Increase provisioned throughput and concurrency in Amazon Bedrock is the correct change to prioritize because increasing capacity reduces queuing and brings p95 latency back under target during high request rates.

Provisioning more throughput and concurrency adds parallel model capacity so requests are handled faster and dispatch decisions are not delayed by backend wait time. This change directly addresses the capacity bottleneck observed when traffic rises to about 450 requests per second and p95 latency grows from 250 milliseconds to 1.4 seconds.

Fine-tune the model with more route data may improve route quality over time but it does not relieve an immediate throughput or concurrency bottleneck. Model training and evaluation are slow to implement and will not reduce peak-time queueing.

AWS Global Accelerator can improve network path performance for some workloads but it will not reduce compute wait time when the model backend is saturated. It does not replace the need to scale Bedrock capacity when the service is the limiting factor.

Lower the maximum tokens per response can slightly reduce per-request compute time but it risks producing incomplete route guidance and usually provides only marginal latency improvement compared to increasing provisioned throughput and concurrency.

Full AWS Practitioner Certification Question