It'll be based on sustained queue length averages rather than any kind of load number. We can also bump it up manually.

Our estimates for some fairly severe traffic based on historic data are quite reasonable for bandwidth costs. But we're a medium sized business, so we have revenue to support that. I think individuals and very small teams can get really surprised by this.

If you're just starting out and don't need dynamic scaling and the tooling that AWS provides, I'd seriously look into colo so that you can buy bandwidth using 95th percentile billing. We use a combination of servers that we own in data centers using 95th billing and AWS. Depends on use case and we have gotten very good at financially modeling future costs for a given config.