OpenRouter researchers analyzed 380 trillion tokens and found an AI premium of 0.641 percentage

points per week Researchers collected real-world usage data from more than 400 models running on OpenRouter, a sample covering roughly two percent of global monthly AI consumption, and distilled it into a single factor capturing growth in tokens, spend, and user counts. That factor let them calculate an AI beta for every public company — a measure of how much a stock's moves track actual model demand rather than media hype.

Firms with high AI beta went on to deliver higher returns in the following period, and a long-short strategy that bought the most sensitive names and shorted the least sensitive earned an average 0.641 percentage points per week. The effect was most pronounced around closed models, paying users, experienced users, and long prompts — in other words, where usage signals real workflow integration. Casual usage and open-weight models produced no comparable premium, suggesting the market rewards deep integration, not just surface-level experimentation.

The phenomenon reaches well beyond the tech sector: communication-heavy, teaching, and human-interaction occupations received relative overvaluation, while analytical, scientific, and operational roles drew extra caution.

The researchers conclude that tokens are becoming a new economic indicator: actual model consumption lets investors flag the companies and occupations they see as winners of the AI economy. The article remains a pre-print on arXiv and the correlation does not prove causation, but the practical takeaway is clear — the market started pricing AI before most companies learned to measure its impact on them.

The most interesting data point missing from the report: the authors did not publish the list of tickers with the highest and lowest AI beta