Competing explanations, external critiques, and adjudication tests

A useful thesis must expose rival mechanisms to tests it does not control. Listing a critique does not rebut it, and assigning every critic a "blind spot" can immunize the document against challenge. Each explanation below therefore states a mechanism, a distinctive prediction, and evidence that would disconfirm it.

The frontier-firm debate contains substantive disagreements and methodological challenges. Substantive explanations differ about cash flow, infrastructure, competition, labor, and security. Methodological challenges ask whether the thesis itself has valid constructs, representative cases, independent evidence, stable thresholds, and prospective tests. The latter cannot be absorbed as one more viewpoint within the market debate.

Competing-explanation map

Table 10.1. Twelve competing explanations with distinct predictions and disconfirming evidence.

The productivity-bull perspective

Mechanism

The bull argues that artificial intelligence is not another application category but a general-purpose technology comparable to electrification, computing, and the internet. Model capability, hardware efficiency, and distribution improve together. Each decline in the cost of intelligence expands the set of economically viable tasks. Frontier firms require large capital bases because the market is correspondingly large. Index inclusion is a sign of successful diffusion, not a systemic defect. Energy demand is evidence that the technology is useful enough to justify a new industrial build-out.

From this perspective, vertical integration is productive. Agents reduce the transaction cost of converting capability into results. Frontier labs, software vendors, and service firms will discover business models through competition. Infrastructure that appears excessive during construction may become the platform for later innovation, just as fiber and cloud capacity enabled unanticipated services. The greatest policy error would be to constrain financing, permitting, or experimentation before the benefits become measurable.

Distinctive prediction

The distinctive prediction is broad, independent diffusion: successful task cost falls; utilization and customer cash rise; process-level productivity appears outside subsidized pilots; and social value grows even if some early investors lose money.

Disconfirming evidence

This explanation is disconfirmed if fully loaded enterprise ROI, successful tasks per kWh, utilization, and revenue per energized megawatt fail to improve after broad deployment, or if gains remain confined to strategic financing and a narrow set of firms.

The bubble-skeptic perspective

Mechanism

The skeptic argues that a real technology can support a financial bubble. Capital providers extrapolate model improvements and user growth into revenues that may not cover compute, depreciation, integration, and customer acquisition. Strategic investors finance labs that purchase their infrastructure, making demand appear more independent than it is. Announced gigawatts, contracted chips, and private valuations reinforce one another before end customers demonstrate willingness to pay for completed work.

The skeptic expects price competition to erode model rents, open-weight systems to weaken scarcity, and enterprises to discover that integration and human supervision consume much of the apparent productivity gain. Public listings may transfer late-stage risk into indexes and retirement portfolios just as the capital cycle peaks. The resulting correction need not invalidate AI; it may simply reveal that infrastructure owners and model providers cannot all earn the margins embedded in their valuations.

Distinctive prediction

The distinctive prediction is that prices, depreciation, weak utilization, circular demand, impairments, cancellations, and refinancing stress appear before broad independent customer cash and process-level return.

Disconfirming evidence

This explanation is disconfirmed by sustained free-cash-flow growth, independently financed customer demand, high utilization, improving return on invested capital, and verified productivity across sectors despite falling model prices.

The financial-stability-regulator perspective

Mechanism

The regulator does not need to predict an equity collapse. The task is to identify plausible channels through which a shock would impair core functions. The Financial Stability Board has highlighted AI-related third-party concentration, market correlations, cyber risk, model risk, and dependence on a small number of providers (FSB 2024; FSB 2025). A regulator therefore asks which banks, insurers, funds, clearing members, utilities, and critical institutions rely on the same firms, clouds, models, or data centers.

From this view, the problem is opacity. Equity investors see public filings; regulators may not see the full network of capacity reservations, guarantees, derivatives, cloud commitments, project vehicles, and operational dependencies. The state should collect exposure data, stress-test common shocks, require continuity planning for critical AI services, and coordinate financial, energy, cyber, competition, and national-security oversight.

