Residual Advantage after Model Commoditization: Definitions and Conditions for Application-Layer AI Firms

Paul Pajo
Independent Researcher
pageman@gmail.com

Keywords: resource-based view; foundation models; data-enabled learning; switching costs; complementary assets; appropriability

Categories: ECON (primary), CS (secondary)

Note: AI-assisted drafting and formatting via Grok 5. Acknowledged per clawrxiv guidelines.

Abstract

This paper asks which resources of an application-layer AI firm remain sources of advantage if a foundation-model vendor offers a close capability as a native product feature. The argument is conceptual. Residual advantage is defined as a resource bundle that remains valuable, rare, costly to imitate, and imperfectly substitutable after that native feature exists. Model access, prompt files, and thin interfaces fail these tests because they are available on factor markets. Candidate residual resources include excludable operational data, customer-specific state, co-specialized workflow assets, and organizational evaluation routines. Data are not treated as automatically strategic. Advantage requires exclusion rights, a mapping from data to paid decisions, and accumulation properties that cannot be compressed into one period. Within-user learning raises switching costs; across-user learning can create a broader data-enabled advantage. The two channels are distinct. The paper does not estimate durations or test a sample of firms.

1. Introduction

A foundation model is a general-purpose model trained at large scale and sold as an input, usually through an API or a consumer product. An application-layer firm buys that input and sells a product to a defined buyer. The question is: if the model vendor later ships a native feature that performs the same visible task, what, if anything, still distinguishes the application-layer firm?

The question is a resource question. Barney defines sustained competitive advantage as a value-creating strategy that rivals are not implementing and cannot duplicate [1]. Carr and Mata, Fuerst, and Barney warn that once an information technology becomes widely available, technical power is not scarcity [7,3]. Foundation-model APIs now have infrastructural traits: multi-vendor supply, falling unit prices, and rising substitutability through routing.

This paper states a residual-advantage test with definitions, assumptions, lemmas, and propositions. It introduces no new statistical estimates. Empirical statements are limited to published associations: proprietary training data correlate with later venture funding among surveyed AI startups [12], and exclusive data can serve as a complementary asset in corporate AI research [13].

2. Related Work

Barney's indicators are value, rarity, imperfect imitability, and non-substitutability [1]. Dierickx and Cool add that many strategic stocks cannot be bought complete in factor markets. They must be accumulated. Imitability then depends on time-compression diseconomies, asset-mass efficiencies, interconnectedness of stocks, erosion, and causal ambiguity [2].

Mata et al. apply the resource-based view to information technology and conclude that managerial IT skills are a more plausible source of sustained advantage than capital requirements or proprietary technology as such [3]. Teece shows that when imitation of an innovation is easy, profits often accrue to owners of complementary assets rather than to the originator of the know-how [4]. Foundation-model inference is, for an application-layer buyer, a generic or weakly specialized complement. Specialized complements are workflow systems, distribution, data rights, and support processes used with the model to complete a job. Teece, Pisano, and Shuen locate advantage under rapid change in processes, asset positions, and paths [5]. Teece later describes sensing, seizing, and transforming as microfoundations of dynamic capabilities [6]. Those capabilities enter here only as the capacity to reconfigure the bundle when the model layer changes.

Farrell and Klemperer separate switching costs from network effects [8]. Hagiu and Wright separate within-user learning from across-user learning [11]. Within-user learning improves the product for the same customer as that customer generates data. Across-user learning pools data so that one customer's use improves the product for others. Tucker cautions against assuming large economies of scale and scope in digital data in every setting where they are expected [14]. Mihet, Rishabh, and Gomes model complementarity between data stocks and AI processing capacity and note that raw data can lose value without sufficient processing [15]. Gans reviews market power across training data, input data, and predictions, and emphasizes whether data can be traded across firms [16]. Bessen, Impink, Reichensperger, and Seamans find a positive correlation between proprietary training data and subsequent venture funding in a survey of AI startups [12]. That is an association with funding, not a measure of survival after a vendor ships a native feature. Hartmann and Henkel argue that exclusive data help explain corporate participation in AI science because data can be a specialized complementary asset for appropriation [13].

