Overview
- Amazon published an engineering memo that was posted publicly on Tuesday directing engineers to measure AI work by long-term customer and business outcomes rather than raw token usage.
- The memo explicitly warns against 'tokenmaxxing,' which it defines as using AI excessively to inflate usage metrics without improving real results.
- Engineering leadership called metrics like token counts, AI-adoption rates, and lines-of-code 'deeply flawed' when turned into targets because they can raise costs and add needless complexity.
- Managers are told to use AI themselves on production work, investigate teams in the bottom and top 10 percent of AI adoption, and judge success by delivery impact and feature velocity rather than token volumes.
- The guidance follows Amazon's May removal of an internal token leaderboard and reflects wider 2026 industry 'token shock' concerns as firms balance heavy AI use, rising model bills, and the role of AWS as both vendor and large internal consumer.