A media startup serves 2M Bedrock image prompts monthly with spiky traffic, repeated similar prompts, and nightly bulk tagging jobs. Inference cost is soaring while p95 latency remains flexible overnight. Which cost-optimized AI design lowers Bedrock spend while preserving quality?

Prompt caching and semantic caching in ElastiCache eliminate repeat inference cost for similar prompts, intelligent routing selects the most economical model per request, and Bedrock batch inference processes deferrable nightly tagging at lower effective cost than real-time calls, directly addressing spikiness and repetition. Provisioned Throughput reserved continuously incurs fixed hourly commitment regardless of spikes and wastes capacity during idle periods for variable workloads. SageMaker real-time GPU endpoints add persistent instance cost, scaling delays, and operational burden when Bedrock serverless inference plus batch already fits the pattern. Pure on-demand real-time inference with maximum tokens maximizes token spend and provides no reuse or bulk discount, ignoring the flexible overnight window. For the exam match steady high-volume baseload to provisioned and variable plus deferrable to on-demand plus batch with caching, and in practice measure cache hit rate, token usage, and routing accuracy before committing.