From National Value Models to Enterprise Value Models

"The company with the most advanced AI is the strongest." The sentence feels natural. As analysis, it does not hold. Where a firm holds its AI capability (its position) and how much that capability bears on its counterparties (its leverage) are two different variables.

This paper takes the nine cells on which the sister paper, National Value Models (No.10), placed nations, and places firms on them. It does not assume a firm sits in its nation's cell. A nation's position sets the cost of standing in each cell, not where a firm stands.

VURA Working Paper No.11. This page explains the core of the paper in English. The full paper (Japanese original, 2,198 pages; English edition, 3,033 pages) is linked at the end.

This paper in three minutes

• The national grid is carried over to firms. The vertical axis is the tier of AI capability (C1–C3); the horizontal axis is the position where value is produced (M1–M3). Same grid, different subject. It is a map of positions, not a league table.

• A nation's position does not determine a firm's position. All five case companies domiciled in Japan led from a cell different from the nation's leading cell. What a nation sets is the cost of occupying each cell.

• Having and mattering are different things. The everyday word "dependence" is split into three: indispensability, lock-in and desirability. And their signs differ: indispensability depreciates when it is exercised; desirability appreciates the more it is used.

• The single strongest claim: AI capability by itself is not a moat. What defends a position is integration cost — the cost of putting AI to work in real operations — and part of that cost does not fall no matter how capable the models become.

• What remains after efficiency advantages fade is attributed to value-definition capability. But reviewers pointed out that this claim risks circularity, and the paper has weakened it.

• The paper declares falsification conditions for the framework as a whole. It forbids itself from surviving awkward observations by adding auxiliary variables.

The strongest claim in the paper

The paper's core is four claims. The one that runs hardest against practical intuition is this.

Core claim | AI capability by itself is not a moat What defends a firm's position is integration cost. Of the six components of integration cost, three — proof of compliance, verification staff and allocation of liability — do not fall as foundation-model capability rises.

In plain terms — As models get smarter, only the "model part" gets cheaper. Proving you comply with regulation, keeping people who check the outputs, and deciding who is liable when something goes wrong: these three stay, whatever the model's intelligence.

An analogy: the active ingredient and the pharmacy's work

An analogy. A drug's active ingredient gets cheaper once its patent expires, because other companies can make the same compound. Three things do not get cheaper with it: that the drug is approved, that a pharmacist checks the prescription, and that someone is answerable if side effects appear.

AI has the same shape, the paper says. The model (the ingredient) is compressed and gets cheaper. What puts it into practice — proof of compliance, verification staff, allocation of liability (the pharmacy's work) — remains. If there is a moat, it is on the pharmacy's side.

The claim carries its falsification condition: if the three components are shown to have fallen systematically through foundation-model capability gains alone, not through a relaxation of the rules, the claim is refuted. Writing down how a theory would be shown wrong, at the same time as the theory, is the discipline of this series.

On model compression — the speed at which foundation-model unit prices and capability gaps shrink — the paper adopts no single point estimate. Reported declines in API unit prices range from one-ninth to one-nine-hundredth per year, and the appendix keeps that range as a range. The boundary between what compression reaches and what integration cost keeps is defined as absorption distance (Definition C30).

The starting point: taking the National Value Model as given

This paper sits in Layer 3 (management of the firm / micro) of VURA's four-layer architecture. It takes National Value Models, at Layer 0 (national structure), as given conditions — premises the firm cannot move on its own — and works the logic down to the firm. The paper calls this back-illumination.

Three things are taken over: the nine cells, the three world scenarios (Fragmentation, Diffusion, Stagnation; Fragmentation is the base case) and the three words of leverage.

M1 ProductionM2 TransformationM3 Utilization

C3 Critical tierNot yet realizedNot yet realizedNot yet realized

C2 Frontier tierOperatingOperatingOperating

C1 Commodity tierOperatingOperatingOperating

M1 produces AI itself; M2 transforms AI into products and services; M3 uses AI in industry and society. C3 (the critical tier) is treated as a category not yet realized, so six cells operate today.

Proposition C69: a nation's position does not determine a firm's position

This is the discipline the paper puts first. Without it, every country-level description invites misreading.

Proposition C69 | A nation's position does not determine a firm's position Confusing levels of aggregation is prohibited. What a nation's position sets is the cell-occupancy cost, not the cell a firm occupies.

In plain terms — "Japan is at M3×C1, so Japanese firms are at M3×C1" is not a valid inference. A nation's position is a quantity close to a weighted average of its firms' positions, and an individual's position cannot be read off a representative value.

