AI writing code burned 12.15 million yuan, 860% over budget: Amazon’s internal accident sounded the alarm for corporate AI costs. Recently, the US media exposed a major accident involving out-of-control AI costs within Amazon.

AI writing code burned 12.15 million yuan, 860% over budget: Amazon’s internal accident sounded the alarm for corporate AI costs. Recently, the US media exposed a major accident involving out-of-control AI costs within Amazon. An engineer used the Claude Sonnet model of Anthropic to write a program with the goal of completing data matching between author information and product lists on the e-commerce platform - a routine data processing task. However, it ended up generating a bill of US$1.8 million (approximately RMB 12.15 million), and the actual expenditure exceeded the original budget by 860%. What is even more surprising is that the project was not successfully implemented in the end, and the high computing power consumption continued to run for up to 5 months. The company did not notice the abnormality until the financial reconciliation. This means that an unattended AI program has been "burning money" day and night for nearly half a year. Code defects were superimposed and monitoring was lacking, and tiny loopholes swallowed up huge budgets. A review showed that the root cause of the accident was not complicated. The programs generated by AI lack a limit on the frequency of calls. The model continues to make requests in a loop, and cloud computing power and large model call costs continue to accumulate. At the same time, the team lacked a real-time cost warning mechanism, and minor code defects went undiscovered for a long time, continuing to eat up the budget. At an internal Amazon meeting, executives defined the accident as a "typical and costly case" and reported that many similar hidden dangers had occurred internally, with one project unexpectedly incurring additional expenses of $540,000. After a series of pitfalls, Amazon began to urgently build a risk control system and planned to launch an automated protection mechanism. Once AI-related expenditures reach a threshold, the system will automatically limit the flow and interrupt tasks to avoid unlimited surges in costs. At the same time, the company has strengthened its regulations and requires that all AI-generated codes must undergo complete manual review to prevent them from being put online directly without verification. Errors in traditional codes cost time, and out-of-control AI programs cost real money. This incident exposed a risk that is ignored by most companies: errors in traditional codes mostly consume human time, while out-of-control AI programs will continue to generate token fees. In the absence of budget monitoring, it can easily lead to huge losses. Currently, many companies simply expect AI to reduce costs and increase efficiency, but ignore the new risk exposures brought by the computing power billing model. From "AI helping to write code" to "AI secretly burning money", Amazon's lesson is not an isolated case, but a new stage that the entire industry is entering - when AI changes from an auxiliary tool to an autonomous executor, the granularity of cost control must also sink from the "project level" to "each call". Before enjoying the benefits of AI efficiency improvement, installing "brakes" on AI may be the most important lesson for enterprises at the moment. via AI News (author: A