亚马逊内部曝光多起AI项目成本失控事件。[1]其中最严重的一次，员工使用Anthropic的Claude Sonnet模型完成数据匹配任务，产生180万美元账单，超预算860%，直到5个月后才被发现。[1]此外亚马逊还发现另外两起AI超支案例，分别产生54.1万美元和13.4万美元的额外费用，其中后者在超支两周后暴露。[1]
亚马逊工程师对此表示，"过去微不足道的错误，在AI时代变成了灾难性的昂贵错误"。[1]这并非亚马逊首次因AI工具遭遇运营困扰——AWS曾因AI编程助手Kiro的错误操作导致13小时服务中断，亚马逊推出的员工AI使用排行榜也因出现"tokenmaxxing"现象而被关闭。[1]
An internal Amazon incident has exposed significant cost control failures in artificial intelligence projects within the company.[1] An employee utilized Anthropic's Claude Sonnet model to complete a data matching task, which resulted in a bill of $1.8 million—exceeding the original budget by 860 percent.[1] The problem went undetected for approximately five months before being discovered.[1]
This was not an isolated occurrence. Amazon identified two additional cases of AI cost overruns during the same period.[1] A financial auditing tools project incurred an unexpected $541,000 in expenses, while a logistics network optimization initiative generated an additional $134,000 in unplanned costs that were exposed after more than two weeks.[1] An engineer at the company characterized the scale of the problem, noting that "trivial mistakes that were inconsequential in the past have become catastrophically expensive errors in the AI era."[1]
The incident underscores broader challenges Amazon faces in managing AI tool deployment.[1] AWS previously experienced a 13-hour service outage caused by errors from its AI programming assistant Kiro,[1] and the company discontinued an employee AI usage leaderboard after observing "tokenmaxxing" behavior among staff.[1]
