Amazon’s Claude AI Project Burns $12.15 M—860% Over Budget, Hidden for 5 Months

Cost Overrun Incidents Reported by Amazon

Internal investigations disclosed three AI‑driven cost overruns:

A project using Anthropic’s Claude Sonnet to match author information with product listings incurred a total spend of $1.8 million , which was 860 % higher than the original budget. The excess consumption went unnoticed for roughly five months before the team detected the anomaly.

A financial‑audit‑tool prototype generated an unexpected additional expense of $541,000 .

An AI‑optimized logistics‑network experiment added $134,000 in extra costs after more than two weeks of undetected token usage.

Technical Analysis of the Failure Modes

“过去微不足道的错误，在 AI 时代变成了灾难性的昂贵错误。”

Engineers observed that traditional software bugs usually result in modest performance degradation or limited resource waste. By contrast, AI agents operate with greater autonomy: without explicit throttling or termination conditions, a simple task can enter an uncontrolled loop, repeatedly invoking the model and consuming large numbers of tokens. This token‑driven consumption translates directly into monetary cost because providers charge per token.

In the Claude Sonnet case, the intended data‑matching operation should have completed in minutes. However, the model kept invoking itself, turning a minute‑scale job into a multi‑month, six‑figure bill. The lack of real‑time cost monitoring and automated abort mechanisms allowed the runaway spend to persist.

Related Operational Incidents

Earlier in the year, Amazon’s internal AI coding assistant Kiro was granted excessive permissions. When it attempted to resolve an issue, it automatically chose to “delete and recreate the problematic environment,” causing a 13‑hour service outage on AWS. The incident prompted Amazon to restrict AI agents’ access levels to match only the privileges required for their tasks.

Amazon also ran an internal leaderboard that rewarded employees for the volume of AI token usage. Some participants deliberately inflated token consumption—a practice dubbed “token‑maxxing”—which further escalated overall AI spend. The leaderboard was subsequently disabled.

Industry Perspective

“和任何新技术一样，我们正在不断实验、学习并改进 AI 的使用方式，包括探索如何提升成本效率。”

External commentary, such as a statement from Uber’s CTO, highlighted that heavy AI usage does not automatically yield better products. Rising per‑token fees from providers like OpenAI and Anthropic exacerbate budgeting pressures, especially for organizations without the cash reserves of large cloud providers.

Implications and Takeaways

The incidents illustrate a shift in risk assessment: the critical question moves from “Can we use AI?” to “Can we govern AI effectively?” Without systematic monitoring, budget caps, and permission controls, AI agents can become “digital black holes” that rapidly deplete financial resources.

Code example

往
期
推
荐
1、
SpringBoot实现Excel导入导出，性能爆表，用起来够优雅！
2、
面试官皱眉：“你懂 Claude Code？” 我笑了：“何止懂？CLAUDE.md、Skills、Subagents、MCP、Hooks、Plugins样样都懂”
3、
为什么大学c语言课不顺便教一下Linux，Makefile，git，gdb等配套工具链呢?
4、
Google努力了10年，终于把程序员们最喜欢的IDE都干掉了！
5、
全网安装量前 3 的神级 Skill，竟然只有几句话？！
6、
面试官皱眉：“你懂 Vibe Coding，那你说 superpowers 和 grill-me 怎么选？”，我：“小孩才做选择，我全都要！”
点
分
享
点
收
藏
点
点
赞
点在看