AWS, Google users hit with surprise AI bills
- The Register reported on May 18 that AWS and Google Cloud customers were hit with unexpected AI bills after inference and training usage spiked. - One AWS customer told The Register an Amazon Bedrock experiment produced a $30,141.33 invoice after credits ran out and alerts failed. - Google Cloud budgets and alerts documentation says spending is not capped automatically; AWS documents Cost Anomaly Detection limits for supported services.
The Register reported on May 18 that some AWS and Google Cloud customers were left with AI-related bills in the tens of thousands of dollars after usage ran far beyond what they expected. Two recent cases described different failure modes: compromised or exposed Google Cloud API keys that were used to run expensive model calls, and Amazon Bedrock usage that a customer said slipped past AWS cost alerts. In both cases, the customers told The Register they struggled to identify what was driving the charges and how to stop them. Google and AWS both offer billing and alerting tools, but their documentation shows those tools do not act as hard spending caps. ### How did Google Cloud users end up with AI bills they say they did not authorize? The Register reported on May 12 that several Google Cloud customers said their API keys were compromised and then used to run inference on high-cost image and video models. The outlet said users who had long paid modest monthly amounts for products such as Google Maps suddenly saw charges in the thousands or tens of thousands of dollars. (theregister.com) Google told The Register that the incidents were not specific to Google and usually stemmed from compromised credentials, including API keys leaked in public code repositories. The company said attackers actively scrape public repositories and urged customers to use stronger security practices, including multi-factor authentication, auditing keys and keeping credentials out of public code. (theregister.com) Rod Danan, chief executive of Prentus, told The Register his company’s Google bill had usually stayed below $50 a month before he received a March alert that he was being charged $3,000. Danan said tracing the source of the charges was difficult once they appeared. After publication, The Register said Google reimbursed Danan and another named source, Isuru Fonseka. (theregister.com) ### What was different about the AWS case? The Register reported on May 14 that one AWS customer ran up $30,141.33 in Amazon Bedrock charges in April 2026 while using Anthropic’s Claude through AWS. The customer had set AWS Cost Anomaly Detection thresholds 33 days earlier at “Absolute ≥ $100 AND Relative ≥ 40%,” according to the report. (theregister.com) The same report said AWS Activate credits covered the first $8,026.54 of charges, after which billed inference costs began accumulating along with another $675.07 in AWS infrastructure charges. The customer told The Register there was no notice when the credits were exhausted and that the first warning was a large invoice. (theregister.com) AWS documentation says Cost Anomaly Detection tracks unusual spend patterns in deployed AWS services and sends alerts when thresholds are met. AWS also says the default AWS-managed alert is set at $100 and 40% of expected spend when Cost Explorer is enabled. ### Why did the AWS alert not catch the Bedrock spending? The Register said the customer expected AWS Cost Anomaly Detection to watch “AWS Services” automatically, but Marketplace charges were not supported in the way the customer assumed. (theregister.com) The report said Anthropic Claude model charges in that case were billed through AWS Marketplace, which meant the anomaly tool did not trigger an alert. (docs.aws.amazon.com) AWS documentation says customers can create monitors and alert subscriptions for anomalous spend, but the service has defined coverage and limitations. The Register cited AWS documentation warning that AWS Marketplace is an unsupported service for that anomaly workflow. Corey Quinn of the Duckbill Group told The Register that Bedrock model spend being treated as Marketplace was not intuitive unless a customer was already deeply familiar with AWS billing. (theregister.com) Quinn said he uses Anthropic directly in part to get real-time billing, alerts and cutoffs. ### Don’t cloud budgets stop this automatically? (docs.aws.amazon.com) Google Cloud documentation says no. Its budgets and alerts page warns that setting a budget “does not automatically cap Google Cloud or Google Maps Platform usage or spending” and that budgets are designed to trigger alerts as costs trend upward. Google also documents programmatic notifications for customers who want to automate cost-control responses. (theregister.com) Google’s billing controls page says disabling billing on a project stops services and eventually deletes resources, while more selective controls require separate notification-driven automation. That means a budget alert by itself is not the same thing as a hard limit. ### What should users watch next? Google Cloud’s current billing documentation points customers with disputed Gemini charges to Cloud Billing Support and to project-level billing controls. (docs.cloud.google.com) AWS’s cost-management documentation points customers to Cost Explorer, Cost Anomaly Detection and alert subscriptions, with coverage defined by each service’s documented limits. The next concrete step for affected customers is likely to be in billing consoles and support queues, not in model dashboards. (docs.cloud.google.com) Google’s budgets API and billing tools remain available for tighter notification setups, and AWS continues to document anomaly monitors, thresholds and supported services for customers reviewing Bedrock-related spend. (docs.cloud.google.com 1) (docs.cloud.google.com 2)