Market-Priced AI Exposure (the AI Premium)

Sources#

• AI Premium

Summary#

Borri, Liu & Tsyvinski (arXiv 2606.30583) use 380 trillion tokens of realized AI consumption — every paid inference request across 400+ LLMs routed through the licensed OpenRouter panel, ~2% of global monthly token volume, Jan 2024–Apr 2026 — to price AI exposure in the stock market. The pipeline: build an AI Factor (a high-frequency index of worldwide AI-consumption growth), estimate each firm's AI Beta (how its returns co-move with that factor), and measure the AI Premium (how the market pays for that sensitivity). The headline: high-AI-beta firms earn higher subsequent returns — a value-weighted long-short quintile strategy makes 64.1 basis points per week — and the premium concentrates on the intensive, frontier-oriented margin of AI use. Mapped onto labor, the market-implied exposure is positive on interactive/communication/hands-on skills and negative on analytical/scientific/operations-control skills, the inverse of the naive "AI automates hands, spares knowledge work" story.

Its methodological significance for this vault: it is a third measurement paradigm for AI exposure. Where the AEI taxonomy measures exposure from surveys (reported/anticipated) and usage telemetry (observed), and task-based measures (Acemoglu-Restrepo, Eloundou GPT-4, Eisfeldt-Schubert-Zhang) measure it from what an LLM could theoretically do, this measures it from market-implied risk — what equity prices reveal investors believe about a firm's AI-driven winners and losers. It is forward-looking and priced, not a capability score.

Evidence note. empirical — realized paid requests (not self-report, not classifier-inferred), a user-model-day panel over millions of anonymized accounts, standard asset-pricing machinery (Fama-French five-factor + momentum adjustment, Fama-MacBeth, block-bootstrap, event studies). But the sample is not representative, and the authors say so: OpenRouter "likely overrepresents sophisticated and developer users," and open-weight/roleplay/coding traffic is disproportionate to the aggregator's niche. So the premium is the market pricing of this developer-skewed slice of realized consumption, over an early, fast-moving diffusion window — read as "the current price of AI exposure," not a long-run estimate. See the representativeness caveat under below.

The measurement innovation: realized cross-provider consumption#

What distinguishes the underlying data from the vault's other usage-telemetry sources:

• Realized, not reported. Every observation is a paid request carrying its own token count and dollar cost — no survey, no self-report, no transcript classifier inferring intent.
• Cross-provider, not single-lab. The panel spans OpenAI, Anthropic, Google, Meta, DeepSeek, Qwen, Hermes, and 400+ models; the average account draws on >5 distinct models and 7 in 10 use ≥2 providers. Consumption "is unlikely to reflect the product cycle of any single provider." The authors contrast this explicitly with the AEI, which "classifies only about one million Claude conversations from a single provider" — a deliberate breadth-vs-representativeness trade-off (see caveat).
• User-model-day granularity lets them split consumption along margins that matter for pricing: closed-source vs open-weight, paid/core vs new, seasoned vs casual, long vs short prompts, content category, and agentic (tool-calling) vs ordinary.

The AI Factor is the first principal component of standardized weekly log growth in three series — total tokens, dollars spent, and distinct active users (PC1 loads 0.665/0.559/0.496, 56.5% of joint variance). Growth, not levels, because returns respond to the unexpected component of adoption, not the mechanical upward trend. Weekly tokens rose from 11.4 billion in week one to 15.6 trillion by the final week (380.8T cumulative).

The AI Premium: AI exposure is priced#

Firm AI Beta is a rolling 13-week regression of weekly log-excess return on the AI Factor, controlling for the market — re-estimated every week, look-ahead-free. Firms are sorted into quintiles weekly.

• Portfolio sort: the value-weighted high-minus-low spread is 64.1 bps/week (t=2.84); 56.3 bps after Fama-French 5-factor adjustment (t=2.43), 55.9 bps adding momentum. Equal-weighted spreads are smaller but still positive (35.0 / 30.9 bps), so it is not just a few mega-caps.
• Fama-MacBeth cross-sectional regressions confirm a positive price of AI risk after controlling for size, value, profitability, investment, momentum, reversal, leverage, and accruals (value-weighted β_AI coefficient 18.25, t=3.58 in the most demanding spec).
• The sign fits a transition-risk channel (Pastor-Veronesi 2009): technological revolutions reallocate rents across firms/sectors/capital vintages/workers; when that reallocation is systematic, risk-averse investors demand compensation to hold the most-exposed firms. The premium is the price of that risk, not a free lunch.

Firm-level exposure is not a tech-sector dummy. The most positively exposed S&P 500 names include AppLovin (top), NVIDIA, Lumentum, Expedia, and NRG Energy (power); the most negatively exposed include Moderna (bottom), Estée Lauder — and, tellingly, AMD and ON Semiconductor. Some semiconductor firms load negatively, so this is genuinely a comovement-with-realized-consumption measure, not a "is it a chip stock" classifier.

