How do you keep an FM application fast and available at the lowest cost? These notes compare the Bedrock options for capacity: reserved model units, inference profiles that route across Regions, asynchronous batch jobs and routers that pick a cheaper model when quality allows.
| Feature | Geographic | Global |
|---|---|---|
| Data residency | Within a geography such as US, EU or APAC | Any supported commercial Region |
| Cost | Standard pricing | About 10% savings |
| SCP requirement | Allow all destination Regions in the profile | Allow aws:RequestedRegion equal to unspecified |
| Best for | Data residency regulations | Cost optimization |
Source: docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html
Provisioned Throughput reserves model units (MUs) at a fixed hourly price, with no commitment or a 1-month or 6-month term (longer = cheaper, and you cannot delete it before the term ends). A customized model can only be invoked through Provisioned Throughput.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/prov-throughput.html
Inference profiles (cross-Region inference) currently do not support Provisioned Throughput. To raise throughput for a model you buy Provisioned Throughput; to spread load across Regions you use an inference profile.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html
A geographic inference profile (for example US, EU or APAC) keeps processing inside that geography; a global profile routes to any supported commercial Region and costs about 10% less. Choose geographic when you have data residency rules.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html
There is no extra routing cost: you pay the price of the Region you call the profile from. Data stays on the AWS network, and CloudTrail records the request in your source Region with the processing Region in additionalEventData.inferenceRegion.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html
Batch inference processes many prompts asynchronously: you upload input files to S3, submit a job and read the results from S3. It is not supported for provisioned models, and Converse-type jobs cannot use tool calling (use the InvokeModel type).
Official source: docs.aws.amazon.com/bedrock/latest/userguide/batch-inference.html
Intelligent prompt routing gives one endpoint that predicts response quality for each request and routes it between two models of the same family to balance quality and cost. It is optimized for English prompts only.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/prompt-routing.html
Default routers use select Amazon Nova and Meta Llama models. A configured router lets you pick the models, a fallback model and the response quality difference: the router only switches to the other model when its predicted quality beats the fallback by more than that percentage.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/prompt-routing.html
Answer practice questions, then review your mistakes with spaced repetition.