Distinctive prediction

The distinctive prediction is measurable spillover from a frontier-firm shock into funding, credit, clearing, margin, or material intermediary losses beyond matched market controls.

Disconfirming evidence

The explanation is disconfirmed if large equity and operational shocks remain contained, substitution works within required recovery times, and exposure maps show little leveraged or common-provider concentration.

The grid-planner and consumer-advocate perspective

Mechanism

The grid planner sees both opportunity and asymmetry. A large, creditworthy customer can justify generation and transmission that benefits a region. But the customer can also announce several possible sites, reserve capacity, delay construction, or leave after dedicated infrastructure is built. Data centers arrive on technology timelines while regulated assets are financed over decades. If contracts are weak, households and ordinary businesses can inherit the cost.

The preferred policy is milestone-based interconnection, transparent queue rules, customer funding of dedicated facilities, termination security, flexible-load incentives, and protection against speculative duplication. Power should be allocated through credible commitments and reliability value rather than through political enthusiasm for any particular industry.

Distinctive prediction

The distinctive prediction is persistent local scarcity: delayed energization, service-level misses, rising rate or reliability costs, water or emissions conflict, and weak transferability of dedicated assets.

Disconfirming evidence

The explanation is disconfirmed where supply, flexibility, efficiency, customer security, and community agreements keep rates, reliability, latency, and environmental burdens within pre-specified limits.

The frontier-platform-strategist perspective

Mechanism

The platform strategist believes that capital, compute, models, consumer distribution, developer adoption, enterprise context, applications, and feedback form a flywheel. Lower cost expands use; more use produces revenue and evaluation data; better products deepen distribution; deeper distribution justifies more infrastructure. Public capital can strengthen the flywheel by funding construction and acquisitions. Chapter 5 applies this mechanism once to the pending boundary case.

The strategist does not necessarily seek to replace all software or services. The objective is to own the orchestration layer: identity, memory, model selection, permissions, tools, evaluation, and billing. Partners can deliver domain work as long as the frontier platform remains the default runtime. In this view, the highest strategic risk is becoming an interchangeable model behind someone else's customer relationship.

Distinctive prediction

The distinctive prediction is that integrated firms gain voluntary user retention, direct billing, owned-model use, better task economics, and proprietary feedback while maintaining application quality and partner participation.

Disconfirming evidence

The explanation is disconfirmed by user departure, dependence on rival models, deteriorating task margins, weak cash conversion, high exception cost, impaired acquisitions, or customer migration to neutral and hyperscaler-managed alternatives.

The hyperscaler-intermediation perspective

Mechanism

The hyperscaler explanation argues that frontier labs do not necessarily own the most valuable layer. Cloud providers already control enterprise identity, data, networking, security, committed spend, marketplaces, systems of record, and procurement. Managed model routers can make several frontier and open models appear behind one endpoint, allowing the cloud to optimize quality, cost, latency, policy, and failover while retaining the master customer relationship.

The result can be model plurality with cloud concentration. A lab may remain scientifically important but commercially interchangeable, while the hyperscaler captures the control plane, billing, observability, and bargaining power. Labs that attempt to bypass the cloud can face hosting-margin changes, bundled alternatives, or restricted enterprise distribution.

Distinctive prediction

The distinctive prediction is rising enterprise spend through managed cloud routers and agent platforms, increasing substitution among underlying models, and a declining share of direct lab contracts, data ownership, and margin even as total inference grows.

Disconfirming evidence

This explanation is disconfirmed if frontier labs retain direct enterprise contracts, workflow data, billing, durable price premia, and customer authority at scale despite hyperscaler routing and bundling.

The competition and conglomerate-governance perspective

Mechanism

The competition and conglomerate-governance perspective argues that cross-layer ownership can convert technical efficiency into foreclosure. A firm controlling compute, a frontier model, and a widely used application can privilege its own model, bundle capacity, restrict interoperability, cross-subsidize entry, and collect workflow data unavailable to rivals. Chapter 5 contains the canonical pending-case test; this perspective supplies the general rival mechanism.

