From e375345bf6109a23425f8a8c2d35fee79eb062c7 Mon Sep 17 00:00:00 2001 From: Hanashi Date: Thu, 6 Aug 2026 20:08:03 -0400 Subject: [PATCH] SERVING-PLAN: per-output-second cost and pricing at all margins Co-Authored-By: Claude Fable 5 --- docs/SERVING-PLAN.md | 45 ++++++++++++++++++++++++++++++++++++++++---- 1 file changed, 41 insertions(+), 4 deletions(-) diff --git a/docs/SERVING-PLAN.md b/docs/SERVING-PLAN.md index 9227184..528d9b6 100644 --- a/docs/SERVING-PLAN.md +++ b/docs/SERVING-PLAN.md @@ -76,9 +76,44 @@ On the 5090-bf16 pool the same ≤480p jobs cost: 5 s $0.030 · 10 s ~$0.074 · 1-ref +~$0.01. Blend accordingly; the pricing below uses **B200 costs** as the conservative basis, so hybrid routing only improves your real margin. -### Pricing table (per video, 480p/SD baseline, t2v/i2v) +### Pricing — per output-second (the unit to reason in) -Margin means gross margin on price (price = cost ÷ (1 − margin)). +All prices are **$ per second of delivered video**. Margin means gross margin +on price (price = cost ÷ (1 − margin)). Note the honest wrinkle: our cost per +output-second RISES with clip length (compute is superlinear in frames), so a +truly flat per-second price quietly gives long clips a discount — the tables +show exact per-duration cost so you can decide flat vs tiered. + +**Our cost per output-second (B200 bf16 basis), and price at each margin:** + +| Tier | Clip | Cost/s | 0% | 15% | 25% | 50% | 70% | +|-------|-------|----------|---------|---------|---------|---------|---------| +| 352p | 5 s | $0.0097 | $0.0097 | $0.0114 | $0.0129 | $0.0194 | $0.0323 | +| 352p | 10 s | $0.0120 | $0.0120 | $0.0141 | $0.0160 | $0.0240 | $0.0400 | +| 352p | 15 s | $0.0142 | $0.0142 | $0.0167 | $0.0189 | $0.0284 | $0.0473 | +| 480p | 5 s | $0.0190 | $0.0190 | $0.0224 | $0.0254 | $0.0380 | $0.0634 | +| 480p | 10 s | $0.0238 | $0.0238 | $0.0280 | $0.0317 | $0.0476 | $0.0793 | +| 480p | 15 s | $0.0307 | $0.0307 | $0.0361 | $0.0409 | $0.0614 | $0.1023 | +| 768p | 5 s | $0.0630 | $0.0630 | $0.0741 | $0.0840 | $0.1260 | $0.2101 | +| 768p | 10 s | $0.0932 | $0.0932 | $0.1097 | $0.1243 | $0.1865 | $0.3108 | +| 768p | 15 s | $0.1234 | $0.1234 | $0.1451 | $0.1645 | $0.2467 | $0.4112 | + +**Hybrid-routing true cost** (≤480p/≤10 s jobs on 5090-bf16 workers): +480p drops to **$0.0059/s (5 s)** and **$0.0073/s (10 s)** — roughly 3× +cheaper than the B200 basis above. If you price from the B200 table and route +hybrid, realized margin is far above nominal (e.g. a "25 %" 480p price +carries ~77 % true margin on the 5090 share). + +**Flat-rate menu option** (blended 60/25/15 duration mix, hybrid routing): + +| Public flat price | True blended margin | +|---------------------|----------------------------------| +| SD 480p @ $0.02/s | ~50 % | +| SD 480p @ $0.03/s | ~67 % | +| HD 768p @ $0.10/s | ~20 % (HD stays on B200 — price HD at $0.13–0.17/s for 40–50 %) | +| Draft 352p @ $0.01/s | ~0 % on B200, ~60 % if drafts route to 5090s | + +Per-video equivalents of the B200-basis table (for invoice display): | Duration | Cost | 0% | 15% | 25% | 50% | 70% | |----------|--------|--------|--------|--------|--------|--------| @@ -158,8 +193,10 @@ Wait-time intuition (176 s free ref job): start in ~4 min (measured: container rebuild ~4 min + first model load ~40 s on a warm volume). -Rule of thumb: **free tier costs ~$45 per 1,000 videos served** — the worker -math above just determines how fast you serve them. +Rule of thumb: **free tier costs ~$45 per 1,000 videos served, i.e. +~$0.0087 per output-second** — the worker math above just determines how +fast you serve them. (For comparison: every paid SD second sold at $0.02 +funds ~2.3 free seconds.) ### Paid pool (hybrid; blended job assumed 60 % SD-5s / 20 % SD-10s / 10 % HD-5s / 10 % ref-heavy)