# The Great Expansion How AI changes the boundary of the firm The investment boom is measurable. Its organizational destination is not predetermined. To understand when AI makes a firm bigger, follow the costs, the contracts and the responsibility together. **Answer.** AI can lower coordination costs inside firms and between them. Selective integration is therefore plausible, but the investment record alone does not establish a general expansion of industry ownership. Suppose an AI system makes a difficult project easier to organize. A manager can search internal knowledge, coordinate specialists and check their work at lower cost. It sounds like a reason to bring more activities into the company. But the same system may help a buyer describe the project to an outside specialist, compare offers and verify delivery. The improvement that makes the firm more capable can also make the market a better alternative. That is the central ambiguity behind the great AI expansion. More computing infrastructure, more applications and more influence over customers do not necessarily point toward the same organizational form. A company may design its own accelerator, buy fabrication, contract for electricity and supply a model to a hospital that remains independent. Ownership can deepen in one place while specialization expands in another. This article puts the economic mechanism and the observed record on the same page. First, it establishes the scale of the build-out. Then a deliberately simple teaching model exposes the missing half of the make-or-buy argument. The financial panel, dated company episodes and research literature show what that mechanism helps us understand, and where a stronger conclusion would outrun the evidence. ## The build-out is large. What does it establish? Across Alphabet, Amazon, Microsoft and Meta, annual cash purchases of property and equipment increased from $64.113 billion in FY2018 to $357.508 billion in FY2025. The endpoint is about 5.6 times the starting value. This is a substantial commitment of resources, measured in completed company accounts rather than by adding announcements of future spending. The annual path matters as much as the endpoints. The cohort total fell from $150.447 billion in FY2022 to $140.353 billion in FY2023, before the subsequent rise. A story of uninterrupted acceleration would erase that reversal. Nor does the aggregate describe four identical businesses: Amazon's physical assets support retail and logistics as well as cloud computing, while the other firms have different operating structures. Sources: [Alphabet annual XBRL facts](); [Amazon annual XBRL facts](); [Microsoft annual XBRL facts](); [Meta annual XBRL facts](). **Exhibit: fig-01-buildout** Read these as annual cash commitments by a fixed four-firm cohort. The composition identifies which disclosed company accounts contribute to the total; it does not identify the AI purpose of each asset. A capital-expenditure account answers a narrower question than a corporate map. It records payments for physical assets. It does not tell us whether a particular server supports AI, which production stage moved under common ownership, or who contracts with the final customer. Cash purchases also omit some assets acquired through financing arrangements. Mixing them with lease-inclusive headlines would change the accounting definition in the middle of the comparison. The clock has a boundary too. Microsoft's fiscal year ends in June; the three peers end in December. These are nominal, consolidated-company amounts, including international and non-AI operations. Their sum is a useful fiscal-year cohort comparison, not a synchronized calendar-year measure of global AI investment. The spending increase establishes capacity and commitment. To infer a larger firm boundary, we need an additional argument: why should owning the next activity outperform buying its output from someone else? That is a question about governance, not simply about scale. Sources: [Alphabet annual XBRL facts](); [Amazon annual XBRL facts](); [Microsoft annual XBRL facts](); [Meta annual XBRL facts](); [The Nature of the Firm](). ## AI works on both sides of the make-or-buy decision. Coase's starting point is straightforward: using a market is not free. Finding a supplier, explaining the task, negotiating terms, checking quality and adapting to surprises all consume resources. A firm replaces some of those transactions with managerial direction. It does not make coordination disappear. Managers must still allocate attention, transmit knowledge, motivate people and correct mistakes. A firm boundary is the point at which an activity is organized under common ownership rather than obtained through a market relationship. The relevant comparison is the full cost of delivering a specified result under each arrangement. A supplier's price is only one component; internal production cost is only another. Verification, delivery risk, adaptation and incremental fixed costs belong in the comparison as well. Sources: [The Nature of the Firm](); [Transaction-Cost Economics: The Governance of Contractual Relations](). ### Transaction costs The resources used to arrange, monitor and adapt an exchange, beyond the direct work of producing the output. Search, specification, negotiation and verification can make a cheap quoted price an expensive arrangement. **Example.