FR-12: Cognitive Ability: From Exogenous Residuals to a Cornerstone Reshaping Economic Foundations

Traditional economic paradigm improvements typically manifest as pursuing cost reduction, efficiency enhancement, or price reduction within existing frameworks. However, a true 'paradigm revolution' signifies a subversive change to the standards of 'what constitutes a problem and what constitutes a good answer.' This research report focuses the criteria for an economic paradigm revolution on two core shifts: firstly, whether the 'identity' of production factors changes, i.e., what can be owned, replicated, depreciated, and traded; secondly, whether the 'method of cost determination' changes, i.e., whether marginal cost is an exogenous technological parameter or an endogenous strategic variable.

Under these standards, the key to a revolutionary shift is not merely that AI makes costs lower, which is just a deepening of capability; rather, it is whether 'cognitive ability' itself has become a replicable, depreciable, and ownable endogenous factor of production, and whether marginal costs are no longer exogenously determined by technology. When 'capability' can for the first time be directly purchased, replicated, and accounted for as an asset on the balance sheet, this will undoubtedly rewrite the structure of the production function, heralding a profound economic revolution.

Prior to the sweeping wave of AI, the core hypothesis system supporting the edifice of modern economics was considered 'self-evident.' Among these, factor dichotomy and exogenous technology were foundational, with standard production functions treating capital and labor as substitutable factors, and technological progress viewed as an exogenous residual, as revealed by Solow accounting. Marginal costs were generally assumed to be positive and diminishing, with prices tending towards marginal cost, forming the core logic of market equilibrium; even the 'near-zero marginal cost of replication' for information goods was often treated as a 'special case.' [Verified - Classic Literature] (Solow Growth Accounting; Shapiro & Varian, *Information Rules* (1999)).

Furthermore, scarcity is an axiom of economics; knowledge, as a non-rivalrous public good, is undersupplied by the market, hence institutions like patents and copyrights are needed to artificially create scarcity to incentivize innovation (Arrow 1962, Romer 1990). [Verified - Classic Literature]. The boundaries of firms are determined by Coase's (1937) transaction cost theory, while competitive equilibrium and diminishing returns ensure the temporary nature of excess profits and the convergence of long-term growth. The dichotomy of labor and capital, and the idea that wages equal the marginal product of labor, form the common basis for taxation, labor law, and corporate governance. [Verified - Classic Literature]. These assumptions collectively sketch an economic landscape of 'matter-information duality,' where knowledge is an exogenous public good, scarcity is taken for granted, and firms are containers for transaction costs.

The impact of AI on traditional economic assumptions is not one-dimensional but unfolds through seven specific mechanisms. First is the reality and complexity of 'cost collapse.' Although reports indicate that LLM inference costs have decreased by 1000 times within three years, with the minimum cost for achieving MMLU 42 points dropping from $60/million tokens to $0.06/million tokens [Verified - Directly validated by this report] (a16z 2024-11-12), the key is that this collapse represents 'capability-adjusted prices,' not a universal reduction to zero. Cutting-edge models show price-performance improvements of approximately 5-10x annually in knowledge, reasoning, mathematics, and software engineering. After controlling for hardware advancements, algorithmic efficiency improves by about 3x/year, but the highest performance tiers see a reduction of up to 31x, while the lowest tiers only improve by 1.7x [Verified - Directly validated by this report] (arXiv:2511.23455, MIT FutureTech). This demonstrates that marginal costs are endogenously determined by 'capability tiers,' breaking through the traditional assumption that marginal costs are exogenously given by technology.

Second, AI extends 'non-rivalry' from content replication to 'productive capacity itself.' Once model weights are trained, they can be infinitely replicated, serving countless tasks in parallel without degradation from use. This enables cognitive labor capabilities for the first time to possess the characteristics of 'non-rivalrous capital,' which can be owned, deployed, and reused, yet still provide labor-like services [Inference, based on classic Romer theory]. This 'dumbbell-shaped' cost structure is particularly prominent: training costs for cutting-edge models have grown at 2.4x/year since 2016, with the largest training runs projected to exceed $1 billion by 2027 [Verified - Directly validated by this report] (arXiv:2405.21015, Epoch AI). Fixed costs are rising exponentially, while marginal costs are falling exponentially, concentrating increasing returns among a few entities that possess computing power [Inference].

Even more counter-intuitive is the 'cheap generation, expensive validation' cost disease migration. Despite plummeting model price-performance, the cost of running benchmark evaluations has remained flat or even increased; for example, a single breakthrough evaluation for ARC-AGI cost approximately $3000 [Verified - Directly validated by this report] (arXiv:2511.23455, MIT team). This echoes Baumol & Bowen's (1966) 'cost disease,' implying that the value bottleneck has shifted from 'production' to 'validation, specification, trust, and accountability.' Simultaneously, the Jevons paradox is occurring: efficiency gains are not leading to deflation but to demand expansion. Inference bills are rising because agent tasks can consume 5-30 times more tokens than equivalent chat tasks, with single-task costs potentially reaching 50-60 times higher [Single source - Awaiting first-hand verification] (Inferbase). These mechanisms collectively undermine the six core assumptions of traditional economics, revealing a profound shift in the underlying logic of economics.

