CoreWeave adds official SkyPilot support for multi-cloud AI workloads

What SkyPilot brings to CoreWeave

SkyPilot was founded at the Sky Computing Lab at UC Berkeley. CoreWeave describes it as a framework designed to abstract cloud complexity and automate the selection, provisioning and management of compute and storage on any infrastructure. With CoreWeave now supported as a backend, users can launch SkyPilot jobs on their CoreWeave cluster from their familiar SkyPilot YAML or CLI.

CoreWeave says SkyPilot auto-selects the best region and GPU, including spot and preemptible instances, and manages job submission for ML training or inference. The company also lists dynamic resource matching based on compute and storage needs and pricing targets, plus built-in failure handling, smart retries and log streaming.

Networking, storage and scaling features

CoreWeave highlights several specific changes in the integration:

• InfiniBand in one line: adding network_tier: best to SkyPilot resources automatically configures InfiniBand, RDMA and environment variables, without manual tuning. CoreWeave says the official NCCL test example demonstrates 3.6Tb/s interconnect bandwidth. Elsewhere the post refers to fully automated configuration of 3.2 Tb/s Infiniband interconnects.
• Object storage: SkyPilot now recognizes CoreWeave buckets and includes install and config steps and a cw:// scheme in its storage docs. According to CoreWeave, fuse-mounted data can reach throughput of up to 7 GB/s per GPU. The company ties this to its LOTA (Local Object Transfer Accelerator) storage layer.
• Autoscaling: SkyPilot scales CoreWeave clusters up or down based on demand, which the company says means users pay only for what they use while keeping GPU utilization high.
• Documentation: CoreWeave is now included in the SkyPilot installation guide, with steps to connect a CoreWeave Kubernetes Service (CKS) cluster and optionally set up CoreWeave AI Object Storage.

Ready-to-use examples and recipes, such as Llama 4 FT and vLLM, are available on the SkyPilot examples page. CoreWeave says they let teams deploy training, development and serving workloads in minutes.

Who it targets and how to start

CoreWeave pitches the integration at both fast-moving AI-native startups and enterprises scaling to thousands of GPUs. The company lists several benefits for users: Slurm-like convenience with the flexibility of Kubernetes, a unified control plane across multiple Kubernetes clusters, and the ability to provision hundreds or thousands of GPUs with a single command. It also points to its Kubernetes-on-bare-metal architecture as a basis for price and performance optimization.

According to CoreWeave, SkyPilot on its platform already powers production workloads for a growing roster of customers, including AI-native startups running jobs across hundreds of GPUs. The company does not name them in the post.

Getting started involves installing SkyPilot and following the CoreWeave setup steps in the installation guide. Users can then validate networking with the InfiniBand/NCCL example and mount CoreWeave Object Storage through the cw:// scheme. CoreWeave also points to a walkthrough for creating dev pods, serving and multi-node training.

What it means for AI teams

For teams already using SkyPilot to spread work across clouds, the support makes CoreWeave a documented target with preconfigured networking and storage. Teams that prefer Slurm-style or Kueue-based workflows keep those options, since SkyPilot is an addition rather than a replacement. The performance figures cited here come from CoreWeave itself and the post does not provide pricing or availability details.