The real cost of self-hosting

The real cost stack

Open-source models like Wan make the generation itself free — and move the bill elsewhere: hardware (a capable consumer GPU runs €1,500–3,000; serious throughput wants more), power (roughly €0.05–0.15 per generation-heavy hour at EU rates), your time (setup, updates, babysitting failed runs), and slower iteration than hosted queues unless you overbuy hardware.

The break-even math

Against a $0.05/s hosted rate, a €2,000 GPU pays for itself after roughly 11 hosted-equivalent hours of kept footage — which sounds fast until you translate it: at a heavy 20 minutes of kept output weekly, break-even sits around 8–10 months, ignoring your setup hours. Rent-a-GPU services (hourly cloud GPUs) split the difference and suit bursty workloads without the capex.

Who should actually do it

• Teams needing fine-tuning — custom style lock is self-hosting's unique product, not the savings.
• True volume pipelines generating hours monthly, where marginal cost near zero changes the business model.
• Privacy/control requirements that rule hosted APIs out.

For everyone else, hosted budget tiers at $0.04–0.05/s have already priced most of self-hosting's advantage away — without the weekend of CUDA errors.