This perspective focuses less on headline purchase price than on post-merger conduct and separability: continued rival access, nondiscriminatory defaults and pricing, data rights, transparent intercompany economics, and the ability to separate critical services in restructuring. Those variables are operationalized in Chapter 5 and H3 rather than repeated here.

Distinctive prediction

The distinctive prediction is observable foreclosure or governance weakness: discriminatory defaults or pricing, reduced rival access, opaque transfer pricing, cross-subsidy, declining separability, or founder decisions that ordinary governance cannot correct.

Disconfirming evidence

The explanation is disconfirmed if model choice, interoperability, independent board oversight, user retention, segment economics, and service separability remain robust and performance improves without exclusionary conduct.

The enterprise-sovereignty perspective

Mechanism

The sovereignty advocate views intelligence as an enterprise dependency comparable to identity, data, payments, or cloud infrastructure. A firm should not let one provider determine which workflows operate, what data is retained, how performance is measured, or whether prices can be changed. Proprietary models can be used, but only behind a customer-controlled architecture with portable context, evaluation, and fallback.

This perspective favors multi-model routing, open-weight alternatives, private inference for sensitive workloads, contractual data rights, and tested exit. It interprets sovereign investment not as an attempt to duplicate every frontier lab, but as insurance against concentration and geopolitical disruption.

Distinctive prediction

The distinctive prediction is that outages, policy changes, price disputes, or ownership changes lead enterprises to buy managed portability, private deployment, approved model subsets, and tested recovery rather than accept unqualified lock-in.

Disconfirming evidence

The explanation is disconfirmed if single-vendor systems remain reliably substitutable and economically dominant after real dependency shocks, or if managed and self-managed alternatives fail to reduce recovery time, legal exposure, or bargaining risk.

The open-ecosystem industrial-policy perspective

Mechanism

This perspective argues that advanced model capability should become a broadly usable input rather than remain a scarce rent controlled by a few frontier providers. Open weights lower the fixed cost of experimentation and adaptation, allow smaller or specialized models to serve routine work, and move competition toward applications, data, integration, and sector knowledge. A cross-layer commercial coalition can support this strategy because chip, cloud, security, software, and application firms may gain from more total deployment even when model-layer rents fall.

The policy program includes access to compute, shared datasets and evaluation tools, plural frontier capability, portable deployment, and targeted rather than sweeping treatment of distillation and model-derived training. Its evidentiary burden is to distinguish coalition advocacy from realized diffusion and to identify whether lower model concentration merely shifts dependency to the substrate.

Distinctive prediction

The distinctive prediction is that capable open-weight models gain production share across smaller firms, public institutions, and specialized workflows; model-layer price premia and concentration fall; application formation and task specialization increase; and a larger share of spending and gross margin moves toward compute, cloud, control planes, data, and applications.

Disconfirming evidence

The explanation is disconfirmed if open-weight systems remain economically marginal, unreliable, or more expensive after security and operations are included; fail to improve switching, application entry, or public-sector access; or leave both model-layer concentration and adjacent-layer value capture materially unchanged.

The labor and public-interest perspective

Mechanism

This perspective begins with incidence. Productivity gains are not socially neutral if workers lose entry paths, communities subsidize infrastructure, ratepayers absorb grid costs, and a small group of equity owners captures the return. Automation can intensify work, expand surveillance, and make performance management less contestable. Errors in benefits, hiring, credit, health, or public services can scale faster than appeals and accountability.

The public-interest response includes worker participation in system design, transparency about automated evaluation, appeal rights, investment in apprenticeship, gain sharing, community-benefit agreements, and competition policy. The objective is not to prevent automation, but to ensure that productivity becomes lower prices, better services, higher wages, or broadly shared ownership rather than only higher rents.

Distinctive prediction

The distinctive prediction is a divergence between firm-level gains and social incidence: reduced entry-level hiring or apprenticeship, weaker labor share, surveillance, regional cost shifting, or unequal access despite rising output.

