Particle.news

Amazon Tightens Controls After Costly Overruns in Internal AI Projects

Mandatory spending caps, peer reviews, and other safeguards aim to curb runaway token consumption to protect Amazon's 2026 capital plan.

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.