Independent Benchmark Finds X25519MLKEM768 Adds ~3ms Latency Over Classical TLS at p99

A developer conducted an independent, reproducible benchmark comparing post-quantum hybrid key exchange X25519MLKEM768 against classical X25519 on two AWS Graviton3 EC2 spot instances, with the full test costing under one dollar. The experiment used nginx built against OpenSSL 3.5.0 as the target server and Gatling as the load generator, both running on c7g.large instances in the same AWS availability zone. An initial test at 1,000 requests per second proved unreliable due to client-side CPU saturation, which masked any meaningful difference between the two algorithms. Reducing throughput to 300 requests per second yielded cleaner results: median latency was nearly identical, while p99 latency for the PQ arm was roughly 3ms higher — a 60% relative increase — though one trial recorded an outlier of 39ms, possibly due to CPU steal or JVM garbage collection. The author notes this reflects real-world 2025–2026 PQ TLS deployments, where classical certificates are still used alongside hybrid KEMs, and plans a second phase to investigate the outlier and run more trials.