CoreWeave (United States) is an AI cloud that added its own model-serving layer, CoreWeave Inference, available in three forms: token-billed serverless endpoints, dedicated deployments with your own weights, and self-managed inference on the platform’s Kubernetes.
What it does
- Calls models through an OpenAI-compatible endpoint without provisioning a GPU
- Publishes your own or fine-tuned models to dedicated deployments using vLLM or SGLang
- Runs inference in the same environment where the team already trains and tunes models
How it works
- The serverless mode is reached with a W&B Inference key and billed per token consumed
- Dedicated deployments reserve GPU capacity and accept your own weights
- Teams that want full control can run serving inside CoreWeave’s managed Kubernetes
- The platform handles scaling, replicas and compatible endpoints across all three modes
Strengths
- Training and serving on the same infrastructure, with no weight transfers across clouds
- OpenAI-compatible endpoints reduce integration work
- Serverless catalog with open models from several vendors
Points of attention
- Per-token billing applies to serverless; dedicated capacity and instances follow per-GPU-hour pricing
- The core business is still GPU infrastructure, so this card covers the serving layer
- Part of the catalog and enterprise features go through sales