Where the premium concentrates: the intensive, frontier margin#

The eight split-factor sorts all point the same way — the priced part of AI consumption is the sophisticated part:

Margin	High-exposure side	Low-exposure side
Model class	Closed-source 53.4	Open-weight 32.3
User tenure	Paid/core 66.8	New 21.9
User experience	Seasoned 59.3	Non-seasoned 29.3
Prompt complexity	Long 54.1	Short 33.0

(bps/week, value-weighted high-minus-low). Casual and open-weight use is not priced. This "intensive margin is what matters" pattern is the asset-pricing echo of returns to expertise and the delegation shift — depth of use, not breadth of adoption, is the informative signal.

Geography — priced near the frontier, absent far from it. In developed markets the spread is 17.9 bps/week (t=2.77); in emerging markets only 5.0 bps (t=0.94, insignificant). China A-shares (even with a China-specific AI factor built from mainland OpenRouter consumption) show a negative, insignificant spread. The authors read this through Acemoglu-Aghion-Zilibotti "distance to frontier": AI risk is systematic — and therefore priced — only where listed firms and investors sit close to leading-edge AI development and deployment. This is the market-pricing counterpart to the complements argument: exposure is priced where the complementary AI economy exists.

Not just the tech rally or AI hype#

The premium survives the obvious "you just repackaged the AI stock rally" objection. Re-estimating AI Beta with each of three traded benchmarks added to the market control leaves the spread intact: 49.5 bps controlling for the high-tech industry return, 45.4 bps for a semiconductor portfolio, 63.4 bps for an AI/robotics ETF basket. It is 60.3 bps after demeaning by Fama-French 30 industry and 69.6 bps after controlling for Google Trends "Artificial intelligence" attention. The premium prices exposure to realized consumption, not to tech-sector membership or AI-themed sentiment.

An event study around 19 frontier model releases (Anthropic, DeepSeek, Google, Meta, OpenAI) shows high-beta firms out-returning low-beta firms by ~3% cumulative over the window (1.9% over five days, t=4.17; 1.1% risk-adjusted), with visible pre-drift and a plateau ~5 days after — versus only ~1.5% for non-frontier releases. But removing frontier-release weeks entirely still leaves a 0.395%/week spread (t=1.96): the premium is stronger around releases but accrues gradually throughout the sample, not only on news dates — evidence it is a priced risk, not just a run of positive surprises.

The market-implied skill map: interactive gains, analytical loses#

Projecting firm-level AI Betas onto occupations (BLS employment weights) and skills (O*NET ratings) yields the first market-priced map of AI exposure — read as "conditional on AI's success, the skills the market expects to gain the most and lose the most." The divergence is sharp and, per the authors, robust across the Autor-Levy-Murnane, Acemoglu-Autor, and Deming taxonomies:

• Positively exposed (expected to gain / complement): nonroutine interactive content (+0.15, t=2.87), social skills (+0.16, t=3.44), and — the single strongest loading — interaction & communication (+0.36, t=4.21). Also installation/repair, programming, persuasion, instruction, and nonroutine manual work (+0.25). An occupation one SD higher in interaction-and-communication content has 0.36 SD higher market-implied AI exposure.
• Negatively exposed (expected to lose / be substituted): nonroutine analytic/math content (−0.15, t=−3.41), information use (−0.29, t=−2.43), deductive/inductive reasoning (−0.27), routine manual (−0.41), and operations-control skills. In the detailed skill ranking, Science is the single most negative skill (−0.021 weekly return response), well below operations-monitoring, quality-control, and operation-and-control.

The counterintuitive takeaway: the market prices AI exposure positively for the interactive, persuasive, hands-on, coordination-heavy work and negatively for the analytical and scientific work — the opposite of the reflexive assumption that AI first automates manual labor and spares cognition. The authors connect the positively-exposed skills (interaction, persuasion, instruction, install/repair) to Jones & Tonetti's "weak links" — the slow-to-automate tasks that bind production even after most others are automated. This agrees with the survey/usage finding that experienced workers name judgment and relational/interpersonal work as what AI can't touch — here it shows up as a market bet, not a self-report.

Crucially, this map is orthogonal to prior AI-exposure measures: its correlation with the Felten AIOE, Eloundou GPT-4, Webb patent, and Eisfeldt-Schubert-Zhang measures is below 0.08 in absolute value, and no existing measure explains more than 2% of the market-implied variation (the GPT-4 measure correlates weakly-positive, the pre-generative Webb patent measure significantly negative). Market-implied exposure is measuring something the automability scores are not.

