How to Use Replicate the Right Way in Your Next.js App (And Ship a Real Product With It)

Most tutorials show you how to call Replicate. Few show you how to use it well inside a real production app. This article covers the mistakes I made and the patterns that actually work — using Goodbye Watermark as a real-world case study. What Is Replicate, Really? Replicate is a cloud API that lets you run AI models — image generation, video, audio, vision — without owning a single GPU. You send an HTTP request, a model runs on their infrastructure, and you get the result back. The business model is pay-per-prediction: you're charged for the time the model actually runs, not idle time. That means cold boots don't affect your cost — only your latency. 1. Understand the Prediction Lifecycle Before Writing Any Code Every Replicate call creates a prediction — an object with a lifecycle: starting → processing → succeeded (or failed / canceled) starting : model is booting (cold start happens here) processing : predict() is actively running succeeded : output is ready — but files are deleted