Under the impact of AI, the assumptions of the traditional economic paradigm are being completely rewritten, and a new economic blueprint is gradually emerging. First, the production function expands from a simple 'capital + labor' to a more complex *Y = F(M, J, C, D, E)*, where M represents replicable cognitive ability (models), J represents human judgment and specification (non-replicable), C is computing power, D is data, and E is energy. Technology is no longer an exogenous residual but a reproducible and ownable input [Inference, based on M1/M2/M3]. Second, marginal cost is reconfigured as a 'task cost curve,' whose calculation encompasses 'capability tier × token × steps × unit price + validation/trust cost.' Nominal unit price and actual task cost can move in opposite directions, implying that 'zero marginal cost' is a special case applicable only to low-difficulty, single-step tasks, rather than a universal rule [Inference, based on M1/M4/M5].

Third, scarcity is repositioned, shifting from traditional 'scarcity of goods/information' to 'scarcity of judgment/validation/trust/accountability/energy/intent.' When generation becomes cheap, deciding 'what is worth doing,' 'whether it's done correctly,' and 'who is responsible' become the true constraints; what is scarce is no longer bits, but credibility and attention [Inference, based on M4]. Fourth, the source of value undergoes a fundamental migration. Excess profits no longer primarily stem from information asymmetry or production-side monopolies, but rather from the ability to translate vague intentions into verifiable specifications, orchestrate cognitive capability assets, and assume responsibility for final outcomes [Inference, based on M4/M6].

Finally, firm boundaries are no longer simply determined by transaction costs but are dually constrained by 'external computing fixed costs' and 'internal validation and trust costs' [Inference, based on M3/M6 and Coase's theory]. Firms will evolve into 'orchestration contract networks' for capability assets, with their economies of scale dependent on capability reuse rates and the marginal expansion of the trust layer. The labor-capital dichotomy also gives way to a continuum, where the core distribution problem becomes 'who owns replicable cognitive capital.' Research by Acemoglu (2024) has already predicted that AI will widen the gap between capital income and labor income [Verified - Directly validated by this report + Inference], foreshadowing the main battleground of future political economy. Thus, the core of this transformation is capability becoming a replicable factor, marginal costs being endogenously determined by capability tiers, scarcity shifting from goods to judgment/trust/accountability, and firm boundaries being subject to multiple constraints; this far exceeds mere 'deepening,' constituting a genuine economic paradigm revolution.

Examining the veracity of this economic transformation hinges on the precise interpretation of data. For instance, the a16z report's 'equivalent performance dropping 10x annually, MMLU 42-point cost dropping 1000x in three years' [Verified], and Epoch AI's observation that 'the price for GPT-4 level performance drops approximately 40x annually' [Verified], both clearly point to a sharp decrease in costs for a given capability. However, MIT FutureTech's research is more rigorous, finding that cutting-edge model price-performance improves by about 5–10x annually, with algorithmic efficiency improving by about 3x/year after controlling for hardware advancements, but the highest performance tiers see an annual reduction of 31x, while the lowest tiers only improve by 1.7x [Verified]. This highlights the essence of 'tiered cost collapse, not universal zeroing,' and also indicates that any assertion of 'AI marginal costs having reached zero' must be accompanied by explicit definitions of scope, otherwise, conclusions will not be comparable [Verified - Directly validated by this report].

Despite significant micro-level cost changes, no statistically significant 'revolutionary' evidence has yet appeared at the macroeconomic level. Acemoglu (2024) estimates that AI's boost to total factor productivity (TFP) over the next decade will not exceed 0.71%, potentially even less than 0.55% [Verified]. This directly refutes the notion that a 'revolution has already occurred statistically,' reminding us of complex transmission barriers between micro-cost collapse and macro-productivity, such as organizational and institutional inertia. Furthermore, the Jevons paradox reminds us that 'falling unit prices' do not equate to 'falling total bills': agent tasks can consume 5-30 times more tokens than chat tasks, leading to single-task costs rising tens of times [Single source - Awaiting verification]. Simultaneously, training costs are growing at 2.4x/year, and fixed costs are concentrated among a few key players, meaning the narrative of 'everyone can build large models' faces capital threshold constraints [Verified]. AI's 'zero marginal cost' also by no means implies 'zero resource consumption'; energy and computing power, as physical constraints, will become an indispensable baseline in long-term cost models [Verified + Inference].

The shift in the underlying logic of economics will have profound impacts across three levels: business, organization, and product. At the business level, pricing power will shift from traditional 'per token/per seat' to 'per task/per outcome/per unit of value,' as nominal unit prices can no longer accurately predict actual bills. Gross margin structures will be reorganized, with profits from replication approaching zero, while validation, evaluation, routing, orchestration, traceability, and liability assurance will become new profit pools. The Jevons paradox will push computing and energy costs back to the core of business models, making business models deeply tied to hardware and electricity potentially gain a structural advantage [Inference].

