pith. sign in

arxiv: 2606.30583 · v2 · pith:LK6IGSY3new · submitted 2026-06-29 · 💻 cs.CY · econ.GN· q-fin.EC· q-fin.GN

AI Premium

Nicola Borri , Yukun Liu , Aleh Tsyvinski This is my paper

Pith reviewed 2026-06-30 03:11 UTC · model grok-4.3

classification 💻 cs.CY econ.GNq-fin.ECq-fin.GN
keywords AI premiumstock returnsAI betalarge language modelstoken consumptionfirm exposureoccupational skills
0
0 comments X

The pith

Firms whose stock returns covary positively with AI consumption growth earn higher future returns, with a value-weighted long-short strategy delivering 64.1 basis points per week.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The authors use data on 380 trillion tokens from over 400 large language models to build an AI factor based on growth in tokens, dollars, and users. Firms whose returns move in tandem with this factor, called high AI-beta firms, generate substantially higher subsequent stock returns. The premium is heterogeneous, appearing strongly for intensive frontier AI use such as closed-source models and long prompts but not for casual or open-weight use, and it extends into consumer-facing and capital-heavy sectors outside technology while remaining absent in emerging markets. Occupational exposure also matters, with roles higher in interaction and communication skills carrying a 0.36-standard-deviation larger implied premium. The work shows how realized AI consumption directly prices into equity markets and labor demands.

Core claim

Using the unprecedented granularity of proprietary AI consumption data, the paper constructs a high-frequency AI factor from aggregate token, dollar, and user growth across hundreds of models. Firms with higher comovement to this factor earn higher subsequent returns, and a value-weighted long-short strategy based on AI betas earns 64.1 basis points per week. The premium is large for loadings on the intensive, frontier-oriented margin of AI consumption but absent for casual or open-weight use; it reaches beyond technology firms into consumer-facing and capital-heavy parts of the economy yet is absent in emerging markets including China; and AI exposure is more positive for nonroutine interac

What carries the argument

The AI factor, built from realized growth in tokens, dollars, and users across large language models, which isolates firm-level exposure through comovement in stock returns.

If this is right

  • High AI-beta firms earn higher subsequent stock returns on average.
  • The premium is concentrated on intensive, frontier AI consumption margins such as closed-source models and long prompts.
  • AI exposure and its associated premium extend into non-technology sectors including consumer-facing and capital-heavy industries.
  • Occupations with higher interaction-and-communication content carry a larger market-implied AI premium.
  • The premium is absent in emerging markets including China.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Portfolio managers could construct tilts toward high AI-beta stocks using consumption-based betas rather than text-based measures.
  • Wage and hiring differentials may emerge across occupations as markets price AI exposure differently by skill type.
  • If consumption data continue to expand, the same factor construction could track how the premium evolves with broader agentic AI adoption.

Load-bearing premise

The AI factor constructed from aggregate token, dollar, and user growth isolates AI-specific exposure rather than capturing correlated movements from other economic or technological trends.

What would settle it

A hold-out test showing that the value-weighted long-short portfolio sorted on AI betas produces zero or negative returns after standard risk-factor controls would falsify the existence of a distinct AI premium.

Figures

Figures reproduced from arXiv: 2606.30583 by Aleh Tsyvinski, Nicola Borri, Yukun Liu.

Figure 1
Figure 1. Figure 1: Weekly Total Tokens The figure plots weekly total-token consumption on a log scale. Total tokens are prompt tokens plus completion tokens. The in-figure annotation reports the first-week level, final observed-week level, cumulative total-token consumption of 380.8 trillion, and growth multiple. The corresponding standardized weekly log growth series is plotted in Online Appendix Figure IA.2. 2.2 Additional… view at source ↗
Figure 2
Figure 2. Figure 2: S&P 500 Firm-Level AI Exposure [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: S&P 500 Firm-Level AI Exposure (continued) [PITH_FULL_IMAGE:figures/full_fig_p024_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Skill Exposure to the AI Factor The figure plots O*NET skill exposure to the baseline AI factor. Skill exposure is constructed by mapping detailed SOC occupation exposures to O*NET skill importance ratings and taking matched-employment-weighted averages across occupations. Panel A uses the current occupation-by￾industry employment structure. Panel B uses BLS projected 2034 occupation-by-industry employment… view at source ↗
Figure 4
Figure 4. Figure 4: Agentic and Non-Agentic Token Consumption [PITH_FULL_IMAGE:figures/full_fig_p032_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Agentic Token Volume and Realized Price per Token [PITH_FULL_IMAGE:figures/full_fig_p033_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Token, Dollar, and Intensity Shares of Agentic AI Consumption [PITH_FULL_IMAGE:figures/full_fig_p034_6.png] view at source ↗
read the original abstract

Using 380 trillion tokens of realized AI consumption across more than four hundred large language models from the licensed proprietary OpenRouter dataset covering approximately 2 percent of current global monthly AI token consumption, we analyze how AI affects firms, markets, and workers. Leveraging the unprecedented size, scope and granularity data, we construct the AI Factor from growth in tokens, dollars, and users, estimate firm-level AI Betas from stock return comovement, and characterize the \emph{AI Premium}. First, we build a high-frequency AI factor and decompose it into salient components. Second, we show that firms whose returns covary more positively with the AI factor -- high AI beta firms -- earn higher subsequent returns, and the AI premium is large and heterogeneous. A value-weighted long-short strategy earns 64.1 basis points per week, and the premium is large for loadings on the intensive, frontier-oriented margin of AI consumption -- closed-source models, paying and seasoned users, and long prompts -- but not on casual or open-weight use. Third, the premium reaches beyond technology firms into consumer-facing and capital-heavy parts of the economy, but is absent in emerging markets, including China. Fourth, the AI exposure is more positive in nonroutine interactive work and more negative in analytical, scientific, and operations-control skills -- an occupation one standard deviation higher in interaction-and-communication content has 0.36-standard-deviation higher market-implied AI exposure. Additionally, we provide early evidence of the rise of the agentic economy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper constructs an AI factor from growth rates in tokens, dollars, and users using 380 trillion tokens of consumption data across >400 LLMs from the OpenRouter dataset. It estimates firm-level AI betas via stock-return comovement with this factor and documents an AI premium: high-beta firms earn higher subsequent returns, with a value-weighted long-short portfolio returning 64.1 bp per week. The premium is heterogeneous (stronger on frontier/closed-source/intensive margins, present outside pure tech sectors but absent in emerging markets) and correlates with occupation-level interaction skills.

