Three economists took real-world AI usage data and pitted it against US stock market returns. The study covers January 2024 to April 2026, using roughly 380 trillion tokens from OpenRouter — a model aggregator that lets developers access more than 400 AI models through a single API.
They combined weekly token volume, spending, and active user growth into something called an "AI usage index," then looked at how each US stock had responded over the prior 13 weeks. Stocks that tended to outperform the market when AI usage accelerated, and underperform when it slowed, were grouped as "high exposure."
The strategy: buy that high-exposure group every week, short the low-exposure group. The backtest found an average weekly return gap of 0.641 percentage points between the two. Even after controlling for company size, valuation, earnings, and price momentum, the gap stayed at about 0.56 percentage points.
But before you add "AI token usage" to your stock screen, consider the caveats. The sample is only 28 months, OpenRouter covers a small slice of global AI activity, and its users skew developer-heavy. Long-term effectiveness is unproven.
The research paper has the details.