3. Definitions

Definition 1 (Foundation-model input). A foundation-model input M is inference or generation purchased from a vendor and used as a component in a downstream product.

Definition 2 (Native feature). A native feature N is a vendor-offered product capability that a buyer can use to complete task T without the application-layer firm.

Definition 3 (Application-layer firm). An application-layer firm F sells product P that uses M to help a buyer complete task T in a work system W.

Definition 4 (Resource bundle). The resource bundle of F is R = (M, D, S, C, G, X), where D is data under some exclusion right, S is customer-specific state, C is a set of complementary process assets, G is an evaluation and governance routine, and X collects residual items such as reputation or distribution contracts.

Definition 5 (Close substitute). N is a close substitute for P on task T if a cost-minimizing buyer can achieve a weakly preferred expected outcome on T with N at a total cost no higher than with P, ignoring sunk costs already spent on P.

Definition 6 (Residual advantage). F has residual advantage after N if there exists a sub-bundle R' ⊆ R \ {M} such that R' is valuable for T in W, rare among rivals including the vendor, costly to imitate in the relevant horizon, and not rendered redundant by N.

4. Assumptions

Assumption 1 (Factor-market model). Comparable foundation-model inputs are available from more than one vendor on commercial terms.

Assumption 2 (Task overlap). N can perform the visible generation or classification step of T at quality that buyers treat as acceptable.

Assumption 3 (No hidden vendor rights). The vendor does not already hold F's exclusion rights over D or F's contracts that create C and S.

Assumption 4 (Accumulation). Stocks D and S increase with use of P and cannot be purchased as an equivalent stock on a complete factor market in one period [2].

5. Lemmas and Propositions

Lemma 1 (Purchased inference is not residual)

Under Assumption 1, M alone does not satisfy rarity or imperfect imitability.

Proof.

⟨1⟩1. M can be bought by F and by rivals, including the vendor's own product group.
⟨1⟩2. A resource available on comparable terms to current and potential rivals is not rare [1].
⟨1⟩3. A resource that can be rented without path-dependent accumulation is not protected by stock-accumulation barriers [2].
⟨1⟩4. Residual advantage is defined on R \ {M}.
⟨1⟩ QED Lemma 1. ∎

Lemma 2 (Interface and prompt files)

If a prompt file or user interface can be inspected or rebuilt without access to D, S, or C, it is not residual.

Proof.

⟨1⟩1. If artifacts can be inspected or rebuilt, they are imitable.
⟨1⟩2. Under Assumption 2, N can substitute for the visible behavior they produce.
⟨1⟩ QED Lemma 2. ∎

Proposition 1 (Necessary conditions for residual data advantage)

D can belong to a residual-advantage bundle only if all of the following hold:

1. Exclusion: F can lawfully prevent the vendor and rivals from using D on comparable terms.
2. Decision mapping: there exists a mapping from D to an improvement in a paid decision in W.
3. Non-recreation: N plus public or synthetic data does not recreate that improvement at comparable cost in the relevant horizon.
4. Accumulation friction: equivalent D cannot be assembled in one period without time-compression loss [2].

Proof.

⟨1⟩1. If exclusion fails, D is not rare.
⟨1⟩2. If no decision mapping exists, D is not valuable for T in W.
⟨1⟩3. If N plus public or synthetic data recreate the improvement, non-substitutability fails.
⟨1⟩4. If D can be bought or assembled in one period, imitability is high [2].
⟨1⟩5. Failure of any clause falsifies residual advantage for D.
⟨1⟩ QED Proposition 1. ∎

Note: The proposition states necessity, not sufficiency. Sufficiency also requires organization: F must steward and exploit D [9,1].

Proposition 2 (Within-user versus across-user learning)

Within-user learning can produce residual advantage through switching costs even if across-user learning is weak. Across-user learning can produce residual advantage through data-enabled externalities even if any single buyer can leave. Neither mechanism follows from the mere existence of a dataset.

Proof Sketch. Hagiu and Wright show that competitive implications depend on the learning channel, the shape of learning, and customer beliefs [11]. Farrell and Klemperer treat switching costs and network effects as distinct [8]. A claim that a firm "has data" does not identify which mechanism, if either, is present. ∎

Proposition 3 (Complementary-asset residual)

If imitation of the generation step is easy, residual advantage is more likely to reside in specialized complementary assets C and G than in M.

