Financial markets are rewarding companies that are well positioned to benefit from widespread adoption of artificial intelligence (AI) with higher returns, according to a new Yale-led analysis of 380 trillion AI tokens, one of the largest datasets of real-world AI consumption ever studied.
The difference in stock returns between companies with the most and least AI exposure is substantial — about 0.64% per week, a financial benefit the researchers call the “AI Premium.”
And it’s not just tech companies that are seeing their stock prices rise as AI adoption increases, but also consumer-facing sectors and capital-intensive industries, the study finds. This suggests that investors believe that many types of businesses stand to take advantage of AI-driven productivity, said Yale economist Aleh Tsyvinski, the study’s co-author.
For the study, researchers analyzed 380 trillion AI tokens — the basic units of data that AI models read or generate while performing tasks — from a licensed, proprietary dataset belonging to OpenRouter, a digital platform that routes requests to more than 400 AI models, including GPT, Claude, and Deepseek, from January 2024 through April 2026. The dataset represents 2% of monthly global AI usage by millions of anonymous users during that period.
The researchers leveraged the dataset’s unprecedented size, scope, and granularity to better understand how the rapidly emerging technology is affecting businesses, markets, and workers in the United States and across the globe.
“Previous studies on the economic effects of AI have mostly relied on surveys and estimates, but our work is based on unprecedented amounts of real-world data, providing a much more granular view of how AI consumption is affecting the economy,” said Tsyvinski, the Arthur M. Okun Professor of Economics in Yale’s Faculty of Arts and Sciences. “Our analysis reveals that AI’s emergence is not just a tech story, it’s a much broader story that affects firms and workers in all parts of the economy. We show that it is already deeply influencing financial markets.”
The study, a National Bureau of Economics Research working paper, is coauthored by Nicola Borri, associate professor of finance at Luiss University in Italy, and Yukun Lui, associate professor of finance at the University of Rochester. Yale’s Data-Intensive Social Science Center supported the researchers in licensing and managing the OpenRouter dataset.
For their analysis, the researchers created what they call the “AI Factor,” a weekly measure of growth in worldwide AI consumption. Then they looked at which companies’ stock prices tend to rise as AI usage increases. The analysis showed that those companies generally earned higher future stock returns than companies with lower exposure to AI consumption, which is the “AI Premium.” Beyond the tech sector, they found that the retail and consumer durables industries benefit from the exposure as well as companies with large physical assets, such as manufacturers.
(In this context, “exposure” is a measure of market expectations. It means that investors expect companies to benefit from increased productivity caused by AI adoption, not that those companies are directly benefitting from AI today.)
The study found that the beneficiaries of the AI “premium” are most concentrated in the United States, Europe, and other developed markets where investors and companies are deeply connected with leading-edge AI development, infrastructure, and construction. It is less pronounced in China and other emerging markets (defined as countries whose economies are transitioning from developing to developed status).
“The equity markets now reward the proximity of companies to the most frontier models in the United State and Europe,” Lui said.
The study also showed that investors value intensive AI consumption that involves proprietary, advanced models, experienced users and paying customers, and longer, sophisticated prompts over people casually experimenting with free or open-source models.
“Despite the widespread and fast adoption of AI tools by everyday users, the AI premium is mostly determined by the exposure to the frontier AI consumption by sophisticated and professional users,” Borri said.
The researchers combined stock market data, government labor data, and the data on AI consumption to gain insights into how AI’s continuing emergence will affect workers across fields and industries.
They found that occupations involving non-routine tasks — such as persuasion, teaching, communication, and interaction with people — have more positive exposure to AI consumption compared to occupations that involve routine work, such as jobs in health care and those that involve scientific analysis and other kinds of analytical work. This means that investors believe that AI will create greater economic opportunity in jobs that feature communication and coordination over those that involve scientific analysis, the researchers noted.
“Our analysis suggests that interactive job skills will be rewarded while analytical skills get penalized with the rise of AI,” Tsyvinski said. “Occupations in science are among the most negatively affected, which may seem surprising. There is a distinction to be made here between more routine lab work and the kind of cutting-edge scientific research that occurs at major research universities. It is the routine work that our analysis indicates will be most affected by AI.”
The study also showed an increase in the use of what is known agentic AI — systems that don’t just answer questions but can operate autonomously to complete a task — over the period covered by the OpenRouter dataset. In 2024, use of agentic AI models accounted for only a small share of AI consumption. But by 2026, more than half of AI tokens involved agentic systems, according to the study. Additionally, the researchers found early evidence that the exposure of companies to the rise of agentic economy is starting to play an important role in determining their equity market valuations.
Source: Original Article
