csQCA · final census · cutoff 2026-09-30

Business Models of AI Foundation Model Vendors

A crisp-set Qualitative Comparative Analysis of the 49 AI foundation model vendors in the population at 2026-09-30, asking which combinations of market position, integration, scope, pricing, knowledge governance and openness are sufficient for the viability of the business model.

0
Vendors
0
Relevant cases
0
Viable
0
Configurations
Status. The newest shipped edition is the monograph v7.3-r3 (October 9, 2026): the analysis and the preliminary findings, which this page states. The April N=49 analysis is superseded; the final v7.3 is in preparation — see Revisions.

AI Foundation Models

Foundation Models (FMs) are large neural networks trained on broad data at scale and adaptable to a wide range of downstream tasks (Bommasani et al., 2021). Two defining properties distinguish them from narrow AI: emergence—qualitatively new capabilities arising at scale—and homogenization—convergence of diverse AI applications toward a shared representational substrate.

4 Structural Economic Features (Burkhardt et al., 2024)
1
Capital-Intensive Pre-Training
Frontier model training exceeds $100M per run, creating structural barriers to entry.
2
Economies of Scale in Inference
Marginal cost per query falls sharply at scale — cost advantage compounds with volume.
3
Data Network Effects (RLHF)
User interaction signals feed back into model quality, compounding advantage for large-base vendors.
4
Non-Rival Knowledge Accumulation
Trained weights serve unlimited concurrent users at near-zero marginal cost after fixed training.
4 Vendor Categories in This Market
Proprietary ClosedRetain weights, access via API subscriptions, accumulate exclusive RLHF signal.
Open-Weight CommunityRelease weights under permissive licenses; monetize through ecosystem and complementary products.
Platform-EmbeddedIntegrate AI into existing ecosystems (Search, Social, OS); monetize via advertising or subscriptions.
Specialized DomainApply FM capabilities to high-value verticals (legal, medical, financial); command premium pricing.

The coexistence and commercial viability of all four categories simultaneously constitutes the central empirical puzzle motivating this study.

Key Result

Of the 44 relevant cases, 36 are viable and 8 died (01.AI/Yi, Adept AI, Databricks, Inflection AI, Recursal AI / Featherless AI, Snowflake, Together AI and Writer). One condition is necessary for viability, ~PAID, free access or payment in data (consistency 0.917: 33 of the 36 viable vendors hold it). The intermediate solution for viability has 2 equally minimal models; the first is

MONOPOLY • INTEGRATED • HIERARCHICAL + INTEGRATED • PAID • HIERARCHICAL + INTEGRATED • HIERARCHICAL • ~OPEN + WIDE • ~PAID • HIERARCHICAL + ~INTEGRATED • ~PAID • HIERARCHICAL • OPEN + ~WIDE • PAID • HIERARCHICAL • ~OPEN → VIABLE

consistency = 0.941 · coverage = 0.889 · 32 of 36 viable vendors covered

No configuration is sufficient for death: the solution for non-viability covers 2 of the 8 (consistency 1.000). Substitution by an adjacent monopoly (P3) is supported; P1, P2, P4 and P5 are contradicted, and the open and free vendors expected to die are viable. All figures →

Theoretical Framework — PEMAK

Brousseau & Pénard (2007) identify three platform functions that apply to all digital platforms — and to AI foundation model vendors.

Matching (M)
User intent → model response. Conditions: MONOPOLY, INTEGRATED
Assembling (A)
Bundling AI across products. Conditions: WIDE, PAID
Knowledge Management (K)
RLHF pipeline governance. Conditions: HIERARCHICAL, OPEN
6 conditions → 64 possible combinations → 17 observed

Complementor Engagement (§2.4)

Third-party complementors building on the model are the mechanism the framework proposes for FM vendor viability. Two ways of attracting them are distinguished below. Engagement is a scope condition and a mechanism examined inside the cases, never a column of the truth table.

Active Orchestration
Proprietary vendors (Anthropic, Cohere): SDK, developer programs, enterprise API — despite closed weights.
closed weights
Ecosystem Generativity
Open-weight vendors (Meta, DeepSeek): permissive license → community complementors at near-zero cost (Zittrain, 2006).
open weights
Chanson-Rocchi structural analogue: CE = API adoption threshold preceding FM vendor viability. Construct content differs from PL (neobank market recognition); threshold logic is preserved.

Key Findings

Most vendors are viable

Outcome

36 of the 44 relevant cases were viable at 2026-09-30 (base rate 0.818); the business model of 8 died (01.AI/Yi, Adept AI, Databricks, Inflection AI, Recursal AI / Featherless AI, Snowflake, Together AI and Writer).

~PAID is necessary

Necessity

Free access or payment in data: 33 of the 36 viable vendors hold ~PAID (consistency 0.917, coverage 0.892, relevance 0.636). It stays necessary in 64 of the 70 robustness variants that analyse PAID.

Sufficient for viability

Solution

The intermediate solution has 2 equally minimal models covering 32 of the 36 viable vendors (consistency 0.941, coverage 0.889). Parsimonious: INTEGRATED + WIDE • ~PAID + ~PAID • OPEN + ~WIDE • PAID • ~OPEN.

No configuration is sufficient for death

~VIABLE

The solution for non-viability covers 2 of the 8 non-viable vendors (consistency 1.000, coverage 0.250): Recursal AI / Featherless AI and Writer.

P3 supported: substitution by an adjacent monopoly

Propositions

MONOPOLY → VIABLE holds for all 10 vendors with an adjacent monopoly (consistency 1.000). P1, P2, P4 and P5 are contradicted: the open and free vendors expected to die are viable.

2 unresolved contradictions

Truth table

17 of 64 configurations are observed. 000010: Black Forest Labs, Inception and Recraft viable, 01.AI/Yi and Inflection AI not; 000111: AI21 Labs viable, Snowflake and Together AI not — below the 0.80 cut-off both ways, so neither row enters a minimisation.

Contributions

01

An outcome for pre-profit markets

Viability as self-sustainability, observed through the death of the business model, with an exhaustive binary coding rule: 44 relevant cases, 36 viable and 8 non-viable.

02

A census under one rule

49 foundation model vendors from the archived record of the largest model gateway, with dated evidence for every entry, death and code; 17 of 64 configurations observed.

03

A configurational test of PEMAK

~PAID necessary (consistency 0.917); an intermediate solution for viability at consistency 0.941 and coverage 0.889; no configuration sufficient for death.

04

A two-level account of viability

Viability is generativity that the configuration lets the vendor appropriate. Of the five propositions derived from it, P3 is supported and P1, P2, P4 and P5 are contradicted.

05

A protocol fixed in advance

Scope, minimum presence, contradictions, the case-to-condition benchmark and 73 robustness variants were fixed before the analysis; 71 were computed and all are reported.

Methodology

Citation

Xu, J. (2026). Model of Models: A Configurational Analytical Framework of Business Models for AI Foundation Models. v7.3-r3 (October 9, 2026).