Early evidence of the agentic economy#

Defining agentic requests as those whose finish reason is tool_calls, the agentic share of tokens rises from near zero in 2024 to 52.2% by the last full week (59.3% peak) — a genuine hockey-stick, cross-provider corroboration of the asking→doing shift measured outside any single lab. Tool-call and cache-read shares each reach ~40–50% of tokens; reasoning tokens rise more slowly. The agentic dollar share lags the token share because realized price-per-agentic-token falls — prompt caching serves prefix tokens without recomputing them, and providers route agentic requests to cheaper backing models (consistent with firms managing "skyrocketing AI token budgets"). Cross-sectional pricing of the agentic factors gives positive point estimates (~0.3–0.5%/week) but imprecise (short sample) — the authors call it "early evidence of a positive agentic premium."

The representativeness caveat#

The load-bearing limitation, stated by the authors and worth carrying wherever this source is cited: OpenRouter is a niche aggregator skewed toward developers, open-weight experimentation, and roleplay/coding traffic — it is ~2% of global tokens and "likely overrepresents sophisticated and developer users." So the "380T tokens of realized consumption" is not a representative census of AI use; it is a large, developer-tilted slice. This matters two ways: (1) the premium reflects the market pricing of that slice, over an early diffusion window ("current price," not long-run); (2) as a measurement paradigm, realized-consumption exposure inherits a different non-representativeness than the AEI survey's heavy-user skew or single-provider telemetry — both are biased, in different directions. No current AI-exposure instrument is representative; each is skewed by its collection mechanism. That is a sharpening of the measurement problem, not a solution to it.

Connections#

• Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — the primary methodological sibling: this is a fifth, market-implied measure on a different axis from the AEI's observed/theoretical/reported/anticipated four — realized-consumption + equity-price risk pricing, forward-looking and priced rather than a capability estimate. It doesn't close the "representative sample" gap; it reframes it (every instrument is skewed differently)
• Telemetry vs. Survey Measurement — realized paid requests are the purest telemetry (behavior, not feeling), pushed to cross-provider breadth and then turned into a market-priced signal; the finance-side entry in the measure-the-real-system-not-the-proxy thread
• Conversation-to-Delegation Shift — the agentic-token hockey-stick (0→52%) is independent, cross-provider corroboration of the Codex-internal delegation shift; and the "intensive-margin, not extensive-margin" pricing here rhymes with that page's "output tokens, not user counts"
• Returns to Expertise in Agentic Coding — the premium loads on seasoned/paid/core users and long prompts, and the skill map rewards interactive/relational work — both are the asset-pricing echo of "depth of understanding, not breadth of adoption, is what matters," and of experts naming relational judgment as the residual
• AI Investment Story, Not Efficiency Story — the market-pricing complement: the AI Premium says equity markets do price AI exposure positively (as a transition risk / growth-option bet), consistent with "AI is an investment phase" — the premium is investors demanding compensation for holding the reallocation risk, not evidence of realized efficiency
• Organizational Complements to AI — the developed-vs-emerging (and China-absent) result is distance-to-frontier: AI risk is systematic and priced only where the complementary AI economy exists
• AI Usage Cadences — a further member of the usage-telemetry measurement family; OpenRouter's realized panel is a second high-frequency instrument for watching diffusion (weekly here, hourly there)
• Anthropic Economic Index — the single-provider usage-telemetry program this paper explicitly contrasts itself against (breadth vs representativeness)

Open Questions#

• Is the premium a durable risk price or an early-diffusion artifact? The authors flag the short, fast-moving sample and call it "the current price of AI exposure." Does the transition-risk premium persist, shrink, or invert as AI diffusion matures?
• How much does the developer skew move the answer? OpenRouter's slice is unrepresentative; would a representative realized-consumption panel (if one existed) price the same firms and skills, or is the frontier/intensive-margin concentration partly a sampling artifact of who uses OpenRouter?
• Why is the market-implied skill map orthogonal to every task-based measure (<2% variance)? Is market-implied exposure capturing genuinely different information (forward-looking rents, complement/substitute value rather than technical automability), or is it noisier — and which should labor-impact forecasts trust?
• The agentic premium is only "early evidence" (imprecise). Does a positive agentic premium survive a longer sample, and does the falling price-per-agentic-token (caching + cheap-model routing) erode the dollar-side signal even as token volume explodes?
• Does "Science most negative / interaction most positive" hold out of sample? The finding cuts against the intuition that AI automates cognitive work last; is the market right, or pricing a transient narrative?

Sources#

• AI Premium — Nicola Borri, Yukun Liu & Aleh Tsyvinski, AI Premium (arXiv 2606.30583, 2026-06-29; empirical). §1 Introduction (four parts + literature); §2 Data (OpenRouter panel, AI Factor construction); §3 The AI Premium (portfolio sorts Tables 2–3, Fama-MacBeth Table 4, robustness Table 5, international Table 6, salient-component splits Table 7, event study Table 8 + Fig 2, firm-level Fig 3); §4 Occupations/Tasks/Skills (Tables 9–10, Fig 4 skill map); §5 The Rise of Agentic AI Consumption (Figs 5–7, Table 11)