Bedrock and a hard place: Claude adventure leaves AWS user staring down a $30 K invoice

Regulatory action → What it requires → Compliance timeline

AWS Cost Anomaly Detection (CAD) limitation – The service’s documentation states that AWS Marketplace charges are not supported for anomaly alerts. This requirement has been in place since the CAD feature launch in November 2023 and remains unchanged as of the April 2026 billing cycle.

AWS Budgets coverage – AWS Budgets does include Marketplace spend, but only when a budget is explicitly configured for the Amazon Bedrock product family. The budget must be created with a service‑specific filter; generic "All services" budgets will not capture Marketplace line items.

Customer‑facing notification rule – AWS Activate credits are applied automatically to eligible services, but the platform does not emit a depletion warning when the credit balance reaches zero. Customers must query the AWS Billing console or enable the Credit Balance widget in the Cost Explorer dashboard.

---

What happened

A Register reader activated an AWS Activate credit package worth CAD $8,026.54 and then launched Anthropic Claude Opus through Amazon Bedrock. The model is billed via the AWS Marketplace, which means each inference request generates a Marketplace line item rather than a native Bedrock charge.

The user had configured CAD with the following threshold:

• Absolute ≥ $100
• Relative ≥ 40 %
• Scope: "AWS Services" (the default setting that tracks all native AWS services)

Because Marketplace charges are excluded from CAD, the system never raised an alert. The first $8 K of usage was covered by the Activate credits, so no invoice appeared. When the credits were exhausted, the Marketplace billing resumed silently, accumulating $30,141.33 in model‑inference fees plus $675.07 in supporting infrastructure costs.

Why it matters

1. Hidden cost channel – AI models on Bedrock are not billed as ordinary EC2 or Lambda usage; they flow through the Marketplace pipeline. If a customer relies solely on CAD, they will miss any surge in Marketplace spend.
2. Credit‑masking effect – Activate credits hide the underlying consumption pattern. Without a dedicated credit‑balance monitor, the transition from “free” to “charged” can go unnoticed.
3. Alert configuration gap – The default "All services" scope in CAD does not capture Marketplace items. Organizations must create a custom budget that explicitly includes the Amazon Bedrock – Marketplace line item.

Steps to avoid a repeat

Action	Tool	Configuration detail
Enable explicit budgeting	AWS Budgets	Create a budget with Service = Amazon Bedrock and Marketplace = true. Set the threshold to the maximum acceptable monthly spend (e.g., $1,000).
Add credit‑balance monitoring	Cost Explorer dashboard	Pin the Credit Balance widget and set a CloudWatch alarm for < $100 remaining.
Supplement CAD with CloudWatch metrics	CloudWatch	Publish a custom metric that aggregates Marketplace line‑item charges (use the aws.billing namespace). Attach an alarm with the same thresholds used in CAD.
Review Marketplace eligibility	AWS Documentation	Verify that any third‑party AI model you plan to use is billed through Marketplace; if not, CAD will work as expected.
Test with a sandbox account	AWS Free Tier	Run a limited number of inferences and confirm that the charges appear in the Billing → Cost Explorer view under Marketplace.

What AWS says

"AWS offers multiple tools to help customers manage spend, including AWS Budgets, which covers Amazon Bedrock spend on AWS Marketplace and other services. As noted in our documentation, AWS Marketplace charges are not currently supported by Cost Anomaly Detection. Customers with questions should reach out to AWS Support." – AWS spokesperson, May 2026

Takeaway for compliance officers

When drafting internal AI‑usage policies, reference the AWS Cost Management documentation (last updated 12 April 2026) and explicitly state that Cost Anomaly Detection does not monitor Marketplace charges. Require the creation of a dedicated AWS Budget for any Bedrock‑based model and mandate weekly review of the credit‑balance widget. By aligning internal controls with the documented limitations, you can prevent surprise invoices and keep AI experimentation within approved financial limits.