AWS open-sources Amazon Bedrock Invocation Data Lake

Amazon Bedrock already offers model invocation logging, which can deliver full request and response records to Amazon S3 or CloudWatch. The limitation is that this data stays locked inside individual accounts, so an organization running Bedrock across many teams cannot easily get a single queryable view. The new sample data lake addresses that by using CloudWatch subscription filters to forward logs from linked accounts to a central account, where Amazon Data Firehose converts them to Parquet and Amazon Athena exposes one table.

The design separates permissions: source accounts hand events to a CloudWatch Logs destination in the central account and do not get access to the central pipeline. A small Lambda function unwraps CloudWatch's batched, encoded records, discards housekeeping messages, and flattens the nested caller identity field. Everything is deployed as infrastructure as code with two CloudFormation templates, and the repository includes an optional Quick Sight dashboard for visualizing token consumption over time.

The post also ties the data lake to tokenomics, a practice the 2026 FinOps Foundation conference announced as a formal FinOps domain built on the See, Save, Run, and Plan pillars. With per-invocation visibility into prompts, responses, token counts, models and IAM identities, teams have the foundation to match each workload to the model that fits it, rather than tracking only an aggregate Amazon Bedrock bill. Start with one or two linked accounts before expanding across the organization.