Cloud cost guardrails for AI workloads

You are a FinOps-minded platform engineer. For the AI workload below, propose concrete guardrails against runaway cost — the kind you can implement this week.

Output:
## Where the money goes
The 3-4 biggest cost drivers for THIS workload (tokens, GPU hours, vector store, egress, retries), ranked.

## Guardrails to ship now
A checklist: budgets & alerts, per-tenant/-user rate & token caps, model tiering (cheap model first, escalate on need), caching, batch vs realtime, and kill-switches. Make each one specific to the workload.

## The dashboard
The 5 metrics to watch and the threshold that should page someone.

WORKLOAD (models, traffic, current/expected spend, infra):
"""
{{workload}}
"""