** A software project can be inexpensive to code but costly to specify and check. AI might help the buyer contract with a specialist just as it helps an internal team coordinate. **Boundary.** These costs are not observed in the financial capex panel. The teaching model makes assumptions about them so the comparison can be inspected. Sources: [The Nature of the Firm](); [Transaction-Cost Economics: The Governance of Contractual Relations](). Consider the teaching model below. Its inputs are arbitrary cost units per equivalent project, not estimates for any company. Internal production starts at 60 units, internal coordination at 30, verification at 10 and expected delivery or liability cost at 10. Integration overhead is a separate assumption. The market alternative has a purchase price of 70, contracting cost of 25, verification cost of 10, expected failure cost of 5 and fixed cost of 5. With overhead set to 10, a 40% reduction in internal coordination cost lowers the internal total to 108, compared with 115 for the unchanged market alternative. The one-sided story therefore favors making the output inside the firm. Now let market contracting costs fall by 40% as well. The market total becomes 105 while the internal total remains 108. The preferred arrangement reverses, although the firm has become more efficient. Sources: [The Nature of the Firm](). **Exhibit: fig-02-two-margins** Change one assumption at a time. The frontier separates lower internal cost from lower market cost under this displayed model. It is not a fitted boundary, a probability of acquisition or a forecast of company behavior. > ILLUSTRATIVE: In the illustrative example, improving internal coordination alone favors making at 108 versus buying at 115. Improving market coordination too favors buying at 105 versus making at 108. The lesson is the comparison, not the chosen numbers. Lower internal costs do not determine the direction of the boundary unless we know what happened to the external alternative. Reducing integration overhead can reverse the result again. In practice, AI may also alter production quality, verification needs, failure risk and capacity constraints; this small model holds those other components at the assumptions shown. This is why general technical capability does not, by itself, predict a new conglomerate. A reusable model may lower the cost of entering another activity, but an API can make a similar capability available to an independent entrant. Ownership needs an additional advantage: difficult-to-contract adaptation, specific assets, privileged information, a scarce complement or a durable quality benefit from common control. Sources: [The Nature of the Firm](); [Transaction-Cost Economics: The Governance of Contractual Relations](); [The Costs and Benefits of Ownership](); [The Vanishing Hand: The Changing Dynamics of Industrial Capitalism](). ## A larger business and a more capital-intensive business are different findings. The financial panel cannot supply the unobserved costs in that teaching model. It can answer a related descriptive question: did physical investment rise only because the firms grew, or did it also become larger relative to their revenue? Dividing cash property spending by the same firm's revenue separates investment scale from the size of the sales base. For the cohort as a whole, the ratio of summed cash property spending to summed revenue rises from 11.964% in FY2018 to 22.310% in FY2025. This is a ratio of sums, not a simple average of four company percentages. It says that a larger share of the cohort's revenue was matched by that year's cash property spending under the chosen accounting definition. Sources: [Alphabet annual XBRL facts](); [Amazon annual XBRL facts](); [Microsoft annual XBRL facts](); [Meta annual XBRL facts](). **Exhibit: fig-03-capital-paths** Moving right means more reported revenue; moving up means more cash property spending. A ray from the origin represents a constant spending-to-revenue ratio. The ordered observations reveal changes in scale and intensity without turning the relationship into a production function. A nearly vertical movement means spending changed much more than revenue over that interval. Movement along a similar origin-ray means greater scale without a comparable change in the ratio. Neither direction identifies which assets serve AI. A ratio can rise because purchases increase, because revenue grows more slowly, or through both channels. The accounting details are therefore part of the explanation. Meta's FY2025 filing distinguishes $69.691 billion of cash property purchases from $72.215 billion including finance-lease principal. This panel uses the former. Its earlier gross purchases were also reconciled against later presentations net of disposal proceeds. Without that reconciliation, a reporting change could appear as an economic change. A higher cash-investment ratio is evidence of a different resource commitment. It is not an ownership score. To move from the accounts to the boundary, we must identify the activity, the operating entity and the responsibility that changed hands. Sources: [Meta annual XBRL facts](); [Meta FY2025 Form 10-K](); [Alphabet annual XBRL facts](); [Amazon annual XBRL facts](); [Microsoft annual XBRL facts](). ## An