📈 economics

AI Premium

Using a massive proprietary dataset of 380 trillion AI tokens, this paper identifies a significant and heterogeneous "AI Premium" where firms with high exposure to advanced, closed-source AI models and interactive work skills earn substantial subsequent stock returns, while such premiums are absent in emerging markets and open-weight or casual AI usage.

Original authors: Nicola Borri, Yukun Liu, Aleh Tsyvinski

Published 2026-06-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicola Borri, Yukun Liu, Aleh Tsyvinski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the global economy as a massive, bustling city. For years, people have been guessing which neighborhoods will boom and which will fade as a new technology arrives. This paper, written by three economists, decides to stop guessing and start measuring. They built a giant, high-definition microscope to watch exactly how much "AI fuel" (called tokens) is being burned in real-time, and then they checked how the stock market reacted to that fuel.

Here is the story of their findings, broken down into simple concepts.

1. The "AI Fuel" Gauge

The authors used a unique dataset called OpenRouter. Think of OpenRouter as a giant gas station that sells fuel to hundreds of different AI engines (like GPT, Claude, Llama, etc.) all at once.

The Data: They watched 380 trillion "tokens" (the basic units of AI work) being used over two years.
The Scale: This isn't just a small sample; it covers about 2% of all AI usage globally. It's like having a camera on every car in a major city, not just a few.
The "AI Factor": They turned this massive amount of data into a single "speedometer" called the AI Factor. This gauge measures how fast AI usage is growing in terms of volume (tokens), money spent, and the number of people using it.

2. The "AI Premium": Who Gets Rewarded?

The core discovery is that the stock market has started pricing in a special reward for companies that are closely tied to this AI fuel. They call this the AI Premium.

The Analogy: Imagine a new, super-efficient electric engine is being invented. Some car companies are already building cars with these engines; others are still stuck with old gas engines. The stock market realizes that the companies building the new engines are taking a risk (what if the new tech fails? what if regulations change?), so investors demand a higher potential return to hold their stock.
The Result: Companies whose stock prices move with the growth of AI usage (high "AI Beta") earned significantly higher returns than those that didn't.
- The Number: A strategy that bought the most AI-exposed companies and sold the least exposed ones made about 64 cents for every $100 invested, every single week. That is a massive amount of money in the financial world.

3. It's Not Just "Tech" Companies

A common mistake is thinking only Silicon Valley companies benefit from AI. This paper shows the AI Premium is much broader.

The Metaphor: Think of AI not as a specific type of car, but as a new type of electricity. It doesn't just power the lightbulb factory; it powers the bakery, the hospital, and the shipping yard.
The Findings:
- Winners: The premium was huge in retail, consumer goods, and heavy industry. These are companies using AI to manage supply chains, talk to customers, or optimize logistics.
- Losers: Surprisingly, some healthcare and non-durable goods companies actually had a negative connection to AI. The market seems to fear that AI might disrupt their specific business models more than it helps them.
- Geography: This premium only exists in developed markets (like the US and Europe). In emerging markets, including China, the effect was invisible. It seems investors only reward AI exposure where the technology is actually being deployed and adopted by sophisticated users.

4. Who Uses the "Good" Fuel?

Not all AI usage is created equal. The paper found that the stock market only rewards serious, heavy-duty AI use, not casual chatting.

The "Intensive" Margin: The premium comes from:
- Closed-Source Models: Using the most advanced, expensive, proprietary AI (like the latest GPT or Claude) rather than free, open versions.
- Seasoned Users: Companies and developers who have been using AI for a long time, not just newbies trying it out for the first time.
- Complex Tasks: Using AI for long, complicated prompts (like coding or scientific research) rather than short, simple questions.
The Takeaway: The market is betting that AI will be a powerful tool for experts and heavy industry, not just a toy for casual users.

5. Jobs: The "Talkers" vs. The "Thinkers"

The authors mapped these company results down to specific job skills. They asked: "Which human skills does the market think will become more valuable as AI grows?"

