Case study // Sable Compute
Metered billing for 180k GPU-hours a day
A usage-based billing pipeline that rates every GPU-second in real time and reconciles to the cent across three clouds.
Revenue leakage
−97%
Rating latency
220ms
Stack
ClickHouse · Kafka · TypeScript · Stripe · Kronos Ledger
01 — Challenge
Sable billed from hourly cloud exports. Preemptions and spot reclaims were lost between files, leaking an estimated 6% of revenue.
02 — Architecture
Per-second usage events are signed at the node agent and deduplicated by monotonic sequence.
A streaming rater applies contract pricing, commits and credits before writing to a double-entry ledger.
The customer-facing usage API reads from the same ledger as invoicing, so there is one source of truth.
03 — Outcome
Leakage dropped from about 6% to 0.18%. Customers see the exact figure that will appear on their invoice, live.