interface, an operating business and an owner are not interchangeable. Three boundaries make the distinction concrete. The production boundary concerns who creates and delivers the substantive output. The decision boundary concerns who selects the provider, routes the transaction or determines the sequence of action. The liability boundary concerns who owes a remedy or carries professional, contractual or regulatory responsibility when something fails. Ownership may connect these boundaries, but it does not make them identical. OpenAI's Instant Checkout illustrates the separation. Its announcement retains the merchant's role in accepting orders, processing payments, fulfillment, returns and customer support. The interface can expand its influence over discovery and the transaction while the merchant remains the merchant. That can be economically important without establishing ownership of the retail business. Amazon's completed One Medical acquisition is a different kind of event. Yet the entity structure still matters: 1Life provides administrative and management services, while affiliated physician-owned professional corporations deliver medical services. The record supports operating participation in primary care; it does not justify saying that Amazon owns the clinical practices. Sources: [Buy it in ChatGPT: Instant Checkout](); [One Medical Joins Amazon](). ### Residual control rights The right to decide how an asset is used when the contract does not specify every contingency. This is a reason ownership can matter even when ordinary transactions are governed by detailed agreements. **Example.** Owning computing infrastructure can affect who decides its adaptation or use. It does not automatically confer control over the professional duties of a customer using it. **Boundary.** Residual control is not unlimited authority. Law, regulation, contracts and independently owned operating entities still constrain decisions. Sources: [The Costs and Benefits of Ownership](); [Property Rights and the Nature of the Firm](); [One Medical Joins Amazon](). The case instrument keeps 21 dated episodes across seven firms visible while you inspect the definition. Its five categories distinguish technology supply, task augmentation, intermediation, production and operating participation. These are ordered descriptions. They are neither equally spaced quantities nor a sequence that every business must follow. The cases were chosen to make organizational contrasts inspectable. Their number and coverage differ by firm, so the display is not a league table of true corporate breadth. A firm with no qualifying selected episode under a particular rule may still have relevant activities outside this collection. Sources: [Buy it in ChatGPT: Instant Checkout](); [One Medical Joins Amazon](); [Isomorphic Labs and Novartis research collaboration](). **Exhibit: fig-04-participation** Each mark is a dated episode, not a unit of market share. Inspect what the source establishes, including the operating entity, rationale and limitation. An uncertain classification remains uncertain rather than being converted into a midpoint. Under the strict operating-participation rule, eight episodes clearly qualify and a ninth may qualify. Isomorphic Labs' collaboration with Novartis remains bounded between direct research production and industry operation. The range records a classification question, not a statistical confidence interval. The denominator changes when legacy and core activities are excluded. Removing nine such episodes leaves twelve selected cases, of which two to three meet the strict threshold. Tightening a definition and changing a sample are different operations. The following comparison keeps both visible. Sources: [Isomorphic Labs and Novartis research collaboration](); [One Medical Joins Amazon](); [Buy it in ChatGPT: Instant Checkout](). **Exhibit: fig-05-thresholds** The changing count shows how the stated rule acts on this audit. It does not estimate the share of the economy owned by AI firms or the probability that an AI supplier will become an operator. ## The strongest case for ownership is often a problem of adaptation. A chip, compiler, training system and model may need repeated adjustment together. When performance depends on that co-design, a complete contract for every future contingency is difficult to write. A party that commits an asset to one relationship may also lose attractive outside options. Common ownership can reduce repeated bargaining and make coordinated adaptation easier. This is a stronger integration argument than the observation that a technology has many applications. It identifies a problem that market exchange may solve poorly. But it still leaves a comparison to make. Long-term procurement, joint development, information rights, multi-sourcing and repeated relationships can provide safeguards without turning one party into a subsidiary. Sources: [Transaction-Cost Economics: The Governance of Contractual Relations](); [The Costs and Benefits of Ownership](); [Relational Contracts and the Theory of the Firm](); [Partnerships Between Cloud Service Providers and AI Developers: Section 6(b) study](). ### Asset specificity and hold-up An investment is specific when it is worth substantially less in another use or relationship. Once that investment is sunk, bargaining can redistribute its value, which may discourage the investment in advance. **Example.