your security needs. CKS Managed Auth provides a managed solution backed by role-based access control (RBAC) for cluster-wide access control and organization management. CKS supports third-party RBAC providers for granular in-cluster permissions.Metrics and observability
CKS provides comprehensive support for auditing and compliance requirements. You can also integrate performance and data metrics with your existing infrastructure, enabling a flexible observability solution.- For enhanced monitoring, CKS grants access to the CKS cluster API server, enabling Control Plane audit logs.
- This extensive access also lets you deploy your own metrics stacks with custom collection tools, including DaemonSet or Loki logging.
- You can also monitor Node logs, GPU utilization, and other Node-level metrics through CoreWeave Grafana.