Why are your cold starts sub-200ms? What tradeoffs did you accept?

Seeing a lot of FaaS providers claim cold starts under 200ms, but the fine print usually excludes real-world conditions (VPC attachments, EFS mounts, custom runtimes). For those who've actually shipped production serverless with aggressive SLAs: 1. What provisioning strategy got you below 200ms? (provisioned concurrency, snapshot-based, something else?) 2. What did you sacrifice to hit that number? (memory ceiling, init complexity, vendor lock-in?) 3. Did you measure p99 or p95? The marketing decks always show averages. Looking for war stories, not benchmark results from empty Lambdas.