The Surprising Twist: You might think AI would value "smart" analytical skills (math, science, logic). The data says the opposite.
The "Talkers" Win: The skills with the highest positive AI exposure are interaction, persuasion, instruction, and communication. Jobs that involve talking to people, negotiating, teaching, or coordinating complex systems are seen as the "weak links" that AI cannot easily replace.
The "Thinkers" Lose: Skills involving deep analysis, scientific reasoning, and operations control actually had a negative exposure. The market seems to fear that AI will automate these tasks, potentially reducing the value of firms that rely heavily on them.
The Analogy: If AI is a super-intelligent calculator, the market is betting that the person who can explain the answer to the client and convince them to buy the product is more valuable than the person who just calculated the answer.

6. The Rise of "Agents"

Finally, the paper spotted a new trend: Agentic AI.

What is it? Instead of just answering a question, these AI "agents" can plan a trip, book a flight, check the weather, and send an email all by themselves. They are like digital employees that take action.
The Growth: In 2024, this was almost non-existent. By 2026, over 50% of all AI tokens were being used for these "agent" tasks.
The Cost: Interestingly, as these agents became more common, the cost per task dropped. It's like a delivery service that gets cheaper as it gets faster and more automated.
The Market Signal: While the data is still early, there are signs that the market is starting to reward companies that are successfully using these autonomous agents.

Summary

The paper argues that we are in a transition period. The stock market is currently paying a "risk premium" to companies that are deeply integrated with real, heavy-duty AI usage. This isn't just about tech companies; it's about how AI is reshaping retail, manufacturing, and services. The market believes that in this new world, human interaction and coordination will be the most valuable skills, while pure analytical work might face the most disruption.

Technical Summary: "AI Premium"

1. Problem and Motivation

The paper addresses a fundamental gap in understanding the economic impact of Artificial Intelligence (AI): while AI consumption is growing rapidly, there is no high-frequency, granular, and realized measure of how this consumption translates into risk and value for firms, markets, and workers. Existing literature often relies on task-based automation scores, patent counts, or survey data, which may not reflect the market's real-time pricing of AI exposure.

The authors posit that AI represents a "transition risk" (Aghion and Howitt, 1992; Pastor and Veronesi, 2009). As AI reallocates rents across firms, sectors, and skills, investors should require compensation for holding firms exposed to this systematic reallocation. The central question is whether a measurable "AI Premium" exists in equity markets—specifically, whether firms whose returns covary with realized AI consumption growth earn higher subsequent returns.

2. Data and Methodology

Data Sources

The core of the study is a proprietary, licensed dataset from OpenRouter, Inc., covering 380 trillion tokens of realized AI consumption from January 2024 to April 2026.

Granularity: The data is structured at the user–model–day level, covering millions of anonymized accounts interacting with over 400 large language models (LLMs).
Scope: It includes both frontier closed-source models (e.g., OpenAI, Anthropic, Google) and open-weight models (e.g., Meta, DeepSeek), as well as agentic requests (tool calls).
Realization: Unlike reported estimates, these are realized paid requests, capturing actual expenditure and token counts.
Equity Data: Mapped to CRSP/Compustat for U.S. firms and Compustat Global for international markets.

Construction of the AI Factor

The authors construct a high-frequency AI Factor representing the growth in global AI consumption.

Inputs: Three weekly series: total tokens ( $\Delta \ln Tok$ ), usage dollars ( $\Delta \ln Dol$ ), and distinct active users ( $\Delta \ln User$ ).
Aggregation: The baseline AI factor is the first principal component of the standardized weekly log growth of these three series. This factor captures the common variation in the quantity, expenditure, and reach of AI adoption.
Sub-Factors: The granularity allows for the construction of sub-factors based on:
- Model type (Closed-source vs. Open-weight).
- User intensity (Paid/Core vs. New; Seasoned vs. Non-seasoned).
- Prompt complexity (Long vs. Short).
- Agentic activity (Tool calls vs. Ordinary requests).

Empirical Strategy

AI Beta Estimation: For each public firm, the authors estimate a rolling AI Beta ( $\beta_{AI}$ ) by regressing the firm's weekly log excess returns on the AI Factor and the market factor over a 13-week window. This beta measures the firm's sensitivity to AI consumption shocks.
Portfolio Sorting: Firms are sorted into quintiles based on their $\beta_{AI}$ . The authors analyze the high-minus-low (H-L) spread in subsequent returns.
Fama-MacBeth Regressions: Cross-sectional regressions are run weekly to estimate the market price of AI risk (the AI premium), controlling for standard factors (size, value, profitability, investment, momentum, etc.).
Mapping to Tasks/Skills: Firm-level betas are mapped to occupations using BLS employment weights and then to skills using O*NET ratings to derive a market-implied exposure map for labor.

