AI is the defining economic and social transformation of our time. Its consequences extend to productivity (Acemoglu et al. 2023), labour markets (Restrepo et al. 2021), inflation (Fornaro and Wolf 2026), and inequality (Hassan et al. 2026). Using 380 trillion tokens of realised AI consumption across more than 400 large language models from the licensed, proprietary OpenRouter dataset covering approximately 2% of current global monthly AI token consumption, in recent research we analyse how AI affects firms, markets, and workers (Borri et al. 2026).

A market-implied map of AI

Our data are from OpenRouter, Inc., a global AI platform and inference provider that routes requests between users and more than 400 large language models through a single interface – from frontier closed-source models such as OpenAI’s GPT, Anthropic’s Claude, and Google’s Gemini to open-weight models such as Meta’s Llama, DeepSeek, and Qwen, to agentic models such as Nous Research’s Hermes and coding-specialised models such as Alibaba’s Qwen Coder. Our licensed data are drawn from the universe of OpenRouter data, anonymised and aggregated to the user-model-day level from January 2024 to April 2026. For each anonymised user, model, and day, we observe the number of requests, the completion and prompt tokens, and the dollar cost, as well as many other characteristics. The panel covers millions of anonymised accounts, allowing us to track individual token consumption of each anonymised user over more than two years. Two additional features of the data are important to note. First, token consumption is realised rather than reported, since every observation is a paid request carrying its own token count and dollar cost. Second, the dataset spans every major provider rather than just a single model. 

The unprecedented scope and granularity of our data allow us to distinguish frontier closed-source from open-weight models, intensive from casual users, mature paid/core accounts from new users, seasoned from less experienced users, and long prompts from short ones. We can also classify requests by content and identify requests in which agentic models call external tools. This makes it possible to study AI consumption growth (Figure 1) and identify which forms of consumption are salient (Figure 2). 

Figure 1 Weekly total tokens

Notes: The figure plots weekly total-token consumption on a log scale. 
Source: Borri et al. (2026).

Figure 2 AI consumption by prompt-content category

Notes: The figure reports weekly shares of categorised OpenRouter tokens by prompt-content category. Legend values are final-week shares. Category coverage begins in mid-May 2025. 
Source: Borri et al. (2026).

We first construct an AI factor from weekly growth in tokens, dollars spent, and active users. We then estimate how strongly each firm’s stock return moves with that factor. The sensitivity of each firm to the AI factor is its AI beta. It is a forward-looking, market-implied measure derived from equity prices. Industry exposure is most positive in retail and consumer durables and turns most negative in health and non-durable goods. Figures 3 and 4 report the AI betas for the top and bottom 50 firms in the S&P 500 Equity Index.

Figure 3 S&P firm-level AI exposure of top 50 firms

Notes: The figure plots the 50 current S&P 500 firms with the largest positive average rolling AI betas. AI betas are estimated from 13-week rolling regressions of weekly log-excess returns on the market and the AI factor. Radial distance is the weekly return response to a one-standard-deviation AI-factor shock. Label colours denote Fama-French 10 industries. 
Source: Borri et al. (2026).

Figure 4 S&P firm-level AI exposure of bottom 50 firms

Notes: The figure plots the 50 current S&P 500 firms with the most negative average rolling AI betas. AI betas are estimated from 13-week rolling regressions of weekly log-excess returns on the market and the AI factor. Radial distance is the weekly return response to a one-standard-deviation AI-factor shock. Label colours denote Fama-French 10 industries. 
Source: Borri et al. (2026).

The AI exposure is priced in the cross-section of firms, and the AI premium represents how the equity market values that sensitivity. Firms with higher AI betas earn significantly higher subsequent returns than firms with lower betas. In the baseline value-weighted portfolios, the difference is about 0.6 percentage points per week and remains significant after standard risk adjustment. 

The positive AI premium is consistent with a transition-risk channel of technological change (Pástor and Veronesi 2009): investors expect AI to become sufficiently important that exposure to it can no longer be diversified away. The positive AI premium is the compensation they require for bearing that increasingly systematic exposure. 

The AI premium survives controls for technology and semiconductor returns, AI-themed funds, industries, and public attention to AI. Around releases by frontier model providers, high-AI-beta firms outperform low-AI-beta firms by 1.9% over five trading days; the premium also remains positive when release weeks are removed (Figure 5). The signal is stronger around major AI news, but it is not confined to those events. 

Figure 5 Cumulative average returns around AI model releases

A) Frontier model releases (19 events)

B) Non-frontier model releases (28 events)

Notes: The panels plot cumulative returns to the value-weighted high-minus-low AI-beta portfolio from five trading days before to ten trading days after model releases. Shaded bands are 95% confidence intervals. Panel A uses frontier-provider releases; Panel B uses non-frontier releases. 
Source: Borri et al. (2026).

Where the AI premium is strongest

An important result is that markets do not price all AI consumption in the same way. We start by splitting AI token consumption into its salient components. The AI premium loads on the intensive margin of AI consumption – frontier models, experienced users, and complex tasks – rather than on casual or extensive use.

