API providers like OpenAI, Anthropic, ElevenLabs and Runway give you a URL, a key, and a price-per-call. Easy. They handle hardware, scaling, model updates. The downside: per-call cost adds up, you cannot customize the model, and your usage is visible to the provider.
Self-hosted open-weights — FLUX.1 Schnell on a 4090, SDXL on H100, MusicGen on Mac M-series — give you full control. The downside: you buy the hardware, manage the inference stack (vLLM, ComfyUI, diffusers), and eat the cold-start latency.
The crossover point is roughly 100K outputs/month. Below that, hosted APIs are cheaper because you do not amortize an H100. Above that, self-hosting starts to dominate.