Overview
- Internal reviews disclosed multiple AI projects that exceeded budgets, including a product‑listing effort that ran 860% over plan and other programs that accumulated unplanned costs.
- Staff reports show employees overused internal AI tools, with an internal leaderboard called KiroRank suspended after workers gamed usage metrics and drove up token consumption.
- A deployment based on Anthropic’s Claude Sonnet generated about $1.8 million in unexpected charges and never shipped, and other projects also missed budget targets.
- Amazon has begun rolling out mandatory guardrails for AI work, requiring peer reviews and spending caps to detect and stop runaway token use before costs escalate.
- The moves come as the company pursues roughly $200 billion in 2026 capital spending on data centers, networking and custom AI chips, raising investor focus on cost discipline and return on that investment.