Amazon Tells Its Engineers to Stop Chasing Tokens: AI Usage Metrics Are "Deeply Flawed"

Amazon has told its software engineers, in writing, that the numbers it spent much of the past year tracking are not a good way to judge them. The guidance, circulated internally in September and published to the company's public website on Tuesday, warns that token consumption, AI tool adoption rates and counts of code changes or reviews remain "deeply flawed" measures of how a team or an organization is actually using AI.

The memo is attributed to Terk Terkowitz, an Amazon vice president and distinguished engineer. It does not forbid the metrics outright. Instead it argues that using them "as a target instead of a measure" can drive negative outcomes, including the practice now known inside the industry as tokenmaxxing, where engineers burn as much model capacity as possible because consumption has been treated as a proxy for productivity.

Amazon has already been through that cycle. Earlier this year the company tracked engineers' use of internal AI tools, and some teams ran leaderboards ranked by token consumption. In May 2026 it shut one such leaderboard down after finding that employees were inflating their usage to climb it. The new guidance states plainly that Amazon has not capped how many tokens an engineer may spend, and that more consumption mainly raises cost while a focus on lines of code creates unnecessary complexity.

What the company wants instead is judgment applied to outcomes. "Using these input metrics, with care and judgement, as you think about your career and your organization's performance is appropriate," the memo says, while concluding that "it is best to measure, and be measured by, long-term outcomes." Marc Brooker, a vice president and distinguished engineer at AWS, told CNBC that the company looks at feature velocity, meaning product launches and successful changes reaching production.

The memo also asks managers to look at the extremes: the bottom 10 percent of teams, to understand why AI is not working for them, and the top 10 percent, to check whether heavier use is really translating into faster delivery. It leaves room for engineers whose work gains nothing from current tools. "If your work does not benefit from current AI tooling, that is a finding, not a failing," it states, adding that "an engineer who can explain where AI does not yet help their work is demonstrating the judgement we are asking for."

The framing throughout is a shift in what a developer is supposed to be. Amazon calls this the era of the "Renaissance Developer," in which human judgment and machine output are combined, and says AI should sit "at the center of software engineering practice." Leaders are expected to use the tools on production work themselves and to explain where those tools fall short. Teams that are not adopting AI at all, the memo concedes, are falling behind.

The wider context is cost. In March, Nvidia chief executive Jensen Huang said on a podcast that he would be "deeply alarmed" if a developer earning $500,000 did not spend at least $250,000 a year on tokens. That framing aged quickly. Uber said in May that it had spent its entire 2026 AI budget by the end of April, with its chief operating officer noting it was getting harder to justify the spending because the company could not draw a line between token consumption and better customer features. An IBM report cited by Indian media found tokenmaxxing peaked in spring 2026 and has been declining since.

For buyers, the practical lesson is that per-token prices are falling while per-task costs are not, because vendors are moving from flat subscriptions to usage-based pricing and agents now spend tokens on retries, tool calls and failed attempts. Measuring consumption rewards exactly the behavior that inflates the bill. Amazon's own labor record sharpens the stakes: chief executive Andy Jassy has said AI will change the scope and speed of work and could shrink the corporate workforce, the company cut fewer than 1,000 white-collar roles this month, and those cuts follow roughly 30,000 reductions that began in 2025 and ran into January. A company that intends to judge engineers on outcomes rather than activity has to mean it, because the activity metrics are the ones that are easy to instrument.