Before RunPod, renting a GPU meant enterprise contracts, opaque pricing, or crying on a forum.
RunPod made GPU rental a commodity. Pick a GPU on their marketplace — H100, A100, even consumer RTX cards — deploy a pod in under a minute, and start working. SSH in, run your training script, shut it down when you're done. Billing stops when the pod stops.
The pricing: spot H100 from ~$1.99/hr, on-demand from ~$2.79/hr. Cheaper than cloud for short bursts by a lot. No minimum commitment, no reserved capacity planning.
The killer use case: fine-tuning and evals. Spin up 8x H100, run LoRA for 6 hours, download the adapter, shut down. Total cost under $100.
The caveats: it's bare metal Linux, not managed. You install CUDA, drivers, Docker images. The docs help but you should know your way around a terminal.
Who it's for: researchers, small teams, and anyone who needs GPU compute occasionally rather than always. Think of it as AWS for the AI-curious, priced for humans.