Skip to main content

​
Marketplace & Billing

Agnitra converts optimization telemetry into billable usage so monetization teams can charge on measurable outcomes (speedups, GPU hours saved) rather than seat licenses. This guide shows how billing artifacts are produced and dispatched.

​
Usage Event Model

Every optimization returns a UsageEvent that includes:
  • project_id, model_name, region, and optional metadata for attribution.
  • baseline and optimized metrics (latency, tokens/sec, GPU hours).
  • gpu_hours_saved, cost_savings, and total_billable amounts in the configured currency.
  • providers — list of marketplace targets that should receive the event.
  • Delivery receipts per provider (status, message, dispatched_at).
The CLI prints a human-readable summary while the SDK returns a structured object (result.usage_event) for programmatic pipelines.

​
Dispatch Flow

  1. Enable marketplace adapters with extras: 
    pip install agnitra[marketplace]
  2. Configure credentials via environment variables (for example AWS_ACCESS_KEY_ID, STRIPE_API_KEY, or custom secrets).
  3. Call agnitra optimize (CLI) or agnitra.optimize (SDK). The usage meter queues dispatch tasks after telemetry is persisted.
  4. Inspect the dispatch_results in the usage event to confirm delivery. Failed dispatches include retry hints or status: "deferred" when credentials are absent.
Marketplace exports are idempotent; events carry a deterministic usage_id so retries do not double-bill customers.

​
Stripe Metered Billing

  • Adapter: StripeUsageDispatcher.
  • Expects stripe_product_id and stripe_meter_id configured per project.
  • Sends usage shaped for Stripe’s metered billing API with timestamps derived from the optimization run.
  • Retries on 429/5xx with exponential backoff and surfaces final status in the usage event.

​
AWS Marketplace

  • Adapter: AwsMarketplaceDispatcher.
  • Requires IAM credentials scoped to the relevant product listing.
  • Converts usage events into BatchMeterUsage payloads grouped by dimension (often project_id + hardware tier).
  • Emits skipped status when credentials are missing so you can reconcile in post-processing jobs.

​
Custom Integrations

  • Implement the UsageDispatcher protocol to route events to internal billing, analytics, or support systems.
  • Reference implementation: agnitra/core/metering/dispatchers.py.
  • Register custom dispatchers via AGNITRA_TELEMETRY_EXPORT=internal,billing to enable multi-target fan-out.

​
Auditing & Compliance

  • Store raw usage events alongside telemetry for at least 13 months (typical cloud marketplace requirement).
  • Sign payloads before transmitting to third parties when operating in regulated industries.
  • Log dispatch_results to your SIEM to trace who was billed, when, and which kernel changes drove the charges.