Disconfirming evidence

The explanation is disconfirmed if productivity gains consistently reduce prices, improve access and job quality, raise wages or mobility, preserve expertise formation, and distribute benefits without extensive corrective policy.

The frontier AI safety and catastrophic-risk perspective

Mechanism

This explanation begins with misuse, malfunction, and the possibility of loss of control rather than valuation or market structure. As general-purpose systems gain autonomy, tool access, persistence, and permissions in critical environments, cyber, biological, fraud, reliability, and control failures can scale through the same concentrated infrastructure and workflows that create economic value. The 2026 International AI Safety Report finds growing real-world evidence for misuse and reliability failures while treating loss-of-control scenarios as uncertain but potentially severe (International AI Safety Report 2026).

This perspective changes the meaning of vertical integration. Owning compute, models, applications, and action channels can improve monitoring and accountability, but it can also concentrate correlated failure and give one system broader access to critical resources. Economic success can increase exposure faster than safeguards mature.

Distinctive prediction

The distinctive prediction is that autonomy, access, and deployment criticality grow faster than reliable evaluation, containment, monitoring, and recovery; severe near misses, restricted releases, safety cases, or mandatory controls become economically material.

Disconfirming evidence

This explanation is weakened if risk-relevant capabilities plateau, tool and permission controls remain effective under adversarial testing, severe incidents remain rare relative to deployment, and independently verifiable safety cases keep exposure bounded.

The national-security perspective

Mechanism

The national-security strategist sees an integrated technology-industrial base. Chips, models, data centers, power, launch, secure communications, talent, and allied adoption shape military readiness, intelligence, science, and economic influence. Frontier firms can become strategic assets before they become financially systemic; failure, cyber compromise, hostile acquisition, or supply interruption can have state-level consequences.

The response is resilient domestic and allied capacity, targeted export controls, continuity planning, secure procurement, supply-chain diversification, and protection of critical talent and infrastructure. Some redundancy and higher cost may be justified because resilience has option value in crisis.

Distinctive prediction

The distinctive prediction is that states finance redundancy, impose access controls, or protect continuity according to defined strategic capabilities even when market returns alone would not justify the action.

Disconfirming evidence

The explanation is disconfirmed if capabilities remain readily substitutable across allies and suppliers, continuity plans work without incumbent protection, and state intervention does not follow the claimed strategic dependency.

Methodological challenges not resolved by the perspective map

Construct validity: A six-input composite without weights, calibration, thresholds, or outcome mapping cannot function as a predictive index. Channel labels therefore require distinct qualification tests and cannot be aggregated to preserve the argument after an adverse result.

Case dependence: SpaceX is an extreme boundary case, and Cursor is an unconsummated transaction at the evidence cutoff. Neither can demonstrate a general pattern without post-close results and comparable cases.

Endogeneity and selection: High valuations, acquisitions, infrastructure, and product success can cause one another. Event studies, matched cases, pre-deal trends, and failed or neutral alternatives are required before assigning causality.

Self-disclosure circularity: A company's rationale shows intent, not validity. Strategic claims require customer, operating, regulatory, or audited corroboration.

Scenario elasticity: A scenario map is useful only if expected indicator signs are recorded in advance and the best-fitting scenario is reported with residuals. Multiple scenarios cannot be declared simultaneously correct whenever the evidence changes.

Prior and threshold discipline: Each research proposition is paired with a rival or null explanation and a decisive observation. Primary hypotheses specify units, horizons, metrics, nulls, and disconfirming outcomes; exploratory propositions are labeled separately.

Synthesis and adjudication rule

The competing explanations demand many of the same measurements but predict different signs, timing, and control points. Enumerating an explanation is not evidence for or against it. Each quarterly review must identify which primary hypothesis changed, compare the observation with its null, report the best-fitting rival explanation, and state which part of the thesis was weakened or rejected. A critique of the method cannot be dismissed as the "blind spot" of a substantive perspective.