** A team may build a component around one partner's changing technical requirements. Ownership can help adaptation, but a credible long-term or relational contract may also protect the investment. **Boundary.** Technical complexity alone does not establish specificity, and specificity does not prove that acquisition beats every contractual safeguard. Sources: [Transaction-Cost Economics: The Governance of Contractual Relations](); [Property Rights and the Nature of the Firm](); [Relational Contracts and the Theory of the Firm](). Ownership also changes incentives. An independent specialist can retain a return on effort, serve several customers and develop knowledge that a parent cannot easily replace. Acquisition may improve coordination while weakening some of the incentives that made the specialist valuable. Grossman and Hart, and Hart and Moore, make that trade-off central to the allocation of residual control. The FTC's study of selected cloud-AI partnerships is useful because it preserves different contractual features instead of compressing them into a single ownership score. Equity, cloud commitments, information access and consultation rights can create influence and dependence in different ways. Formal independence does not establish easy substitutability; contractual influence does not automatically establish legal integration. Sources: [The Costs and Benefits of Ownership](); [Property Rights and the Nature of the Firm](); [Relational Contracts and the Theory of the Firm](); [Partnerships Between Cloud Service Providers and AI Developers: Section 6(b) study](). **Exhibit: fig-06-relationships** Read the date together with the source's relationship and status. A first-party design, a minority investment, a capacity agreement and an operating disclosure are different observations. Announced capacity remains announced until operation is documented. The 17 selected milestones also prevent the stack from appearing to begin with generative AI. Earlier capabilities and operating relationships remain in the chronology. Their presence is evidence of precedents, not an estimate of the rate at which firms integrated before or after a particular launch. History supplies a useful restraint. The Henry Ford's account says that the Rouge complex never had fewer than 6,000 suppliers, despite its reputation as a monument to integration. It also records the separation of steelmaking in 1981 and its sale in 1989. The example does not establish that AI resembles steel production. It shows why extensive integration and extensive external supply can coexist, and why a boundary can change again when the economic comparison changes. Sources: [Ford Rouge Factory Tour: History and Timeline](); [The Vanishing Hand: The Changing Dynamics of Industrial Capitalism](); [Partnerships Between Cloud Service Providers and AI Developers: Section 6(b) study](). ## A narrow bottleneck can support a very large business. Economic influence need not come from operating every industry a technology reaches. A supplier can provide a difficult-to-replace complement to many independent producers. The relevant questions then concern substitutability, access terms, switching costs and the distribution of bargaining power, rather than the number of customer businesses under common ownership. NVIDIA offers a concrete organizational possibility. Its FY2026 filing reports $193.737 billion of Data Center revenue out of $215.938 billion in total, about 89.7%. The filing also describes a fabless, contracted-manufacturing strategy. A tightly integrated design, systems and software offering can therefore coexist with external fabrication and independent customers. Sources: [NVIDIA FY2026 Form 10-K: Revenue by End Market](); [The Vanishing Hand: The Changing Dynamics of Industrial Capitalism](). **Exhibit: fig-07-specialist** The composition describes NVIDIA's own reported revenue, using its company-defined platform categories and late-January fiscal-year ends. It is neither the share of customer industries NVIDIA owns nor a measure of market concentration. This evidence does not, by itself, establish market power. It establishes why broad downstream ownership is an inadequate prerequisite for economic scale. A narrow position in an important complement may matter more than a wide portfolio of operating subsidiaries, while technological integration can be much greater than ownership of physical production stages. The welfare implications remain conditional. Integration can improve reliability, protect specific investment or remove duplicated markups. It can also weaken access for rivals, discourage independent innovation or concentrate operational risk. Contractual networks preserve incentives and permit specialization, but can conceal dependence in exclusivity, switching costs or control of customer access. The relevant outcomes are prices, quality, innovation, reliability and entry. Ownership breadth alone settles none of them. Sources: [NVIDIA FY2026 Form 10-K: Revenue by End Market](); [Vertical Integration and Firm Boundaries: The Evidence](); [The Costs and Benefits of Ownership](); [Partnerships Between Cloud Service Providers and AI Developers: Section 6(b) study](). ## A launch archive can only grow. An