The cases became direct evidence. Of the first set of 13 companies, the five domiciled in Japan (FANUC, Yaskawa Electric, Yokogawa Electric, Shimadzu, Nomura Research Institute) were all led from M2; none led from the nation's leading cell, M3×C1. Of the seven domiciled in the United States, only one led from the nation's leading cell, M1×C2.

An analogy: a district's rents, and which shop takes which floor

An analogy. Land prices in a district set the rent for opening a shop there. They do not decide which shop takes which floor. Expensive districts have small cafés; cheap districts have big stores.

The paper's cell-occupancy cost (Definition C34) is this rent. It is described for ten jurisdictions, Japan included (Definition C33, jurisdiction profiles). These are not evaluations of nations; they describe the conditions a firm domiciled in that jurisdiction faces. Every conclusion is a conditional: "if efficiency is the goal / if a moat is the goal, the comparison of cell-occupancy costs comes out like this."

Having is not mattering — position and leverage

A position on the nine cells says where a firm produces value. How much that position bears on counterparties is a separate variable, which the paper calls corporate leverage. One of its four core claims is that position and leverage are independent.

WordWhose question it isIn everyday terms

IndispensabilityA third party"Can they go elsewhere?"

Lock-inA counterparty already trading with you"Can they switch?"

DesirabilityThe counterparty's own behavior"Do they still choose you?"

Collapse the three into "dependence" and the judgment is lost. Being impossible to bypass, being hard to leave and still being chosen are entirely different states.

Core claim | The asymmetry of signs Indispensability depreciates when exercised; desirability appreciates the more it is used. But the self-reinforcement of desirability has a ceiling.

In plain terms — Every time you say "you would be in trouble without us," the other side starts looking for a way around you. "They still choose us" gets stronger each time they do. But it does not grow without limit.

The ceiling was formalized as an inequality after review (Proposition C46): desirability holds as long as the counterparty's switching cost exceeds the sum of the initial cost of switching and the present value of post-switch operating costs and losses. The paper acknowledges, however, that this proposition has only weak falsifiability.

H2 | Trust infrastructure and the portability dilemma — where to place the interface

If integration cost is a moat, a firm can sell integration. The paper calls this the supply of integration, and its accumulated result trust infrastructure. This is why some firms are bought even when they cost more.

There is a dilemma. Trust infrastructure is stronger the more specific it is to a jurisdiction, a customer, a regulation — and the more specific, the less portable. Make it generic and it travels, but it can also be substituted. The paper calls the ease of carrying it portability, and the boundary between the specific part and the portable part the interface.

The contradiction the adversarial review found, and how it was resolved

The adversarial review (a review whose job is to break the paper's claims) argued that "the five components of trust infrastructure are inseparable" (Proposition C21) and "they can be separated at an interface" (Proposition C30) contradict each other. The paper showed the critique carried an unstated premise: the contradiction arises only if the portable core is assumed to be handed to the customer on its own. Proposition C30 says no such thing.

Proposition C57 | The location of the interface Proposition C21 does not deny that an interface exists; it fixes where the interface goes. The interface is placed not at the point that maximizes portability but at the outermost point at which the portable core still cannot be handed to a customer on its own. A system whose portable core functions alone is portable and substitutable at the same time.

In plain terms — If you hand over only the portable part and the customer can run it in their operations, the interface is too far out.

The three propositions on the interface thus each govern a different judgment: Proposition C31, location (separation after the fact does not recover it); Proposition C50, timing and width (while deployment targets are fewer than the critical number, separation is a net cost); Proposition C57, the outer limit of location (separate too far and the moat disappears).

What AI does not replace — enterprise brain capital and authentic domain data

Part V carries the sister paper's national brain capital down to the firm. Enterprise brain capital is the hard-to-copy accumulation of cognition: tacit knowledge on the shop floor, trust in institutions, long domain experience. Authentic domain data arises from doing the work itself and cannot be bought from outside.

Neither is treated as something to protect unconditionally. The adversarial review argued that enterprise brain capital becomes a license for organizational rigidity. The paper accepted this and set a discipline: being hard to copy does not imply being useful to current operations. Where none of three observations is on record, retention is not treated as maintenance of enterprise brain capital (Proposition C58).

Investments that hold up in every scenario — the sovereignty premium and value-definition capability

Part VI deals with no-regret actions: investments whose value is positive whichever of the three world scenarios arrives. Proposition C55 records the danger that they squeeze fixed costs. "No regret" does not mean "no cost."

Can the sovereignty premium be explained as insurance?

