For a while, the volume of "token" consumption was often regarded as a key indicator of a tech company's AI ACTivity, technical Investment scale, and even its innovation capability. However, this consensus is quietly unraveling.
According to reports from media outlets like Business Insider, Amazon notified employees this week that its internal "Kirorank" leaderboard has been discontinued. Created by internal developers, the leaderboard was designed to track employees' Token Consumption with the initial goal of promoting the APPlication of AI in business scenarios. However, the initiative backfired. In pursuit of higher rankings and better performance reviews, some employees began treating AI Agents as tools to mindlessly "burn tokens" without regard for cost, leading to a significant waste of the company's computing resources.
In response, Amazon Senior Vice President Dave Treadwell advised employees: "Please don't use AI just for the sake of using AI." In his view, while Amazon can decide how to utilize and track AI tools—and does monitor Token consumption to measure computing costs—the company does not encouRAGe "Token Maxxing." This is a practice that somewhat equates Token consumption with AI innovation capability, leading to a blind pursuit of maximizing usage.
Beyond the act of "burning Tokens" itself, what truly makes tech giants anxious is the lack of visible business and financial returns from this consumption.
Recent research from the enterprise management platform Jellyfish highlights this inefficiency. It shows that high-frequency users of Claude Code consume an average of 225 million Tokens per week, whereas ordinary software engineers consume only 32 million. While the top 10% of claude Code users consume roughly ten times the Tokens of average developers, their ouTPUt is only double.
Meta faced a similar situation with an internal leaderboard called "Claudeonomics." Data revealed that Meta employees consumed 60.2 trillion Tokens in just 30 days. Calculated at anthropic's public API pricing, this bill amounts to approximately $900 million. An insider bluntly stated that the output from those at the top of the leaderboard was mostly "disposable garbage." Consequently, Meta also shut down its leaderboard.
Uber's Chief Operating Officer, Andrew Macdonald, recently noted that the ride-hailing giant has not seen a corresponding increase in Productivity or other benefits to match its rising AI investments. Previously, the company's CTO, Praveen Neppalli Naga, mentioned that as of April, Uber had already exhausted its entire annual budget for Claude Code.
OpenAI CEO Sam Altman also publicly questioned this trend during a recent online event: "I hear a lot of good news from companies, but there's also negative news, like 'Our spending keeps increASIng, and people feel highly productive... but where is the revenue? Where is the actual productivity gain?'"
On the other hand, "Token Maxxing" does have its supporters. GARRy Tan, CEO of the famous US startup incubator Y Combinator, has embraced the term, stating, "We've been Token Maxxing longer than most people."
Jellyfish suggests in its report that business leaders should focus on outcome-oriented metrics, such as the cost per code merge request, rather than simply tallying total Token consumption. The firm emphasizes that while heavy AI use does improve efficiency, the benefits have diminishing returns. The optimal path forward is to promote AI programming applications widely, encourAGIng more engineers to operate within a moderate and reasonable usage range—avoiding both underutilization of AI and excessive resource consumption.
Comments & Questions (0)
No comments yet
Be the first to comment!