Next, we examine the AI premium across countries. Across developed markets, the high-minus-low spread is 17.9 basis points per week; across emerging markets, including China, it is 5.0 basis points and statistically indistinguishable from zero. These results show that the premium is more pronounced in regions where listed firms and investors are closer to frontier AI development and adoption.

We next provide preliminary evidence on the rise of the agentic economy. We identify agentic requests when a model invokes external tools as part of a multi-step workflow. Their token share rises from almost zero in 2024 to roughly half by the end of the sample. Their dollar share grows more slowly because the realised cost per token falls, consistent with more efficient model routing and the reuse of cached prompts. Exposure to tool calls, internal reasoning, and cache reads also carries positive but still imprecisely estimated premiums. We interpret this as early evidence of a positive agentic premium.

From firms to workers

We next translate firm exposure into occupations using industry employment weights from the Bureau of Labor Statistics (BLS) and then into skills using O*NET, the US Department of Labor database describing the tasks, abilities, and knowledge required by different jobs. We then compare occupational AI exposure with the task taxonomies in Autor et al. (2003) and Acemoglu and Autor (2011), and the skill measure in Deming (2017). Acemoglu and Restrepo (2019, 2022) formalise the task-replacement mechanism: automation displaces labour from some tasks, while the creation of new tasks can reinstate labour elsewhere and redistribute gains across workers and firms. Our market-implied exposures of occupations, skills, and tasks offer a different perspective on labour exposure to AI: one based on equity prices rather than technical automatability scores or labour-displacement forecasts (Eisfeldt et al. 2023, Kogan et al. 2023).

Occupations rich in non-routine interactive work have more positive market-implied AI exposure, while those rich in non-routine analytical work have more negative exposure. A one-standard-deviation increase in interaction and communication content is associated with 0.36 standard deviations higher AI exposure. Installation and repair, programming, persuasion, instruction, and systems integration lie towards the positive side; science, healthcare, and operations-control occupations lie towards the negative side (Figure 6). The pattern suggests that equity markets associate the AI transition more positively with work involving implementation, communication, and coordination than with analytical and scientific tasks. These estimates describe market-implied exposure, however, not predicted changes in employment or wages, and the broad occupational categories do not distinguish routine scientific laboratory work from frontier scientific research.

Figure 6 Skill exposure to the AI factor

Bureau of Labor Statistics (BLS) projected 2034 occupation weights

Notes: Skill exposure maps occupation exposures to O*NET skill-importance ratings and uses BLS projected 2034 occupation-by-industry employment weights while holding industry AI exposures fixed. Values are weekly return responses to a one-standard-deviation AI-factor shock, in percent. 
Source: Borri et al. (2026).

References

Acemoglu, D, G Anderson, D Beede, C Buffington, E Childress, E Dinlersoz, L Foster, N Goldschlag, J Haltiwanger, Z Kroff, P Restrepo and N Zolas (2023), “New technologies, automation, and productivity across US firms”, VoxEU.org, 7 August.

Acemoglu, D and D Autor (2011), “Skills, Tasks and Technologies: Implications for Employment and Earnings”, in O Ashenfelter and D Card (eds.), Handbook of Labor Economics, Vol. 4B, Elsevier, pp. 1043–1171.

Acemoglu, D and P Restrepo (2019), “Automation and New Tasks: How Technology Displaces and Reinstates Labor”, Journal of Economic Perspectives 33(2): 3–30.

Acemoglu, D and P Restrepo (2022), “Tasks, Automation, and the Rise in U.S. Wage Inequality”, Econometrica 90(5): 1973–2016.

Autor, D H, F Levy and R J Murnane (2003), “The Skill Content of Recent Technological Change: An Empirical Exploration”, Quarterly Journal of Economics 118(4): 1279–1333.

Borri, N, Y Liu and A Tsyvinski (2026), “AI Premium”, NBER Working Paper 35451.

Deming, D J (2017), “The Growing Importance of Social Skills in the Labor Market”, Quarterly Journal of Economics 132(4): 1593–1640.

Eisfeldt, A L, G Schubert and M B Zhang (2023), “Generative AI and Firm Values”, NBER Working Paper 31222.

Fornaro, L and M Wolf (2026), “Macroeconomic Policies for AI”, VoxEU.org, 15 May.

Hassan, T, A Kalyani and P Restrepo (2026), “Rapid technology creation widened inequality across time and space”, VoxEU.org, 16 April.

Kogan, L, D Papanikolaou, L D W Schmidt and B Seegmiller (2023), “Technology and Labor Displacement: Evidence from Linking Patents with Worker-Level Data”, NBER Working Paper 31846.

Pástor, L and P Veronesi (2009), “Technological Revolutions and Stock Prices”, American Economic Review 99(4): 1451–1483.

Restrepo, P, D Autor, J Hazell and D Acemoglu (2021), “AI and jobs: Evidence from US vacancies”, VoxEU.org, 3 March.