The world's enterprises will spend approximately $47 billion on AI API calls this quarter. By itself, that number tells us almost nothing. Yet it dominates board presentations, investor calls, and vendor pitch decks—because it's the only metric everyone agrees to collect.
Reid Hoffman recently entered the "tokenmaxxing" debate with a calibrated take: token consumption tracks adoption reasonably well, but treating it as a productivity proxy is misleading without broader context. He's right, and the implications are more uncomfortable than most in the industry want to acknowledge. The real problem isn't that we're measuring the wrong thing—it's that we're not measuring the right thing at all.
Token tracking answers a simple question: how much AI are we using? It tells operations teams that the API is live, that employees are prompting the model, that tokens are flowing. This has genuine value for capacity planning and adoption monitoring. A company where token consumption is flat while headcount grows tells a story worth telling. But this metric has a ceiling. Once you know AI is being used, knowing it more precisely adds zero insight into whether it creates value.
The productivity question is harder because it requires counterfactual reasoning. What would this worker have produced without AI assistance? What is the baseline? Most enterprises have no answer. They run quarterly reviews, write performance evaluations, and miss the crucial variable: the same sales rep closed 23% more deals this quarter, but was that the AI assistant, a strong lead flow, seasonal patterns, or improved territory mapping? Attributing outcomes to AI inputs remains methodologically messy in ways that frustrate CFOs and excite academic researchers.
This measurement gap creates a dangerous asymmetry. Leadership sees token bills and assumes value accumulation. The mechanism feels obvious: more AI usage should produce more output. But adoption curves and value curves are not the same shape. Early AI adoption often shows high token consumption with modest productivity gains—the organization is still learning workflows, developing prompts, and building institutional habits. Later adoption can reverse: employees become skilled, use fewer tokens per task, and produce more. A token counter misses this entirely.
The companies genuinely solving for AI productivity measurement are doing uncomfortable work. They're running controlled experiments, assigning matched teams, tracking time-in-task metrics, and accepting that the data won't be clean. Some are embedding measurement directly into workflows—asking knowledge workers to log estimated time saved, then aggregating across roles. Others are linking AI usage patterns to downstream outcomes like deal velocity or code deployment frequency, accepting the correlation as sufficient signal.
None of these approaches scale to the board level in a single dashboard. That's precisely why token counts persist. They are legible, exportable, and defensible. "We used 2.3 billion tokens this quarter" sounds like progress. "We cannot definitively quantify AI's contribution to revenue growth" sounds like a consulting retainer waiting to happen.
Hoffman's caution about context is the right frame, but context alone won't close the gap. The industry needs standardized productivity measurement frameworks that accept messiness as the price of accuracy. Until then, the $47 billion quarterly spend will continue to float in a data vacuum—impressive to count, impossible to justify.