Amazon’s $1.8m Claude blunder shows AI’s runaway costs

Amazon's internal metrics revealed a series of AI cost overruns, most notably an $1.8m Claude deployment that ran 860% over budget for five months before being detected — and ultimately failed. The root causes include misconfigured retry loops, using frontier models by default without guardrails, and per-token billing that accumulates silently rather than crashing like buggy code. AWS ironically sells the tools to prevent exactly this: batch inference, prompt caching, cheaper model tiers like Haiku. Amazon has since scrapped an internal AI usage leaderboard and is building automated spend caps. The broader lesson is a governance one — runaway AI jobs that hide for months are a manageable rounding error for Amazon but a potential crisis for smaller enterprises.