Lightspeed Partner: Blindly stacking tokens is like frantically grinding APM in games, AI's real bottleneck lies in verification capabilities

TechFlow news, July 30, according to Fortune, Lightspeed Venture Partners partner Nnamdi Iregbulem wrote that the AI field is repeating the mistakes of the gaming industry—using "Tokenmaxxing" (i.e., the belief that consuming more Tokens is necessarily better) as the core metric for measuring AI efficiency, just as early gamers regarded Actions Per Minute (APM) as a proxy indicator for skill level, which is essentially a vanity metric.

Research data shows that the Token consumption of AI Agents is 4 times that of ordinary conversations, with multi-agent systems reaching up to 15 times; while a 2025 analysis on AI-assisted software development found that the time engineers spent reviewing and modifying AI-generated code (accounting for 9% of programming time) was more than double the time spent waiting for AI to generate code (4%). The author cited an economics paper by researchers from MIT, the University of Washington, and UCLA pointing out that "the bottleneck constraining growth is no longer intelligence itself, but human verification bandwidth."

The article suggests that enterprises should prioritize building scalable verification mechanisms when deploying AI, rather than blindly pursuing Token consumption, and calls on the industry to establish measurement standards similar to "effective APM" in the gaming industry as soon as possible, to distinguish truly valuable Agent output from invalid compute consumption.