CoreWeave
GPU cloud infrastructure for AI training, inference, and HPC workloads.
About CoreWeave
CoreWeave is a cloud platform for AI and high-performance computing workloads. It runs managed Kubernetes on bare metal for training, inference, and HPC, while Slurm on Kubernetes handles batch and burst jobs. Teams can choose serverless or dedicated inference, S3-compatible object storage, shared POSIX file storage, local scratch storage, and isolated virtual networks. The service suits organizations that need recent NVIDIA GPU infrastructure and the operational controls for large containerized workloads. Pricing is usage based and changes with the hardware and service selected. Hourly rates are available for some configurations, while newer systems may require a sales quote. CoreWeave can fit teams with substantial accelerator demand, but a useful budget still depends on the exact GPU, region, storage design, and purchase option.
Best for
Best use cases
A distributed training run needs multiple GPU nodes, shared checkpoints, and scheduling within one managed environment.
An inference service must move between serverless and dedicated capacity as request volume and latency needs change.
A batch pipeline needs Slurm scheduling and temporary GPU capacity without building a separate physical cluster.
Compare before choosing
Compared with a general-purpose cloud, CoreWeave concentrates its product set on GPU-intensive AI and HPC infrastructure.
Published hourly rates help compare selected hardware, but configurations marked contact sales need a quote before budgeting.
Managed Kubernetes and Slurm options support different operating models, though both require cloud infrastructure expertise.
Features
Large model training
Run distributed AI training on managed bare-metal Kubernetes with storage and scheduling options designed for multi-node jobs.
Production AI inference
Serve models through serverless or dedicated inference options when workloads need scalable GPU-backed deployment.
Batch and HPC workloads
Use Slurm on Kubernetes for scheduled batch work, burst capacity, and compute-intensive scientific or engineering jobs.
Pros
- Specialized NVIDIA GPU infrastructure supports demanding training, inference, and HPC workloads.
- Managed Kubernetes and Slurm options accommodate both container services and scheduled batch computing.
- Published hourly prices for selected configurations can support direct workload cost estimates.
Cons
- Hardware-dependent pricing makes costs harder to estimate without a detailed workload and capacity model.
- Some newer GPU configurations require a sales quote, which slows direct price comparison and procurement.
- Kubernetes, Slurm, storage, and networking choices require experienced infrastructure operators.
Pricing Plans
Usage-based infrastructure
- GPU and CPU configurations priced by selected hardware
- On-demand and eligible inference rates
- Some configurations require contacting sales
* Pricing information is for reference only. Please refer to the latest announcement on CoreWeave's official website.