Amazon discovered it had spent $1.8 million deploying Claude Sonnet to match author details with product listings, only to watch the project fail to ship. That’s the visible wound. The deeper problem: internal reports flagged widespread cost overruns across AI projects, with token consumption becoming a silent budget killer. Employees were casually deploying expensive API calls for non-critical tasks. Amazon’s response was implementing mandatory spending caps and peer reviews, guardrails any sophisticated team should have had from the start.
The real signal here: if Amazon, with world-class infrastructure teams and serious budget discipline, got caught out by AI cost overruns, the implication is clear. Enterprise AI deployments across tech and business are probably hemorrhaging money too. They just haven’t discovered it yet. The $200 billion Amazon committed to AI capex in 2026 only works if the company can actually control what that money buys.
According to the Financial Times, senior engineers at Amazon reported widespread difficulty understanding and tracking AI-related costs. The pattern is familiar: a new tool arrives, teams reach for it without the cost discipline they’d apply to infrastructure decisions, and budgets blow out silently.
The takeaway: AI isn’t expensive because of the models themselves. It’s expensive because teams deploy premium tools without guardrails. Until spending caps and peer reviews become standard practice for AI projects, expect more surprises like Amazon’s.