Amazon Suffers Massive AI Budget Overruns, with One Project Exceeding Budget by 860%

Internal Amazon reports revealed multiple instances of severe AI budget overruns, including a project using Anthropic's Claude Sonnet that went $1.8 million over budget and went unnoticed for five months. Other affected projects included a financial auditing tool and a logistics optimization system, which also incurred hundreds of thousands of dollars in unplanned expenses. This incident highlights the critical financial risks of deploying Large Language Models (LLMs) without proper cost controls, as minor coding errors that cost nothing in traditional systems can trigger catastrophic expenses under token-based pricing. It underscores the growing necessity for FinOps practices and automated guardrails in enterprise AI engineering. The overruns were driven by the shift in AI pricing models from flat subscriptions to token-based billing, where costs scale with the volume of data processed. In response, Amazon engineers are developing automated guardrails to monitor and restrict AI API usage, while some enterprises are shifting to smaller models or open-weight alternatives to manage costs.