Sessions on saturating GPUs, batching, KV-cache reuse, and the unit economics of inference on Lumenrack Compute.