Skip to content
Einsora

0006. Bill GPU allocation time, not utilization

  • Status: Superseded in part by ADR-0011 (no billing; usage accounting kept)
  • Date: 2026-10-22

Superseded in part by ADR-0011 (2026-10-23): Tensora is free, and there is no billing. The mechanism below, allocation-based records from lifecycle events, is kept as usage accounting for quotas, fair share, and reports (usage-accounting.md). Read every “bill” or “price” below as historical.

Context

Section titled “Context”

The original idea was to track GPU-hours from DCGM exporter metrics. But:

  • GPUs passed through to VMs are bound to vfio-pci and are invisible to host DCGM.
  • A reserved GPU costs us the same whether it’s busy or idle, so utilization-based billing invites “reserve and idle”.
  • Metrics pipelines are lossy (scrape gaps, restarts), which isn’t acceptable for invoices.

Decision

Section titled “Decision”
  • Bill wall-clock time in a billable state (running), per second, aggregated per UTC hour per workload. The source is the append-only workload_events table written by service.Transition.
  • Provisioning time is not billed. That’s a customer-friendly cost we absorb, and can revisit.
  • DCGM is used for observability and idle notifications only.

Consequences

Section titled “Consequences”
  • Billing correctness depends on lifecycle correctness. GPUs must be released before stopped or failed is recorded. A domain test checks that billable states hold capacity.
  • Controller downtime can overbill a workload that died. A crediting process is planned (see metering-and-billing.md).
  • Usage records are immutable and idempotent, so they’re auditable.

Alternatives considered

Section titled “Alternatives considered”
  • DCGM utilization: doesn’t work for VMs, and is lossy.
  • Kubernetes pod lifetimes from kube-state-metrics: closer to right, but not durable and not tied to our tenant model.

Update 2026-10-23 (superseded by ADR-0011)

Section titled “Update 2026-10-23 (superseded by ADR-0011)”

There are no prices and no billing. Usage is accounted by the class a workload actually ran on (placed_class, ADR-0010) and by inference gateway token counts, for quotas, fair share, and reports.