Significance. If the AI factor isolates AI-specific exposure, the scale and granularity of the consumption data would represent a substantial advance for asset-pricing and labor-market studies of technological adoption. The out-of-sample return predictability, cross-sectional heterogeneity, and occupation-level patterns would provide falsifiable, economically large evidence on how realized AI use affects firm valuations and skill premia.

major comments (3)
  1. [Factor Construction] Factor-construction section: the AI factor is formed directly from aggregate token/dollar/user growth without reported orthogonalization to market returns, industry portfolios, or other innovation proxies (e.g., R&D intensity or patent-based factors). Because the central claim is that the 64.1 bp premium reflects AI-specific comovement rather than correlated tech or macro trends, this omission is load-bearing for interpretation.
  2. [Portfolio Sorts and Returns] Portfolio-formation and return section: the value-weighted long-short results on AI betas do not report alphas after controlling for Fama-French-Carhart factors or industry fixed effects in either beta estimation or portfolio returns. Without these controls, it remains unclear whether the premium is incremental to known risk factors.
  3. [Heterogeneity Analysis] Heterogeneity tests: while the paper distinguishes closed-source vs. open-weight margins, it does not test whether the factor loadings remain significant after including interactions with existing technology or innovation factors; this test is needed to confirm that the reported heterogeneity isolates AI-specific exposure.
minor comments (2)
  1. [Data and Factor Construction] Abstract and data section: state the precise sample period, rebalancing frequency of the factor, and any winsorization or standardization applied to the growth rates used in factor construction.
  2. [Results Tables] Table/figure captions: ensure all reported returns include t-statistics or standard errors and clarify whether the 64.1 bp figure is raw or risk-adjusted.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. The suggestions for additional robustness checks on factor construction, portfolio returns, and heterogeneity tests will improve the clarity of our results. We address each major comment below and commit to incorporating the requested analyses in the revision.

read point-by-point responses

  1. Referee: [Factor Construction] Factor-construction section: the AI factor is formed directly from aggregate token/dollar/user growth without reported orthogonalization to market returns, industry portfolios, or other innovation proxies (e.g., R&D intensity or patent-based factors). Because the central claim is that the 64.1 bp premium reflects AI-specific comovement rather than correlated tech or macro trends, this omission is load-bearing for interpretation.

    Authors: We agree that showing the AI factor captures incremental variation is central. The factor is built from consumption growth (tokens, dollars, users), which is independent of returns by construction. In revision we will (i) report pairwise correlations of the AI factor with the market, Fama-French factors, and available innovation proxies, and (ii) produce an orthogonalized AI factor by regressing the raw factor on the market and Fama-French factors and re-estimate all betas and premiums with the residual series. These steps will directly address whether the documented premium reflects AI-specific exposure. revision: yes

  2. Referee: [Portfolio Sorts and Returns] Portfolio-formation and return section: the value-weighted long-short results on AI betas do not report alphas after controlling for Fama-French-Carhart factors or industry fixed effects in either beta estimation or portfolio returns. Without these controls, it remains unclear whether the premium is incremental to known risk factors.

    Authors: We will add the requested controls. In the revised manuscript we will report alphas from time-series regressions of the value-weighted long-short portfolio on the Fama-French-Carhart factors. We will also re-estimate firm-level AI betas after including industry fixed effects and repeat the portfolio sorts with industry-adjusted betas. These additions will demonstrate whether the 64.1 bp premium survives standard risk-factor and industry adjustments. revision: yes

  3. Referee: [Heterogeneity Analysis] Heterogeneity tests: while the paper distinguishes closed-source vs. open-weight margins, it does not test whether the factor loadings remain significant after including interactions with existing technology or innovation factors; this test is needed to confirm that the reported heterogeneity isolates AI-specific exposure.

    Authors: We will extend the heterogeneity analysis by adding interactions between AI betas and existing technology/innovation measures (R&D intensity and patent-based factors) in the cross-sectional return regressions. This will test whether the differential premia on closed-source versus open-weight margins, and on intensive versus casual use, remain significant after these interactions. The results will clarify whether the heterogeneity is incremental to broader tech exposure. revision: yes

Circularity Check

0 steps flagged

No circularity in AI factor construction or premium measurement

full rationale

The derivation chain uses an AI factor built directly from independent external consumption data (aggregate token, dollar, and user growth in the OpenRouter dataset), estimates firm betas via return comovement with that factor, and measures the premium via actual subsequent returns on a value-weighted long-short portfolio sorted on those betas. This is a standard empirical asset-pricing test that does not reduce any result to its inputs by construction, nor does it rely on self-citations, fitted parameters renamed as predictions, or imported uniqueness theorems. The central claim remains falsifiable against external return data and is not self-definitional.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract alone supplies no explicit free parameters, axioms, or invented entities; full methods section would be required to audit estimation choices such as beta window length or factor orthogonalization.

pith-pipeline@v0.9.1-grok · 5803 in / 1169 out tokens · 45105 ms · 2026-06-30T03:11:56.048584+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.