Most surprise GPU bills have the same story: a test cluster that kept running, a job that scaled further than planned, a new model that quietly doubled inference costs. Nobody decided to spend the money; nobody was asked.
Automate execution, not judgment
You don't need a committee for every GPU hour. You need a person to approve the moments when spend changes in kind, and automation for everything else.
The checkpoints that matter
- New cluster or new GPU type: someone confirms the size, the term and who owns it.
- Moving from on-demand to a reservation: check utilization data before committing for months.
- Scaling past an agreed ceiling: autoscaling stops at a limit and asks before going further.
- A new model in production: estimate cost per request before launch, not after the first invoice.
Make approval cheap
An approval should take a minute: a message with the size, the hourly rate, the expected duration and the total. Budgets per team, alerts at 50%, 80% and 100% of a monthly limit, and automatic stop rules for idle resources handle the rest.
What it looks like with Strata
On-demand servers run from a prepaid balance, so spend can never silently outrun what you topped up. Reserved clusters come with a fixed invoice for a fixed term. Both make the moment of decision explicit, which is the point.
