AI workload spend caps (filled example)

Acme's AI features call Bedrock (production) and Vertex AI (a batch enrichment job). Both have per-environment quotas and budget alerts so a runaway loop or abuse can't produce an unbounded bill, and…

Download whole kit View on GitHub

Illustrative example for a fictional company. See the blank template at ../06-ai-workload-controls/spend-caps.md.

Evidence on file

Artifact	What it shows
`bedrock-quotas-2026-06-20.json`	Provisioned/throughput limits + per-model access policy
`budget-alerts-2026-06-20.json`	Budget thresholds (50/80/100%) + notification channels
`alert-routing-2026-06-20.md`	Where a cap breach pages (PagerDuty → on-call SRE)

Gathered from

# AWS Bedrock spend alarm + service quotas
aws cloudwatch describe-alarms --alarm-name-prefix "bedrock-spend" > bedrock-alarms.json
aws service-quotas list-service-quotas --service-code bedrock > bedrock-quotas-2026-06-20.json

# GCP budget for the AI project
gcloud billing budgets list --billing-account=01ABCD-234567-89EFGH > budget-alerts-2026-06-20.json

Controls in place

Per-model rate limits on the production Bedrock endpoint (capped at 60 req/s, well above steady-state, low enough to bound abuse).
Budget alerts at 50 / 80 / 100% of a $4,000/month AI cap, routed to #ai-oncall and PagerDuty.
Hard stop: at 100% a Lambda disables the non-critical Vertex batch job (the production inference path stays up; only discretionary spend is cut).

Buyer questions answered

"How do you prevent runaway AI spend or abuse?"

Per-model rate limits bound request volume; tiered budget alerts (50/80/100%) page the on-call SRE; the 100% threshold automatically disables the discretionary batch workload. The most recent alert (2026-05-30, 80% threshold) was a legitimate traffic spike and resolved without action.

Refresh

Quarterly. Last refreshed 2026-06-20 by L. Favre. Next due 2026-09-20.