# Lambda On-Demand vs Provisioned Concurrency — Worksheet (July 2026)

Fill with **us-east-1** (or your region) prices from [Lambda pricing](https://aws.amazon.com/lambda/pricing/). Rounded to five minutes for PC duration.

## Inputs

| Input | Value |
| --- | --- |
| Memory (GB) | |
| Average duration (s) including INIT when cold | |
| Steady concurrent executions (avg) | |
| Peak concurrent executions | |
| Hours/day PC would be enabled | |
| Need sub-100ms p99? (Y/N) | |

## Rough monthly PC cost (always-on)

`PC units × memory_GB × 3,600 × hours_per_month × PC_GB_second_price`

Example (illustrative): 10 units × 0.5 GB × 3,600 × 720 × $0.0000041667 ≈ **$54/mo** (verify current rate).

## Decision

| Condition | Prefer |
| --- | --- |
| Spiky / idle nights, latency OK | On-demand (+ SnapStart where supported) |
| Steady baseline concurrency, latency SLA | PC on the **baseline**, not the peak |
| Peak-only windows | Scheduled Application Auto Scaling on PC |
| PC utilization ≪ ~60% of provisioned time | Usually on-demand is cheaper |

## Mitigations before buying PC

- [ ] Power Tuning memory (cost ≠ latency optimum)
- [ ] arm64 / Graviton where compatible
- [ ] SnapStart (Java 11+, Python 3.12+, .NET 8+)
- [ ] RDS Proxy / connection reuse to cut INIT