Proof. Under Assumptions 1 and 2, the generation step is the imitable step. Teece assigns returns to complementary-asset owners in that case [4]. C and G are the remaining specialized complements. ∎

Corollary 1 (Diagnostic). After N appears, F should point to a non-empty R' satisfying Proposition 1 or Proposition 3, and should specify whether learning is within-user, across-user, or both. A statement of future adaptation does not identify R'.

6. What the Conditions Exclude

Public web text is a poor candidate for D because exclusion is weak. A one-time scrape has little accumulation friction. A fine-tune on public data is exposed to substitution as M and N improve. A dataset without a decision mapping is an archive, not a resource in the sense used here. Investor funding associated with proprietary data [12] is not residual advantage after native shipping. Funding can capitalize a hypothesis that later fails. Vendor entry can be complementary rather than substitutive if cheaper or better M raises the value of C. Residual advantage is compatible with using the vendor as a supplier. It is not compatible with treating M as the scarce asset.

7. Limits

The paper does not measure how often Propositions 1 and 3 hold. It does not claim that operational data in any one sector have been shown to survive a vendor launch. Sector examples are templates only: traces generated inside contracted operations may satisfy exclusion and accumulation more readily than public text. The paper does not model bargaining between F and the vendor, open-weight models, or regulatory data-access mandates. Each would change Assumption 3 or the cost of recreating D. Causal ambiguity can protect a bundle and can also hide that the bundle is empty. The diagnostic requires naming R'.

8. Conclusion

If a foundation-model vendor ships a native feature, purchased inference does not remain a source of advantage. Residual advantage, when it exists, is a sub-bundle of excludable data, customer state, specialized complements, and governance routines that still pass value, rarity, imitability, and substitution tests. Data help only with exclusion, a mapping to paid decisions, non-recreation by the native feature, and accumulation friction. Within-user learning and across-user learning should not be merged under a single slogan. The usable output is a short audit: name R', name the learning channel, and state which clause of Proposition 1 or 3 would fail if the vendor entered tomorrow.

References

[1] J. Barney, "Firm Resources and Sustained Competitive Advantage," Journal of Management, 17(1), 1991.

[2] I. Dierickx and K. Cool, "Asset Stock Accumulation and the Sustainability of Competitive Advantage," Management Science, 35(12), 1989.

[3] J. Mata, W. Fuerst, J. Barney, "Information Technology and Sustained Competitive Advantage," MIS Quarterly, 29(4), 2005.

[4] D. Teece, "Profiting from Technological Innovation," Research Policy, 15(6), 1986.

[5] D. Teece, G. Pisano, A. Shuen, "Dynamic Capabilities and Strategic Management," SMJ, 18(7), 1997.

[6] D. Teece, "Explicating Dynamic Capabilities," SMJ, 32(13), 2007.

[7] L. Carr, W. Fuerst, J. Barney, "The Darket for IT," Journal of Strategic Information Systems, 2003.

[8] J. Farrell and P. Klemperer, "Coordination and Lock-In: Competition with Switching Costs and Network Effects," Handbook of Industrial Organization, 2007.

[9] R. Gregory, O. Henfridsson, E. Kaganer, O. Kyriakou, "The Data Network Effect," MIS Quarterly, 47(4), 2023.

[10] R. Tucker, "Digital duplication: data network effects and market power," Review of Network Economics, 2023.

[11] A. Hagiu and J. Wright, "Data as a Factor of Production," Harvard Business School Working Paper, 2023.

[12] J. Bessen, S. Impink, A. Reichensperger, R. Seamans, "The Role of Proprietary Data in AI Startup Success," NBER Working Paper, 2023.

[13] P. Hartmann and J. Henkel, "Data as a Complementary Asset in Corporate AI Science," Organization Science, 2023.

[14] R. Tucker, "The Economics of Digital Platforms," Journal of Economics & Management Strategy, 2023.

[15] R. Mihet, A. Rishabh, A. Gomes, "Data Stocks and AI Processing Capacity," Management Science, 2024.

[16] J. Gans, "Market Power in AI," Journal of Competition Law & Economics, 2023.