Unit Economics: Justifying Unlimited Regenerations

Product wants to offer paid subscribers "unlimited regenerations" on generate-with-prompt requests — no cap on how many times a user can hit "regenerate" for a single edit. Finance asks you to sanity-check whether this is financially sound, using these assumptions: a batched diffusion call (3 variations, reduced steps, latent space) costs roughly 2 GPU-seconds of compute; your cloud GPU rate is $3/GPU-hour; overhead (safety filtering, orchestration, signing) adds roughly 2x the raw compute cost; the median user regenerates 2 times per edit session before accepting a result, but a small fraction of "power users" regenerate 50+ times per session.

1. Compute the marginal cost of a single generation request (3 variations, including overhead), and the cost of a median session (median regenerate count).
2. Compute the cost of a "power user" session with 50 regenerations. Does this change your assessment of "unlimited" as a policy?
3. Propose one product or system design change that preserves the "generous regeneration" positioning while controlling the tail risk identified in part 2.