AI Usage Billing

A reference use case for token-metered products that need hard quota enforcement and a durable billing ledger.

AI usage billing is one of the clearest fits for Parix.

The current product contract already targets AI teams that need real-time quota enforcement and a durable usage ledger without running TigerBeetle themselves.

Where this pattern fits

Use this shape when you need to:

meter requests, tokens, credits, or model usage as ledger events
reject over-limit traffic deterministically instead of reconciling later
retain billable usage history for invoicing and support workflows
separate customer-facing product logic from the underlying ledger infrastructure

Typical flow

A request enters your application with an organization, workspace, or API key identity.
Your application checks quota state and writes the usage event to the ledger path that backs billing.
Parix stores the durable event trail used for reconciliation and downstream billing jobs.
Aggregation jobs convert raw usage into priced line items and invoice-ready totals.

What the repo already supports

The repository already contains the pieces for an hourly usage-billing flow:

D1 usage tables for bucketed usage and priced ledger events
queue-backed aggregation for profile-hour billing windows
cron-driven retry and backfill behavior
Stripe billing setup and invoice generation workflows

For operational detail, see the existing runbooks:

Current managed cluster pricing reference

The current catalog and billing-rate source of truth lives in the canonical product spec:

docs/product-specs/pricing-model.md

Current active single-replica cluster floors:

Provider	PX-5	PX-10	PX-20	PX-40	PX-80	PX-160
GCP	$85	$160	$210	$600	$1210	$2420
AWS	region-aware	region-aware	region-aware	region-aware	region-aware	region-aware

Current sizing policy:

PX-5: 1 vCPU / 4 GiB and treated as development-only
PX-10: 2 vCPU / 8 GiB
PX-20: 2 vCPU / 16 GiB
PX-40: 2 vCPU / 32 GiB
PX-80: 2 vCPU / 64 GiB
PX-160: 2 vCPU / 128 GiB

Practical deployment guidance:

If you need a portable production default, start at PX-10 or PX-20.
Memory-heavy tiers above PX-20 are now available on both GCP and AWS, but the quoted monthly price depends on the selected region and storage backend.
The usage-ledger invoice path is seeded separately from the catalog and carries provider-scoped markup through billing_rate.

Why TigerBeetle helps here

This pattern benefits from strong ordering and durable writes under contention:

quota checks and ledger writes need clear failure semantics
burst traffic cannot tolerate race conditions around remaining balance
reconciliation should be driven from append-only facts, not reconstructed state