operating boundary can shrink. Counting announcements accumulates activity by construction. A measure of active operations must also record closures, sales, suspensions and changes of scope. The difference is the familiar distinction between a flow of events and a stock at a date: new entries can be positive while exits prevent the active stock from growing. This audit retains Stadia, Mixer and the original standalone Sora product rather than erasing their withdrawals. But their event clocks are not interchangeable. The first retained event for Stadia and Mixer is an exit announcement; for Sora, it is a launch. Drawing three equal-looking business-lifetime bars would turn different source meanings into a false comparison. Sources: [A message on Stadia and our long-term streaming strategy](); [Bringing More Players Into Our Gaming Vision](); [Sora is here](). **Exhibit: fig-08-exits** Keep the event label attached to each date. These three selected withdrawals demonstrate reversibility and a measurement requirement; they do not estimate a venture failure rate. A product closure also has a scope. Ending Mixer did not mean that Microsoft withdrew from every activity in gaming or streaming. Conversely, a familiar product name does not establish uninterrupted operation through the cutoff. A defensible ownership or operating-stock panel needs stable entity identifiers, completion and operating dates, exits and a rule for renamed or successor products. Those records are not supplied by this selected event audit. Keeping that absence visible is more useful than constructing a smooth historical expansion index from today's list of capabilities. Sources: [A message on Stadia and our long-term streaming strategy](); [Bringing More Players Into Our Gaming Vision](); [Sora is here](). ## Better task performance does not identify a change in ownership. The productivity literature makes the organizational question plausible without resolving it. The published customer-support study by Brynjolfsson, Li and Raymond covers 5,172 agents and estimates a 15% average increase in issues resolved per hour. Its effects differ by experience and skill. That is evidence about performance in a particular organization, not an estimate of a 15% expansion in the boundary of the firm. The Cybernetic Teammate experiment examines a different outcome. Its analysis covers 776 randomized professionals from 811 participants, while the main quality regressions use 550 experimental-unit observations because teams submit one solution. Relative to unaided individuals, the reported baseline quality point estimates rise by 0.373 standard deviations for individuals with AI and by 0.245 for teams without AI. These comparisons concern a one-day product-development exercise. They do not test whether AI replaces a legal organization. Sources: [Generative AI at Work: published version](); [Generative AI at Work: revised preprint v2](); [The Cybernetic Teammate, Working Paper 33641](); [The Cybernetic Teammate: workplace collaboration and expertise](). ### What is being estimated? An estimand is the specific effect a study is designed to learn about: a treatment, an outcome, a population and a time horizon. Changing any of those changes the question. **Example.** The effect of access to an AI assistant on issues resolved per hour differs from the effect of owning an AI asset on acquisitions or industry entry. **Boundary.** A credible estimate of one effect is not automatically evidence for another. Different outcomes should not be pooled merely because they all concern AI. Sources: [Generative AI at Work: published version](); [The Cybernetic Teammate, Working Paper 33641](); [Still Waters, Rapid Currents, Working Paper 33777](); [Artificial Intelligence, Firm Growth, and Product Innovation](). The revised Danish study, Still Waters, Rapid Currents, follows another margin again. Its latest wave covers 25,000 workers across 7,000 workplaces in 11 highly exposed occupations. The authors report near-zero average effects on earnings and recorded hours, with intervals ruling out effects larger than 2% two years after ChatGPT's launch, despite changes in tasks and AI-related work. Task reorganization, measured productivity and earnings are distinct outcomes. At the firm level, Babina and coauthors associate AI-skilled employment with innovation and growth in a principally 2010-2018 US panel. Their university-exposure instrumental-variable strategy has its own identifying assumption. It does not identify a post-2022 effect of owning the AI stack on legal entry into new industries. Nor should percentages, standard deviations and earnings estimates from these settings be averaged into an organizational effect. Sources: [Still Waters, Rapid Currents, Working Paper 33777](); [Artificial Intelligence, Firm Growth, and Product Innovation](). The older integration literature reinforces the conditional answer. Acemoglu and coauthors find different associations for upstream and downstream technology intensity in UK manufacturing; whose investment matters changes the prediction. Lafontaine and Slade's review does not support a context-free presumption that integration is efficient or harmful. Ide and Talamas model knowledge hierarchies, including differences between autonomous AI and copilots, but a modeled hierarchy is not an observed