Skip to content

CoreWeave: An AI Cloud Built from Kubernetes, GPU Fabric, and Storage

Aug 22, 2026 1 min
TL;DR CoreWeave is more than rented GPUs: it combines Kubernetes, GPU networking, storage, and inference into AI infrastructure, while platform engineering and capacity governance remain yours.
Table of Contents
  1. CKS is a platform, not a model API
  2. Inference has two responsibility boundaries
  3. Capacity, portability, and failure tests
  4. References

🌏 中文版

CoreWeave is cloud infrastructure focused on AI and HPC. It is more than hourly GPUs: the platform includes CoreWeave Kubernetes Service (CKS), Slurm on Kubernetes, object/file/block storage, VPCs, and serverless or dedicated inference.

CKS is a platform, not a model API

CKS integrates a Kubernetes control plane, GPU nodes, and high-performance networking for distributed training, self-hosted inference, and shared platforms. Teams still own workload specs, requests and limits, queues, checkpoints, autoscaling, RBAC, and rollouts. Without a Kubernetes platform owner, GPU access can still stall on scheduling and recovery.

For training, test GPU topology, collective communication, shared-filesystem throughput, and checkpoint restore—not only a single-GPU benchmark. Tier data: durable corpora in object storage, multi-node hot data on file storage, and local or ephemeral disks as cache. A bottleneck anywhere leaves expensive GPUs waiting for I/O.

Inference has two responsibility boundaries

Serverless inference reduces idle capacity for request-driven workloads. Dedicated inference suits steady traffic, latency SLOs, specific hardware, and isolation. Measure cold and model-load time, tokens per second, batching, queue delay, errors, and cost per successful output.

Hourly GPU price is insufficient. If a serverless wrapper blocks runtime, parallelism, or observability requirements, return to CKS. If a product only calls a few foundation models, a managed model API is simpler than owning weights, runtimes, and capacity.

Capacity, portability, and failure tests

The same GPU model does not make clusters equivalent. Regional capacity, reservations, interconnect, CPU/RAM ratios, storage egress, support, and quotas affect delivery. Terraform and Kubernetes manifests, OCI images, portable checkpoints, and S3-compatible artifacts reduce switching cost, but network and performance tuning remain provider-specific.

Fail a node, storage mount, and network path before launch. Confirm jobs resume from checkpoints and inference drains and reschedules. CoreWeave fits teams for which GPUs are production infrastructure; small experiments and API-first products may prefer RunPod, Replicate, or a general hyperscaler.

References