Video is a long job on a GPU we have not published a number for.

Video generation holds a GPU for a long time and uses a lot of memory. A node that dies at minute forty wastes the forty minutes. A provider with no free cards wastes the launch.

This site has no video recipe and no published video throughput, latency, or cost. The H100 results are a text model, Qwen3.8-27B. They do not describe a video model. I am not going to put a tokens-per-second figure here, because we did not measure one.

What would still be true if you brought the GPUs

denv does not rent the GPU. Lambda and Nebius list B200. H100 and H200 are in the catalog too. Your own server can join. You pay whoever owns the machine.

A job that must stay on one fast interconnect stays in one site. The cluster article says not to stretch a single replica across providers. A second site is a second replica, or the next job, not half of this one.

Health checks and repair still apply to the node: XID errors, a GPU that falls off the bus, a driver reinstall, a replacement when the provider account can launch one. They do not resume a half-rendered video. They stop you serving the next request on a dead card.

A cold start measured on one fresh H100 for a text model was 12 minutes 46 seconds, then 99 seconds with a warm cache. A video checkpoint is larger. We have not measured it. Treat the text number as a floor you should not plan around, not as a video startup time.

If you have a video model and the GPUs, write to anup@denv.ai. Until a recipe has been tested on that GPU type, it is not in the catalog, and this page will stay this short.