3. Key Contributions

Unprecedented Data Granularity: The paper leverages the largest cross-provider record of realized AI consumption (380T tokens) to move beyond theoretical or survey-based measures of AI exposure.
Market-Implied Exposure: Unlike task-based measures (e.g., Acemoglu and Restrepo, 2019) or technical automability scores, this study derives AI exposure directly from stock price comovement, reflecting the market's forward-looking valuation of AI risk.
Decomposition of AI Risk: The study distinguishes between different margins of AI consumption (e.g., frontier vs. open-weight, intensive vs. casual), revealing that the market prices specific components of AI adoption differently.
Agentic Economy Evidence: The paper provides early empirical evidence on the rise of "agentic" AI (models executing multi-step workflows), tracking its rapid growth from negligible to over 50% of total token volume.

4. Key Results

The AI Premium in Equity Markets

Positive and Significant: Firms with higher AI betas earn significantly higher subsequent returns. A value-weighted long-short strategy (Long High $\beta_{AI}$ , Short Low $\beta_{AI}$ ) earns 64.1 basis points per week (t-statistic 2.84).
Robustness: The premium persists after controlling for Fama-French five factors, momentum, size, value, profitability, investment, leverage, and accruals. It remains significant even after controlling for high-technology sector returns, semiconductor portfolios, and AI-themed ETFs.
Heterogeneity: The premium is not uniform; it is driven by the intensive, frontier-oriented margin of AI consumption:
- Closed-source models drive a larger premium (53.4 bps/week) than open-weight models (32.3 bps/week).
- Paid/Core and Seasoned users generate a stronger premium than new or casual users.
- Long prompts (complex tasks) are more priced than short prompts.

Geographic and Sectoral Distribution

Geography: The premium is concentrated in developed markets (17.9 bps/week) where firms are closer to the AI frontier. It is absent in emerging markets, including China, where the H-L spread is statistically insignificant.
Sectors: The premium extends beyond technology firms. It is most positive in retail and consumer durables and most negative in health and non-durable goods. Notably, the technology sector shows mixed exposure, with semiconductor firms having positive exposure but other tech firms showing negative or neutral exposure.

Labor Market Implications

Skill Mapping: The market-implied AI exposure is positive for non-routine interactive skills (communication, persuasion, instruction, installation/repair) and negative for non-routine analytical and scientific skills (deductive reasoning, information use).
Quantitative Impact: An occupation one standard deviation higher in interaction-and-communication content has a 0.36 standard deviation higher market-implied AI exposure.
Comparison to Existing Measures: The market-implied exposure explains less than 2% of the variation in existing technical AI exposure measures (e.g., Felten et al., 2021; Eisfeldt et al., 2023), suggesting the market prices a distinct dimension of risk related to realized consumption patterns rather than just technical automability.

The Rise of Agentic AI

Growth: Agentic token share (requests invoking external tools) grew from near zero in 2024 to 52.2% of total tokens by April 2026.
Pricing: Exposure to agentic consumption components (tool calls, cache reads, reasoning tokens) shows positive, economically sizable point estimates (0.3–0.5% per week), though statistical significance is limited by the short sample. This suggests an emerging "agentic premium."

5. Significance and Claims

The paper claims to provide the first market-implied mapping of AI exposure across firms, tasks, and skills using realized consumption data. Its significance lies in:

Validating Transition Risk: The existence of a positive AI premium supports the view that AI is a systematic risk factor requiring compensation, consistent with theories of creative destruction and risk reallocation.
Refining the Narrative: The results challenge the notion that AI is purely a "technology sector" phenomenon. Instead, it is a broad economic force affecting consumer-facing and capital-heavy industries, with distinct winners and losers based on the nature of their tasks (interactive vs. analytical).
Data-Driven Insight: By using realized token consumption rather than stated intentions or technical capabilities, the study offers a more accurate, albeit forward-looking, measure of how the market perceives the economic impact of AI.

The authors remain modest, noting that the sample covers an early, fast-moving phase of AI diffusion, so the estimates reflect the current price of AI exposure rather than a long-run equilibrium. They also acknowledge that OpenRouter data may overrepresent sophisticated users, implying the premium is tied to the market's pricing of observed, intensive AI consumption.