acquisition or outsourcing contract. A stronger empirical test would first construct consistent firm-sector histories. It would distinguish owned from contracted capabilities, inherited core activities from new participation, and active operations from cumulative announcements. Then it would ask whether control of a particular capability precedes operating entry where that capability matters, using a credible comparison group and preserving exits. Timing alone is not enough. A firm may invest upstream because it already plans to enter downstream; cash, distribution and expected demand can drive both decisions. Moving the integration variable one year earlier does not remove that anticipation. An elaborate event-study chart cannot supply the missing counterfactual. A smaller comparison may be more informative: assign comparable projects to AI-assisted internal teams and AI-assisted external suppliers, then measure full quality-adjusted delivery cost, verification, adaptation and failure. That would not answer every question about conglomerates. It would answer the relative-cost question that the broader organizational claim often assumes away. Sources: [Vertical Integration and Technology: Theory and Evidence](); [Vertical Integration and Firm Boundaries: The Evidence](); [Artificial Intelligence in the Knowledge Economy](); [The Nature of the Firm](). > INTERPRETATION: The reviewed record establishes a large investment build-out and varied governance arrangements. It does not identify a general causal movement from AI capability to ownership of downstream industries. ## Inspect the dependency that actually moved. The economic answer is conditional but useful. Common ownership becomes more attractive when difficult adaptation, specific investment, verification or responsibility make contracts relatively costly. Independent specialists become more attractive when AI improves specification, comparison and monitoring, or when their knowledge and incentives are valuable outside a hierarchy. Both movements can occur inside the same technology system. A buyer should therefore test substitutability and responsibility rather than count product logos. An investor should distinguish capital intensity from ownership and interface reach from operating exposure. A regulator needs evidence about access, prices, quality and entry before inferring either efficiency or foreclosure. A firm considering expansion should compare ownership and contract for the same output, including integration and failure costs. The great expansion is already visible in investment and technical reach. Whether it becomes an expansion of the firm is a separate, inspectable question. The useful next step is to trace the cost or obligation that changed, preserve the alternative arrangement, and look for the evidence that could reverse the conclusion. ## Sources and methods The empirical source remains the reviewed R1 snapshot, with a cutoff of September 18, 2026. It contains 32 fiscal-year financial observations, 21 selected company episodes, 17 milestones, 13 studies and 61 distinct source URLs. No new observations or causal estimates are introduced by the R5 presentation. Financial amounts are nominal consolidated-company cash purchases of property and equipment and revenue. Microsoft ends in June and the other financial-cohort firms in December. Cash purchases are not lease-inclusive announced capex and are not an AI-only spending measure. The company cases are purposive dated records with source-audited, single-pass editorial classifications. Stage bounds describe unresolved classification, not sampling uncertainty. Legacy/core scope and participation threshold are separate choices. The comparative-cost instrument is an explicitly illustrative teaching model. Its displayed assumptions, formulas and exact values are downloadable separately from the empirical tables. It is not fitted to firms, a welfare calculation, an ownership forecast or an investment recommendation. Sources, exact-value tables, reproduction files and the original research manuscript remain available below. The original manuscript is retained as a versioned research source, not offered as a competing reading mode. ## Glossary ### Transaction costs The resources used to arrange, monitor and adapt an exchange, beyond the direct work of producing the output. Search, specification, negotiation and verification can make a cheap quoted price an expensive arrangement. **Example.** A software project can be inexpensive to code but costly to specify and check. AI might help the buyer contract with a specialist just as it helps an internal team coordinate. **Boundary.** These costs are not observed in the financial capex panel. The teaching model makes assumptions about them so the comparison can be inspected. Sources: [The Nature of the Firm](); [Transaction-Cost Economics: The Governance of Contractual Relations](). ### The boundary of the firm The set of activities organized under common ownership, rather than obtained through exchange with independent parties. **Example.** Using a model through an API does not put the user inside the model provider's legal organization. **Boundary.