SkyPilot launches jobs across clouds. denv serves the model.

SkyPilot is the closest thing in this list to infrastructure rather than a token API. It helps you run jobs on more than one cloud. Teams use it to find GPUs and to start a vm or a cluster without rewriting the launch script per provider. That is a real problem, and SkyPilot is a reasonable answer to it.

denv answers a narrower problem: serve an open-weight model, continuously, behind one endpoint.

What denv is opinionated about

• A recipe pins the engine, the weights and the GPU type, and it is tested before it serves.
• The cluster is Kubernetes, including llm-d when you want cache-aware routing.
• Health checks and auto-repair are in the product, not a script you attach later.
• Failover is across sites in one cluster. GPU clouds are providers. The phrase we use is “like EKS on top of EC2, but for inference.”
• You pay the GPU provider. denv does not resell the machine.

What this page will not claim

It will not claim SkyPilot cannot do failover, serve vLLM, or span regions. People build all of that on it. The difference is who assembles it. SkyPilot leaves the serving stack to you. denv ships one. If you want the launcher and you are happy to own vLLM, repair and the gateway, stay on SkyPilot. If you want those decisions already made and pinned, look at denv.

denv is early. SkyPilot is open source and widely used. Read both docs before you pick. The decision map is the inference guide.