Some jobs need a GPU. On the Premium tier, a job can request an NVIDIA T4 node, submitted through the same job API as any other container - you add a GPU flag and the job lands on GPU hardware.
GPU capacity is usage-billed and scales to zero when idle, so there is no always-on GPU cost between runs. A GPU job has a whole node to itself, so its CPU and memory are included in the GPU rate rather than billed on top.
GPU jobs run on your own dedicated cluster inside your isolated VPC, the same as every other workload, and appear on the same single invoice.