VBH package margin tool

Interactive version of the ai-cost-per-user model (Aug 26, 2026: post-repricing $3/$5/$15, Premium budget 750k, all AI pools capped). Edit any input — every table recomputes live. Single file, no network calls.

Inputs

Per-tier price, match-AI token budget, blended $/1M and user-mix share are editable directly in the first table.

Per-tier economics @ 40% utilization

AI $/mo = match budget × utilization × blended rate + doc-AI cap × utilization × $1.50/1M + scoring (typical runs × 1.5k tokens × $10/1M). Blended = 0.8 × input + 0.2 × output price.

If EVERY quota is used 100% — worst-case cost per user

Adversarial floor: full match budget + full doc-AI cap + max scoring runs, every month. Narration caching (168h) and the 24h scoring cooldown make this hard to reach legitimately. Red line = margin target.

Overall platform spending per month

Net earnings — what you keep after every cost

Net = price − AI spend − payment fees (toggle above). The Docker host is self-owned (sunk), so there is no other per-user cost. "Typical" uses the utilization slider; "worst case" assumes 100% of every quota by every user.
How to read these numbers
Net per user / mo, typical — the profit you keep on an average user in a normal month: mix-weighted revenue minus AI spend at the utilization set on the slider, minus payment fees. This is the number to expect in practice.
Net per user / mo, worst case — the same figure assuming every single user burns 100% of every quota, every month: full match-AI token budget + the entire document-AI cap + the maximum scoring runs. It is an average across the user mix, not any single tier, and it is a floor, not a forecast — the caps make it mathematically impossible to earn less per user than this, even under coordinated abuse. (In reality the 168h narration cache and the 24h scoring cooldown make reaching 100% very hard.) Live breakdown:
Max abuse damage $/user — typical net minus worst-case net for that tier: the most extra money one deliberately abusive user can cost you beyond normal usage. It is a hard ceiling enforced by the spend caps, not a risk estimate. Before the Aug 2026 hardening this was effectively $32,400+/month per abusive user (uncapped document extraction); now the worst possible user in every tier is still profitable. Exception: Enterprise match AI is unmetered by contract, so its figure rests on the ~3M-token assumption rather than a code-enforced cap.