Skip to content

PRICING.md

Latest commit

 

History

History
142 lines (105 loc) · 5.42 KB

PRICING.md

File metadata and controls

142 lines (105 loc) · 5.42 KB
Raw

DockVision — Pricing

Model: cost-plus, transparent

Every job shows both the real GPU cost charged to us by RunPod and the DockVision markup that funds platform overhead, failure absorption, and margin.

billed_cents = ceil(executionTime_seconds × gpu_rate_cents_per_sec × 1.5)

The 1.5× covers:

  • Failed jobs we still pay RunPod for (~5% of runs)
  • MSA service compute (amortized across Boltz/Protenix runs)
  • S3 egress to user downloads ($0.09/GB)
  • Reconciliation / monitoring infra on www0
  • Margin

If you want to skip the markup, the platform is AGPL-3.0 — self-host. Bring your own RunPod, Stripe, and S3 — pay providers directly. We sell ergonomics, not access.

GPU rates (Flex, RunPod, May 2026)

GPU VRAM $/sec (RunPod) $/sec (you pay) $/hr (you pay)
RTX 4090 24 GB $0.000310 $0.000465 $1.67
L40S 48 GB $0.000530 $0.000795 $2.86
A100 80GB 80 GB $0.000760 $0.001140 $4.10
H100 80GB 80 GB $0.001160 $0.001740 $6.27
H200 141 GB $0.001550 $0.002325 $8.37
B200 180 GB $0.002400 $0.003600 $12.96

Rates re-synced weekly from RunPod's pricing endpoint.

Per-tool typical compute cost

Calibrated against CASP15/16 benchmark data from ~/data/vaults/casp — these are real observed costs from prior runs.

Tool GPU Typical runtime Typical cost (user)
GNINA RTX 4090 1-5 min $0.03 - $0.14
Boltz-2 H100 2-8 min $0.21 - $0.84
Protenix H100 3-10 min $0.31 - $1.05

Each tool's landing page shows a live pre-flight estimate based on input size at submit time.

Storage

We also bill for storage, separately from compute. Most users won't notice until they accumulate big runs.

  • First 5 GB: free. Covers most users for their first month or two.
  • Above 5 GB: $0.04 per GB-month, billed daily as (GB - 5) × $0.04 / 30 per day.
  • Cold tier at 60 days: S3 lifecycle moves objects to S3 Infrequent Access. Cold pricing drops to $0.0125/GB-month internally — we pass that through as $0.015/GB-month with a 1.2× markup on cold files.
  • Auto-delete at 180 days: anything older than 180 days and pinned = false is deleted. pin <path> makes a file immune.

A daily billing_events row with kind='storage' and a brief description ("Storage: 12.4 GB · $0.030") keeps the ledger auditable.

Failure billing policy

  • Worker crashes, OOM, network errors: no charge to user. We eat the RunPod cost.
  • Tool returns failure status: no charge.
  • User cancels mid-run: charged for elapsed executionTime up to cancel.
  • Job exceeds tool-specific max-runtime cap: hard-killed, no charge.

We track absorbed cost in billing_events with kind='adjust' and a negative "platform absorbed RunPod cost" entry to a system account — keeps the books straight.

Pre-flight estimate

Before any run, the CLI shows:

$ run boltz2 --target /inputs/T1313.json
estimated cost: $0.42  (240s typical × $0.00174/s × 1.5)
balance after: $9.58
proceed? [y/N]

Pre-flight uses typical_runtime × gpu_rate × 1.5. Actual cost is computed from real executionTime and can differ by ±50% in worst case (model convergence, input size).

Hard caps

  • Submission rejected if balance_cents < estimated × 1.2 (20% safety buffer).
  • Mid-run auto-cancel: cron checks running jobs every 30s; if balance drops negative, RunPod cancel sent.
  • Max 5 concurrent running jobs per user.
  • Max 100 job submits per rolling hour per user.
  • Max 100 GB per file upload.

Top-up

  • Minimum top-up: $10.
  • No free tier in v1.
  • Stripe Checkout in test mode initially; switch to live before public launch.
  • Balance never expires.

Refunds

  • Reversed on request for documented platform bugs within 30 days.
  • Top-ups non-refundable after 30 days of inactivity (state aggregator law in some US states + Stripe rules).
  • Storage charges incurred on data that became inaccessible due to our error are auto-refunded.

Self-host pricing

The AGPL-3.0 source at github.com/sness23/dockvision builds a fully working instance. Self-hosters pay:

  • AWS EC2 + S3 (or equivalent)
  • RunPod (per-second GPU)
  • Stripe (transaction fees only; no need to use the prepay flow internally)

No DockVision license fee. AGPL requires that if you operate a modified version as a network service, you publish the modifications. Read the license.

Why 1.5× and not 2× or 3×

  • Fovus runs Boltz-1 at $0.10/structure as a public floor.
  • Neurosnap academic credit ≈ €1.10 = ~$1.20, and a typical Boltz run is 1-3 credits = ~$1.20-3.60 — that's a ~3-5× markup over raw RunPod.
  • 2× to 3× is the easy-money zone, but it invites a transparent competitor.
  • 1.5× undercuts Neurosnap meaningfully while still funding overhead. If reconciliation shows we're losing money on failures or MSAs, tune up to 1.6-1.8× with a public changelog entry. If we discover a moat (better viewers, better CASP integration), we can tune up; if churn says the markup is felt, tune down.

The markup is a config constant. Changes are logged.