Public firms that say whether they rent or host AI models show no measurable gap in equity risk; but most firms omit model provenance from 10‑K filings, and noisy disclosure means investors cannot price vendor-concentration risk today.

Main Finding

Using manual, rule‑based textual classification of Item 1 and Item 1A in 1,487 annual reports (508 U.S. registrants, fiscal years 2023–2026), the paper finds no statistically detectable difference between firms that disclose renting AI models (vendor‑dependent) and firms that disclose training/serving their own models (self‑hosted) on three equity‑risk measures: realized post‑filing return volatility, market beta, and implied cost of equity. Crucially, the paper shows that this null result is driven in large part by disclosure practice and measurement error: most firms that discuss AI do not say whether their models are rented or owned, and independent readers disagree on architecture labels at a nontrivial rate (raw agreement low; Cohen’s kappa ≈ 0.571 out‑of‑sample), leaving the design underpowered to rule out economically meaningful differences.

Key Points

• Conceptual mechanism: Renting firms pay a provider price per inference that may be subsidized below provider cost; repricing risk (and correlated repricing across providers) raises input cost variance and, if non‑diversifiable, equity beta and cost of capital relative to owning firms who have capitalized model costs.
• Three testable predictions derived: (1) renting firms have higher return volatility around filings; (2) renting firms have higher market beta and implied cost of equity; (3) ordering pattern should distinguish whether ownership of cost structure or mere exposure to AI drives differences.
• Empirical finding: Vendor‑dependent and self‑hosted firms are statistically indistinguishable on realized volatility, market beta and implied cost of equity across many specifications (including firm fixed effects, relabeling exercises, and cleaned comparison groups).
• Measurement and disclosure problem:
  • ~75% of registrants that mention AI do not disclose model provenance (i.e., they say they have AI‑powered products/platforms but stop short of saying who trains/serves the underlying models).
  • Independent classification of the same filings is noisy: raw agreement is low and out‑of‑sample reproduction of the author’s rubric yields Cohen’s kappa ≈ 0.571 (moderate agreement).
  • Carrying this measured classification error through the estimator shows the study is underpowered: the observed null cannot exclude a true volatility difference up to roughly a quarter of the comparison‑group average volatility, and cannot detect differences smaller than about one‑sixth even under optimistic assumptions.
• Non‑disclosure signal: Firms that mention AI but do not disclose provenance tend to have higher pre‑existing market betas (~+0.10) and a higher implied cost of equity (about 1/15 ≈ 6–7% relatively) than firms that do not mention AI at all; these differences are present before the filings and therefore reflect firm characteristics, not immediate market reaction.

Data & Methods

• Data
  • Sample: 1,487 registrant‑years from 508 U.S. SEC registrants (fiscal years ending 2023–2026).
  • Text source: Item 1 (Business) and Item 1A (Risk Factors) of Form 10‑K; Item 1A is primary, Item 1 used when decisive language appears there.
• Classification
  • Rule‑based rubric, pre‑specified and applied to each filing: classifies each registrant‑year as vendor‑dependent (explicit third‑party provenance), self‑hosted (explicit in‑house training/serving), or no material AI exposure.
  • If both disclosed, coded as vendor‑dependent (conservative, since any third‑party dependency creates the exposure).
  • Boundaries: naming a cloud provider ≠ vendor dependency (focus is on model provenance/inference pricing); provable product‑level statements required to code architecture.
• Reliability assessment
  • Out‑of‑sample validation: independent readers/classifiers applied to held‑out filings. Raw agreement on architecture low; Cohen’s kappa ≈ 0.571 for reproducing the rubric’s label.
• Empirical strategy
  • Outcomes: post‑filing realized volatility, market beta, implied cost of equity.
  • Controls: firm size, profitability, leverage, industry exposure, volume of risk disclosure, and specification robustness (fixed effects, relabeling, cleaned comparison groups).
  • Measurement‑error adjustment: the paper propagates classification error through the estimation to convert null results into bounds on what effect sizes remain consistent with the data and measured unreliability.
• Robustness checks: many specification variants, repeated relabeling, firm‑level pre/post comparisons, and cleaned comparison groups.

Implications for AI Economics

• For researchers
  • Textual methods applied to optional disclosures can be severely limited by selective omission and labeling ambiguity; validation and propagation of classification error are essential. A null from optional text disclosure does not imply absence of economic effect—measurement bounds are necessary.
  • Studies of input‑side counterparty risk (vendor concentration in AI) require either better disclosure or alternative data (contracts, procurement, vendor invoices, or direct survey) because corporate filings do not reliably reveal model provenance.
• For investors & asset pricing
  • Current public filings are insufficient for investors to assess vendor‑concentration risk from AI inference markets. Pricing of this exposure cannot be reliably done from Form 10‑K text alone.
  • The observed higher pre‑existing betas among firms that mention AI without disclosing provenance suggests that selective disclosure itself may be informative about firm risk characteristics; investors should treat voluntary AI mentions with caution.
• For regulators & policy
  • Vendor‑concentration (provenance of embedded models and dependence on third‑party inference providers) is a potentially systemically relevant attribute that is not captured by existing mandatory disclosure rules. If policymakers want markets, supervisors and researchers to price and monitor this exposure, a reporting rule requiring structured disclosure of model provenance (or at least material third‑party AI dependencies and major providers) would materially improve measurability.
  • Absent mandatory disclosure, supervisory assessment of concentration risk in AI supply chains will be hampered and may require targeted data‑collection initiatives (e.g., regulator surveys, contract reporting).
• For firms / corporate governance
  • Firms implicitly choose what to reveal about their AI supply chains; non‑disclosure behavior correlates with higher underlying market risk and may reflect governance, business model or product risk differences. Boards and risk committees should consider whether voluntary disclosure about AI dependencies is in investors’ and counterparties’ interests.
• Broader point
  • The paper reframes the counterparty‑dependence literature by moving it from observable customer concentration to an input (AI model) whose provenance is largely invisible in voluntary corporate text—highlighting an important gap in the information environment around fast‑adopted, input‑sensitive technologies like AI.

Limitations noted in the paper - The study tests a coarse implication (architecture label) rather than the full model (subsidy gap, quantities, provider repricing hazard) because those economic primitives are not in public filings. - Classification noise and voluntary disclosure mean the estimates are underpowered; the null is informative only after converting it into bounds that account for measured unreliability.

Suggested next steps (implicit in the paper) - Obtain structured data on vendor relationships (through regulatory reporting, surveys, procurement data or transaction‑level records). - Advocate or experiment with a standardized disclosure regime for material AI vendor/provenance information to enable market and supervisory pricing of vendor concentration risk.