The extra cost of owning a capability in-house is the sovereignty premium. The formula that justifies it as "insurance" is the language of management; accounting standards have no asset class called insurance. So the sovereignty premium shows up in the financials twice: as lower ROIC and as impairment losses (Proposition C51).

How, then, to tell insurance from over-defense? The paper answers with an audit formula.

Proposition C66 | The insurance audit p × L × f ≥ C (p: annual probability of occurrence / L: loss if it occurs / f: share of the loss avoided / C: annual cost). If any of the four variables is not on record, the investment has been evaluated neither as insurance nor as efficiency, and is a cost without a purpose.

In plain terms — If you say "we own it in-house in case something happens," you must hold four numbers: the probability of that something, the loss, the share of the loss that owning it avoids, and the annual cost. Missing even one, it is not insurance. It is spending without a purpose.

An investment that cannot be explained to shareholders is pushed outside the feasible region, whether or not it is justified. The paper does not say disclosure lowers the cost of capital, because no study has measured that.

Value-definition capability — the circularity objection, and a claim downgraded

The fourth core claim is that the difference remaining after efficiency advantages have diminished is attributed to value-definition capability: the capacity to decide for oneself what to make and what counts as value.

Here the review pressed hardest: "If every remaining difference is called value-definition capability, that is circular and cannot be identified." The paper accepted this and re-read Proposition C42.

Proposition C42 (as re-read) Not "the residual is attributed to value-definition capability," but: only where the identification procedure (the three series and the precedence relation of Definition C29) is satisfied may part of the residual be attributed to it. Where no precedence relation is observed, the paper does not treat it as one of its propositions.

In plain terms — The paper does not say "unexplained difference = value-definition capability." Only when a change in definition is observed to come first, and the difference to follow, is part of that difference attributed.

To the objection that "value-definition capability turns into poetry," the paper answers with organizational reproducibility (Definition C32, Proposition C64): is there a recorded procedure the organization can repeat?

Divergence between nation and firm is not an anomaly; it is the default

If a nation's position and a firm's position are separate coordinates, they will diverge. The paper sorts this positional divergence into six types (Definition C36, Proposition C78).

TypeWhat is happeningWhere the excess appears

I Upward divergenceThe nation does not lead C2 production, but the firm aims at M1×C2On the capital side

II Trailing complianceRequirements are fixed within the jurisdiction, but procedures are obtained after the factOn the time side

III Leading complianceCompliance completed in advanceNo excess (works as a moat)

IV Jurisdiction straddlingA single specification deployed across jurisdictions with different conditionsRebuilding the specific layer in each jurisdiction

V Downward retentionThe firm meets the conditions for transformation but stays at M3×C1No excess (the transformation margin is not captured)

VI Upstream specializationPositioned in M3, the firm becomes a supplier to producers outside the jurisdictionIn the composition of the counterparties it negotiates with

Of the first set of 13 companies, divergence was observed in 12 (I = 1 / III = 4 / IV = 5 / V = 1 / VI = 1). For five companies the records needed for a judgment did not exist, and the paper writes exactly that. The types are a diagnostic tool, not a ranking.

Where the 31 companies stand

The cases come in two sets. The first set of 13 was selected on disclosure regime and size. The second set of 18 comprises the companies covered in the sister book series, the AI Management Library; from it the paper took only the list of companies and re-established every fact from primary sources such as Form 10-K filings and Japanese securities reports.

Leading cellFirst set (13)Second set (18)Total (31)

M1×C211011

M1×C1000 (empty)

M2×C28210

M2×C1/C2 boundary404

M2×C1 (alone)000 (empty)

M3×C2022

M3×C1044

The first set skewed toward regulated industries and suppliers of integration, and lacked production-side cases. The second set filled that gap, but M1×C1 and M2×C1 (alone) remain empty.

Some observations from primary sources, stated without evaluative language. TSMC: its top ten customers account for 78% of revenue, the largest for 19%. Costco: membership fees are 1.93% of total revenue but 51.3% of operating income — a contrasting case in which the unit of charging is independent of model usage. BYD: customer concentration of 18.12% is reported only in the Chinese-language annual report, not in the English results announcement.

Turning what could not be observed into an instrument

Of the 18, two — OpenAI and Anthropic — have no statutory periodic disclosure. Three had been assumed; primary verification showed that SpaceX had listed on the NYSE in June 2026.

From this the paper built disclosure availability (Definition C37) into an instrument. Disclosure availability is not a property of the firm but a consequence of the regime and capital structure it sits in. Because the principal occupants of M1×C2 include entities with thin disclosure, the description of that cell rests on weaker evidence than the others, and no change in case selection removes the bias (Proposition C79).

