Running an AI side project profitably in 2026
The cost surfaces that will eat your margin if you're not watching them.
Token costs are the obvious surface but rarely the killer. The killers are silent: a free tier abused by a single user generating a million tokens overnight, a webhook that retries forever, a background job that re-embeds the entire corpus on every deploy.
Put hard per-user rate limits in place before you launch, not after. Cap maximum generation length per request. Cache aggressively, especially embeddings — they almost never need to be recomputed for the same input.
On pricing: usage-based looks fair but is hard to predict for buyers. Tiered with a generous-but-finite quota converts better and lets you forecast revenue. Charge per seat where the customer is a team, per request only when the customer is a developer.