** Legal ownership, operational control and technological dependence can differ. A single product interface cannot identify all three. Sources: [The Nature of the Firm](); [The Costs and Benefits of Ownership](). ### Residual control rights The right to decide how an asset is used when the contract does not specify every contingency. This is a reason ownership can matter even when ordinary transactions are governed by detailed agreements. **Example.** Owning computing infrastructure can affect who decides its adaptation or use. It does not automatically confer control over the professional duties of a customer using it. **Boundary.** Residual control is not unlimited authority. Law, regulation, contracts and independently owned operating entities still constrain decisions. Sources: [The Costs and Benefits of Ownership](); [Property Rights and the Nature of the Firm](); [One Medical Joins Amazon](). ### Asset specificity and hold-up An investment is specific when it is worth substantially less in another use or relationship. Once that investment is sunk, bargaining can redistribute its value, which may discourage the investment in advance. **Example.** A team may build a component around one partner's changing technical requirements. Ownership can help adaptation, but a credible long-term or relational contract may also protect the investment. **Boundary.** Technical complexity alone does not establish specificity, and specificity does not prove that acquisition beats every contractual safeguard. Sources: [Transaction-Cost Economics: The Governance of Contractual Relations](); [Property Rights and the Nature of the Firm](); [Relational Contracts and the Theory of the Firm](). ### Relational contracting Cooperation supported by repeated interaction, credible future rewards and the parties' outside options, including commitments that cannot be completely specified or enforced in a formal contract. **Example.** A cloud provider and a model developer can coordinate engineering and capacity while remaining separate firms. **Boundary.** An alliance is not necessarily a temporary stage before acquisition. Its stability depends on incentives and credible alternatives. Sources: [Relational Contracts and the Theory of the Firm](); [Partnerships Between Cloud Service Providers and AI Developers: Section 6(b) study](). ### Cash investment intensity In this article, cash purchases of property and equipment divided by the same firm's revenue in the same fiscal year. **Example.** A ratio can rise because investment grows faster than sales, even if both amounts increase. **Boundary.** It includes non-AI assets, omits some financed additions and uses mixed fiscal-year ends across the cohort. A percentage-point change is not the same as percentage growth. Sources: [Alphabet annual XBRL facts](); [Amazon annual XBRL facts](); [Microsoft annual XBRL facts](); [Meta annual XBRL facts](). ### Production, decision and liability Three different questions: who delivers the output, who directs the transaction, and who is responsible when the result fails. **Example.** An AI purchasing interface can guide demand while a merchant continues to accept orders and handle fulfillment and returns. **Boundary.** Influence over a transaction can be substantial without making the intermediary the owner or operator of every participating business. Sources: [Buy it in ChatGPT: Instant Checkout](); [One Medical Joins Amazon](). ### Stocks and flows A flow counts events over a period. A stock describes what exists at a date. An active operating stock requires entries and exits, not just an accumulating archive of launches. **Example.** A discontinued service remains a historical episode but should not silently remain an active business in a later ownership count. **Boundary.** The present audit is a selected event record, not a reconstructed annual stock. Its three withdrawal records do not form a population failure rate. Sources: [A message on Stadia and our long-term streaming strategy](); [Bringing More Players Into Our Gaming Vision](); [Sora is here](). ### What is being estimated? An estimand is the specific effect a study is designed to learn about: a treatment, an outcome, a population and a time horizon. Changing any of those changes the question. **Example.** The effect of access to an AI assistant on issues resolved per hour differs from the effect of owning an AI asset on acquisitions or industry entry. **Boundary.** A credible estimate of one effect is not automatically evidence for another. Different outcomes should not be pooled merely because they all concern AI. Sources: [Generative AI at Work: published version](); [The Cybernetic Teammate, Working Paper 33641](); [Still Waters, Rapid Currents, Working Paper 33777](); [Artificial Intelligence, Firm Growth, and Product Innovation](). ### Complementarity and specialist incentives Choices are complementary when making one work better raises the value of another. Joint control may help coordinate them, while independence can preserve effort, learning across customers and a residual return. **Example.** A chip and its software may benefit from close co-design even when fabrication is contracted out. **Boundary.** Technical integration and legal integration are different. A better integrated system does not show that every stage should share an owner. Sources: [The Costs and Benefits of Ownership](); [Property Rights and the Nature of the Firm](); [NVIDIA FY2026 Form 10-K: Revenue by End Market]().