Yale Analysis of 380 Trillion AI Tokens Quantifies 'AI Premium' in Stock Returns and Agentic AI Surge

Introduce the scale of the measurement problem. Until now, measuring the economic impact of artificial intelligence has relied heavily on corporate surveys, executive sentiment, and broad estimates. A new working paper from researchers at Yale University and the National Bureau of Economic Research fundamentally changes this approach by analyzing actual, realized AI consumption at an unprecedented scale. By moving away from subjective self-reporting, economists are finally able to quantify exactly how the integration of advanced machine learning models is altering global financial markets and corporate valuations.[1][2]

The research team, led by economists Nicola Borri, Aleh Tsyvinski, and Yukun Liu, analyzed 380 trillion AI tokens generated between January 2024 and April 2026. These tokens—the fundamental units of data that AI models process and generate—were sourced from OpenRouter, a platform that routes requests to more than 400 different large language models. This dataset captures approximately two percent of global monthly AI token consumption, providing a highly granular view of how businesses and individuals are actually utilizing the technology in real-world applications.[1][2][3]

By tracking this massive volume of usage across millions of anonymized accounts, the researchers constructed a high-frequency AI Factor based on the aggregate growth in tokens processed, dollars spent on API calls, and the number of active users. They then mapped this granular consumption data against the broader stock market to observe how equity valuations responded to AI exposure over time. The primary claim emerging from this evidence pack is the existence of a substantial and measurable AI Premium in the cross-section of equity returns, proving that markets are actively pricing in this technological shift.[1][4]

The evidence for this premium is exceptionally robust, grounded in actual usage rather than executive sentiment. The study demonstrates that firms whose stock returns covary more positively with the AI factor—designated mathematically as high AI beta firms—earn significantly higher subsequent returns. Specifically, a value-weighted long-short trading strategy based on this exposure yields a premium of 64.1 basis points per week. This indicates that financial markets are actively and aggressively rewarding companies that are positioned to benefit from AI adoption, translating technological integration directly into sustained shareholder value.[1]

However, the data reveals that not all artificial intelligence usage is treated equally by the market. The financial premium is highly concentrated on the intensive, frontier-oriented margins of AI consumption. Investors are specifically rewarding corporate exposure to closed-source, proprietary models, seasoned paying users, and complex, long-form prompts that require significant computational resources. In contrast, casual experimentation or the use of free, open-weight models does not drive the same equity premium, suggesting that markets strictly value deep, professional integration over superficial or exploratory adoption.[1][2][4]

The sectoral distribution of this AI Premium also challenges conventional narratives that restrict the technology's benefits to Silicon Valley. While technology firms naturally benefit, the evidence shows the premium extending deeply into consumer-facing sectors like retail and consumer durables, as well as capital-heavy industries such as traditional manufacturing. This broad distribution indicates that investors expect AI-driven productivity gains to materialize across the wider physical economy, fundamentally altering operational efficiencies and cost structures far beyond the immediate software and cloud computing ecosystems.[2]

Geographically, the evidence points to a stark divide in how these benefits are being realized. The beneficiaries of the AI Premium are heavily concentrated in the United States, Europe, and other developed markets that are closely tied to frontier AI development and the massive data center infrastructure required to support it. The premium is notably absent in emerging markets, including China, suggesting that global equity markets are currently rewarding strict proximity to the most advanced models and the localized infrastructure required to run them at scale.[1][2]

Beyond corporate valuations, the Yale study provides compelling empirical evidence regarding artificial intelligence's immediate impact on the labor market. By combining the token consumption data with government labor statistics and occupational classifications, the researchers mapped AI exposure to specific human skills. The findings upend some early assumptions about which jobs are most exposed to automation, providing a data-driven look at the future of work and the specific capabilities that will retain their value in an increasingly automated economy.[2][4]

The data shows that occupations heavily reliant on non-routine interactive skills—such as persuasion, teaching, negotiation, and complex human communication—have a strongly positive exposure to AI consumption. Conversely, jobs centered on analytical, scientific, and operations-control skills face negative exposure. For example, routine laboratory work, standard data analysis, and basic coding are highly exposed to AI substitution, while roles requiring nuanced human interaction and emotional intelligence appear to be complemented and enhanced by the technology, rather than replaced by it.[1][2]

A critical technological shift documented in the evidence pack is the rapid and accelerating rise of the agentic economy. Agentic AI refers to advanced systems that do not merely generate text in response to a prompt, but autonomously execute multi-step tasks, call external software tools, and manage complex workflows without constant human supervision. In early 2024, these agentic models accounted for only a very small fraction of total AI consumption, as the industry was still focused on basic conversational interfaces.[2]

By the spring of 2026, the consumption data shows a dramatic transformation in how the technology is being deployed. Tool-call and cache-read shares rose to account for roughly two-fifths to one-half of all processed tokens on the OpenRouter platform. This indicates that more than half of current AI usage involves agentic systems actively performing tasks, retrieving data, and interacting with other software, rather than passively answering user queries. The era of the simple chatbot has rapidly given way to autonomous digital workers.[3][4]

The researchers found early but compelling evidence that exposure to this specific agentic consumption carries its own positive premium in equity valuations. The point estimates for this agentic premium range from 0.3 to 0.5 percent per week for individual agentic factors. Because widespread agentic token consumption is a very recent phenomenon, the researchers note that these estimates remain somewhat imprecise, but they clearly signal that financial markets are beginning to differentiate between basic generative AI and highly autonomous, action-oriented systems.[3]

While the dataset is unprecedented in its size and scope, the researchers are transparent about the uncertainties inherent in measuring such a nascent and rapidly evolving phenomenon. The long-term stability of both the general AI Premium and the newer agentic premium remains an open question for economists. As the technology matures and becomes universally adopted across all sectors, the unique competitive advantage it provides may diminish, potentially compressing the outsized equity returns currently enjoyed by early adopters and heavily exposed firms.[4]

Furthermore, it is crucial to distinguish between market expectations and realized, bottom-line productivity. The exposure measured by the AI betas in this study reflects investor expectations that these companies will benefit from AI adoption in the future. It does not necessarily prove that these firms are already seeing proportional increases in their fundamental earnings, profit margins, or operational efficiency today. The stock market is a forward-looking mechanism, and the current premium represents a collective bet on future transformations rather than a receipt for past performance.[2][4]

Despite these inherent uncertainties, the Yale analysis represents a landmark achievement in the empirical study of artificial intelligence and its economic footprint. By moving away from subjective corporate surveys and anchoring their findings in 380 trillion verifiable data points, the researchers have provided the clearest, most granular picture yet of how AI is reshaping the global economy. The evidence confirms that the AI transition is well underway, actively pricing into financial markets, and fundamentally altering the value of both corporate assets and human labor.[1][2][4]