CoreWeave vs Replicate Deployments

CoreWeave

Per-hour prices for 8-GPU H100, H200 and B200 nodes are public, the docs ship as Markdown with llms.txt and embedded OpenAPI specs, and IAM has read-only roles per service. An organisation must be approved by CoreWeave's sales team before any token exists, and no request rate limits were found in the reviewed documentation.

Replicate Deployments

OpenAPI file, llms.txt and an MCP server with a two-tool code mode. Private instances bill set-up and idle time, H100 at $5.49 an hour.

CoreWeave

1. Create an API access token at console.coreweave.com/tokens with an expiry and send it as Authorization: Bearer to https://api.coreweave.com
2. Add GPUs by applying a NodePool resource (compute.coreweave.com/v1alpha1) to a CKS cluster with instanceType and targetNodes. instanceType can't be changed afterwards
3. Set targetNodes to 0 or delete the Node Pool when a job ends. Nodes are whole 8-GPU machines billed by the hour
4. Never pass dry_run to the MCP tool coreweave_kubectl_apply expecting a preview. The reference says the parameter is ignored and the manifest is applied
5. Use a separate Forge API key and https://api.inference.wandb.ai/v1 for Serverless Inference. A CoreWeave API access token doesn't work there

Replicate Deployments

1. List GET /v1/hardware first and use the returned sku in the deployment body
2. Set min_instances to 0 for bursty work; a warm H100 bills $5.49 an hour whether called or not
3. Send Prefer: wait on deployment predictions to block instead of polling
4. Copy outputs within an hour; API prediction data is deleted after that
5. Wait for the reset time in the 429 body before retrying; prediction creates cap at 600 a minute