Delivery modes and operating boundary
CoreWeave currently documents three inference paths: Serverless Inference, Dedicated Inference, and self-managed inference on CoreWeave Kubernetes Service (CKS), alongside GPU cloud instances and bare-metal accelerated clusters.
Serverless Inference is an API-managed path with no cluster operation; Dedicated Inference lets customers choose model weights, GPU class, runtime and scaling while CoreWeave operates routing and lifecycle; inference on CKS leaves the serving stack, scheduling and autoscaling with the customer.
Computraq is prelaunch. Naming this provider here does not mean that Computraq hosts workloads on it, has tested its integration, or has a commercial relationship with it.
Buyer and procurement preparation
Use these questions to prepare a comparable, workload-specific evaluation. Product names describe the provider’s published offering; they do not establish availability for a particular account.
- Choose serverless API, managed dedicated serving, or self-managed CKS based on model control and infrastructure responsibility; compare the customer’s operational duties explicitly.
- For Dedicated Inference, verify model artifact requirements, GPU class, supported runtime, zone and scaling controls; for CKS, scope cluster operations and reserved/on-demand capacity.
- Ask for the service commitment, capacity term, data-handling terms, support scope and total cost inputs in a written quote.
Developer and operator evidence
Before selecting capacity, create a small acceptance plan that records the model and revision, input shape, runtime, GPU configuration, concurrency, target latency or job completion, and recovery expectations. Provider documentation is evidence of product design, not proof that this workload has been tested.
- For the managed API, test request compatibility, model version and rate/usage limits; for Dedicated Inference, pin model weights, GPU class and vLLM or SGLang runtime.
- Test startup, model loading, autoscaling, gateway routing, latency and model rollout with representative request sizes before production.
- For CKS, own and test the serving stack, scheduler, autoscaling and multi-node topology; measure network communication separately from device compute.