Product · CloudNerve
TokenLedger
What each customer, feature and model costs you on Amazon Bedrock, every day.
A daily job that joins your AWS cost data with your Bedrock invocation logs and splits the real Bedrock bill across tenants, features and models, so you know your margins each day, not at month end.
Who it's for
SaaS companies and ISVs that sell features built on Amazon Bedrock and need cost and margin per customer.
When teams call us
Bedrock spend growing faster than revenue, a pricing decision for an AI feature, or a customer whose AI usage nobody can price.
What it does
Inside TokenLedger
Reconciled allocation
Every euro goes to a tenant, a feature or a visible unattributed line, and the total matches the invoice.
Unit economics
Cost per 1,000 requests by tenant, feature and model.
Gross margin per tenant
Add a monthly revenue file and see which customers are profitable.
Prompt-cache savings
What prompt caching saved, measured from your own traffic.
Provisioned throughput break-even
When buying provisioned throughput would pay off.
Routing what-if
What a cheaper model would have cost on your real requests.
Tenant anomalies
Tenants whose cost moved unusually, flagged daily.
What you receive
On every run
- A single-file HTML report, printable, with a reconciliation box and data freshness
- A CSV file for every table
- A daily dataset that later runs use for trends and month-to-date margin
- An optional e-mail with a link to the report
How it works
Step by step
- 01
Turn on the data
Bedrock model invocation logging to Amazon S3, and a CUR 2.0 export with resource IDs.
- 02
Stamp requests
A small open helper (Python and TypeScript) adds tenant and feature to each Bedrock call. Without it, cost is still split by IAM principal and model.
- 03
Deploy
One CloudFormation stack; a Lambda function runs daily on a schedule.
- 04
Read the report
Finance and engineering get the same numbers every morning.
Built on AWS
The AWS services it uses
- Amazon Bedrock
- AWS Data Exports (CUR 2.0)
- Amazon Athena
- AWS Glue Data Catalog
- AWS Lambda
- Amazon EventBridge Scheduler
- Amazon S3
- AWS KMS
- Amazon SNS
Your data
Runs serverless in your account and uses token counts and request metadata, not prompt text. Nothing is sent to CloudNerve.
Scope, stated plainly
What it does not do
- Change your models, prompts or traffic
- Invoice your customers
- Cover AI providers outside Amazon Bedrock
- Split cost per request without Bedrock invocation logging turned on
Questions
Common questions
- Do we need code changes?
- Only to stamp tenant and feature on each request, with the helper. Everything else reads data AWS already produces.
- How accurate is it?
- The allocated total matches the Bedrock line in your bill. The report shows the rounding residual and the unattributed share instead of hiding them.
- Where does it run?
- In your AWS account, in the Region you choose, as a scheduled Lambda function with Athena queries.
Works well with
Pair it with an engagement
See TokenLedger on your own AWS data.
Book a free Intro Call with a CloudNerve architect, or ask us anything about how it would fit your accounts.