Recently, U.S. media exposed a major AI cost overrun incident within Amazon. An engineer used Anthropic's Claude Sonnet model to write a program aimed at matching author information with product lists on an e-commerce platform—a routine data processing task—that ultimately resulted in a bill of $1.8 million (approximately 12.15 million RMB), with actual spending exceeding the original budget by 860%.

What was more surprising is that the project was never successfully implemented, and the high computing power consumption ran continuously for five months. The company only discovered the anomaly during financial reconciliation. This means an unmonitored AI program was "burning money" non-stop for nearly six months.

Code flaws combined with lack of monitoring, small vulnerabilities consuming huge budgets

Analysis showed that the root cause of the incident was not complicated. The AI-generated program lacked a limit on the frequency of calls, and the model continuously made requests, causing cloud computing and large model call costs to accumulate. At the same time, the team lacked a real-time cost alert mechanism, and small code defects went unnoticed for a long time, continuously consuming the budget.

In an internal Amazon meeting, executives defined this incident as a "costly case study" and reported that there were multiple similar risks internally, with some projects unexpectedly incurring an additional $540,000 in expenses. After repeated problems, Amazon began to urgently build a risk control system, planning to launch an automated protection mechanism—once AI-related expenses reach a threshold, the system will automatically limit traffic and interrupt tasks to prevent costs from skyrocketing. At the same time, the company strengthened regulations, requiring all AI-generated code to be fully reviewed by humans before going live, to avoid direct deployment without verification.

Traditional code errors cost time,失控 AI programs cost real money

This incident revealed a risk that most companies have overlooked: traditional code errors mostly consume human time, while an uncontrolled AI program continuously generates token costs. Once budget monitoring is lacking, it can easily lead to huge losses. Many companies today simply expect AI to reduce costs and improve efficiency, but ignore the new risk exposure brought by the computing power billing model.

From "AI helps write code" to "AI secretly burning money," Amazon's lesson is not an isolated case but rather a new stage that the entire industry is entering. When AI transitions from an auxiliary tool to an autonomous executor, the granularity of cost control must shift from the "project level" to the "each call." Before enjoying the benefits of AI, installing an "emergency brake" for AI may be the most important lesson enterprises need to learn right now.