AI products require compute in three different scenarios

Model vendors need to build, train, and operate their models, and all of this requires large amounts of fast compute.

1 · Model builderPre-training

Rare, expensive work that creates the general base model.

2 · Builder or customerFine-tuning

Smaller adaptation work that shifts behavior and format.

3 · Product userInference

Daily usage: prompts in, tokens out, paid per request.