亞馬遜內部工程師在員工會議上揭露，公司在導入AI技術時出現多起嚴重的成本超支事件。其中最突出的案例是使用Anthropic的Claude Sonnet模型進行電商網站作者資料比對，花費高達180萬美元，超出預算860%，且耗時五個月才被發現。此外還有約54.1萬美元的財務稽核工具開發超支，以及13.4萬美元的物流配送AI部署意外支出。資深工程師指出，這些並非個別事件，而是反映了AI部署中缺乏有效支出管控的系統性問題。
亞馬遜工程師表示，傳統程式設計中「微不足道的廉價」編碼錯誤，在AI模型部署中卻可能造成「災難性的昂貴」後果。公司正在開發自動化的成本防護機制，以防止未來再度發生類似的失控支出。亞馬遜官方回應稱，這些只是少數團隊在學習過程中的個別案例，並不代表公司整體AI運用的常態，並強調公司正持續改進AI的使用效率。（關鍵數字：300,000, 180）
在更宏觀的背景下，亞馬遜今年預計資本支出將達到2000億至2200億美元，其中絕大部分投入AI與資料中心基礎設施，同時正進行大規模裁員以控制成本。整個產業也面臨AI部署成本上升的挑戰，主要AI供應商已從固定訂閱制轉向按代幣（token）計費模式。亞馬遜此前還因員工在內部開發平台上進行「代幣灌水」競賽而被迫關閉相關排行榜，凸顯了AI成本管理的複雜性。
Amazon engineers have disclosed multiple instances of severe cost overruns stemming from AI deployments lacking adequate spending controls. The most notable case involved spending $1.8 million using Anthropic's Claude Sonnet model to match author details with ecommerce listings—an 860 percent budget overrun that went undetected for five months. Additional incidents included $541,000 in unexpected costs for financial auditing tools and $134,000 in accidental spending on logistics AI, with the latter taking over two weeks to discover.
Senior engineers warned colleagues that coding mistakes considered trivially cheap in traditional systems were proving catastrophically expensive when teams deployed AI models. The company is now developing automated guardrails to prevent future uncontrolled spending. Amazon responded that these were small, isolated examples of teams learning from one another and did not reflect how AI is broadly used across the company, which employs roughly 300,000 corporate staff and generates about $180 billion in quarterly revenue.
The incidents occur against a backdrop of Amazon's massive AI investment push, with expected capital expenditure of $200–220 billion this year, predominantly directed at AI and data center infrastructure, even as the company undertakes sweeping layoffs to manage costs. The broader industry is also contending with rising AI deployment expenses as providers like Anthropic and OpenAI shift from flat subscriptions to token-based billing. Amazon previously had to shut down an internal leaderboard ranking employees' AI usage after it incentivized token inflation, underscoring the persistent challenge of governing AI costs at scale. (Key numbers: 2000, 2200)