Counterparty disclosure — the results of OpenAI Global, LLC reported in Microsoft's Form 10-K, for example — can serve as a source, but it allows position to be judged, not the two components of leverage (Proposition C80).

The cases illustrate the theory; they are not a list of recommended stocks. The paper introduces corporate neutrality as a new discipline and prohibits, without exception, language of superiority or inferiority, skill or clumsiness, attribution of motive, performance forecasts and investment judgments.

What the paper admits it cannot yet show

The paper incorporates every point raised in internal review, adversarial review and a further round of criticism. This working paper has not been externally peer-reviewed. The points a reader should know first are these.

The response to the "over-construction" critique — falsification conditions for the whole framework

The heaviest criticism: "Attaching a falsification condition to each proposition does not guarantee that the framework as a whole is falsifiable. If an auxiliary variable is added every time an awkward observation appears, the theory survives forever." The paper accepted this as valid and declared three macro-level falsification conditions.

Proposition C76 | Falsification conditions for the framework as a whole If any of the following is observed, the paper will not maintain the framework by adding auxiliary variables. (i) Reduction to a single coordinate — if most of the variance in leverage is shown to be explained by position on the nine cells, so that a second coordinate is unnecessary (ii) Disappearance of integration cost — if the three components (proof of compliance, verification staff, allocation of liability) are shown to have fallen systematically through foundation-model capability gains alone, not through a relaxation of the rules (iii) Identity of signs — if indispensability and desirability are shown to respond to exercise with the same sign

In plain terms — If any of these three happens, the theory is discarded, not repaired.

In Proposition C77 the paper also sorts all 80 propositions into 12 methodological conventions, 14 definitional implications and 54 empirical propositions, accepting the criticism that "every proposition carries a falsification condition" is merely formal while the three kinds are mixed.

There is no control group, and adding the second set does not create one

The first set of 13 was not selected on performance, but it was selected on disclosure regime and size. Having no control group, the paper claims no causation. It claims only two things: that the configurations in question are observed, and that they can be described with the paper's instruments (Proposition C59).

Adding the second set of 18 does not resolve this. Because the sister series is about "the companies that created industries," survivorship bias is stronger in the second set, not weaker. And, as the paper itself writes, since the sister series is by the same author, selection having taken place outside the paper is no guarantee that convenient cases were not chosen.

It does not transfer as-is to unregulated industries

To the criticism that "a trust-infrastructure moat is a privileged strategy available only to regulated industries," the paper did not stop at admitting a limitation. It set out a reduced set of five instruments for domains without institutional preconditions (independence of the unit of charging / buyer concentration / absorption distance / where enterprise brain capital sits / where authentic data sits). It also judges that the AI Foundry Model does not hold in its existing form in domains that lack institutional preconditions (Proposition C68).

Distinguishing what it does not write from what it cannot

On the useful life of AI investments, the paper offers no alternative accounting procedure. Standard-setting lies outside its unit of analysis, and proposing an alternative would come close to accounting advice. Not "we do not write it because we cannot," but "we do not write it because we are not in a position to." With that distinction made explicit, it places only three observations (Proposition C75).

Other weaknesses the paper lists itself

• The unqualified form "Layer 0 is exogenous" is not the paper's claim. It does not specify the scale threshold below which Layer 0 is treated as exogenous (Proposition C73)

• It identifies four places where the exclusion of speculative evidence (▽) is incomplete, and lists 37 circulating claims that could not be verified against primary sources

• Of the 31 companies, it lists 17 whose leading cell could be judged differently (Proposition C72)

Conflict of interest

The author is the representative of VURA Capital, whose business is enterprise-redefinition assessment and management participation. If the paper's framework is read as encouraging investment in "integration cost" and "enterprise brain capital," the author has an incentive to overstate it. The conflict-of-interest clause declares that no arrangement exists between the author and the 31 case companies (verbatim wording to be confirmed against the original).

Implications for strategy leads, CFOs and investors

For strategy and corporate planning. Where your firm stands on the nine cells and how much it bears on its counterparties are two different questions. The twelve-indicator self-assessment (Appendix D) records them separately. The interface question — if you hand over only the portable part, can the customer run it in their operations? — should be answered before any discussion of going generic.

For CFOs. An investment made "in-house, in case" needs four numbers: p, L, f and C. Missing even one, it is spending that is neither insurance nor efficiency.

For investors. The 31 case companies are not recommended stocks, and positions are not a ranking. What cannot be read from the record is not risk; it is an information gap.