At the organizational level, we will witness a 'dumbbell-shaped' organizational structure: a few super-platforms with massive fixed costs coexisting with numerous extremely small, trusted units (e.g., One-Person Companies) specializing in specific domains and possessing non-outsourcable judgment and trust. The rationale for middle management and executive layers will be diminished. Coase's theory of firm boundaries will be rewritten; the internalization or externalization of capabilities by firms will no longer solely depend on transaction costs, but more crucially on the 'unownability of external computing power' and the 'non-outsourcable nature of internal validation and trust' [Inference].

At the product level, the core capabilities of products will be manifested in 'verifiable delivery and cost observability.' Products will need to provide clear cost-per-task dashboards, inference budget management, prefix caching, and explicit lists of assumptions and uncertainties. 'Capability routing' itself will become core intellectual property, as the cost for the same model can vary by several orders of magnitude across different task modalities [Inference, based on M1/M4/M5]. For One-Person Companies (OPCs) such as those under study, their core footing will shift from 'being better at production' to 'possessing non-outsourcable judgment and trust, and productizing validation capabilities,' thereby seeking sources of excess profit in a world of concentrated fixed costs and expensive validation [Inference].

Facing profound changes in the economic paradigm, OPCs under study must adopt forward-looking strategies. First, elevate 'marginal cost' to a first-order observable variable, establishing a meticulous cost-per-task ledger that records the capability tier, token consumption, processing steps, cache hit rate, validation time, and specific fees for each deliverable. Differentiate between 'unit price' and 'task cost' to avoid budget predictions based solely on pricing pages [Actionable]. Second, when tracking trends, always use 'capability-adjusted costs' and explicitly state the methodology, e.g., controlling for baseline performance, distinguishing between Pareto frontiers and full samples, and noting the sample period, to avoid treating partial data (such as '1000x reduction') as a universal trend [Actionable].

Third, establish a dual accounting system for 'fixed costs vs. marginal costs' to calculate reuse rates and break-even points. Treat model or agent capabilities as depreciable fixed assets (including acquisition, fine-tuning, and maintenance costs), while each run and validation is considered a marginal cost. An OPC's viability will directly depend on whether its capability reuse rate can cover and exceed the marginal costs of validation [Actionable]. Fourth, manage 'validation and trust' as core assets and moats. Given that the cost disease has migrated to the validation end, OPCs should accumulate replayable check scripts, evaluation baselines, traceability records, and failure/withdrawal logs; these are assets difficult to replicate in the short term and capable of supporting a premium [Actionable].

Fifth, establish a 'task budget' rather than a simple 'token budget' for OPCs, and actively combat the Jevons paradox. Through measures such as setting inference budget caps, limiting agent steps, and maintaining context hygiene, regularly examine whether efficiency improvements merely lead to an explosion in task volume rather than actual profit growth [Actionable]. Sixth, re-evaluate 'build vs. buy' decisions using Coase's transaction cost theory and a dual framework of fixed costs. The core basis for deciding whether a capability should be outsourced to an API/model or internalized as a self-built agent/private model is the 'affordability of external computing fixed costs' and the 'non-outsourcable nature of validation and trust in that particular step,' rather than solely looking at unit prices [Actionable]. Seventh, treat 'capability' as a depreciable asset, conducting depreciation and substitution analysis. Track the depreciation rate of models/tools used, the speed at which they are replaced, and instances where competitors gain an advantage, guarding against the high concentration and high depreciation risks associated with betting on a single cutting-edge model [Actionable].

Eighth, clearly define the OPC's role as 'specification definer + orchestrator + accountability bearer.' In a world where generation capabilities are increasingly cheap, what is scarce are clear intentions, precise specifications, unique taste, and accountability for results. OPCs under study should upgrade their methodology from 'AI writes for me' to 'I define the problem, AI generates, I independently validate, I take ownership' [Actionable]. Ninth, reuse internal experimental designs to explicitly incorporate economic factors into the OPC production function. Referring to research series such as FR-09/10/11, treat 'computing power/inference budget' and 'validation cost' as independent variables to quantify their impact on delivery quality and unit cost [Actionable]. Tenth, continuously monitor energy and computing power as the physical floor for OPC costs. Incorporate electricity prices, computing availability, and inference energy consumption into the cost model to identify tasks where 'nominal unit prices decrease but physical costs remain rigid,' avoiding misinterpreting short-term price reductions as changes in long-term parameters [Actionable].

Facing this profound economic transformation, numerous unverified propositions and potential risks remain. For example, data concerning 'surging agent task token consumption and single-task costs' primarily originates from vendor blogs and their cited industry analyses; the raw data has not yet undergone independent verification [Single source - Awaiting first-hand verification]. Furthermore, different institutions (e.g., a16z, Epoch AI, MIT) present significantly varying estimates for the rate of cost reduction, such as '1000x in 3 years,' '9-900x/year,' and '5-10x/year.' These discrepancies stem from differences in measurement objects, sample periods, and fidelity, thus requiring explicit methodological statements when cited to avoid confusion [Verified - Methodological differences].

It is noteworthy that Epoch AI's estimates for training costs (e.g., absolute amounts for GPT-4/Gemini) have large error margins, and their authors explicitly acknowledge these estimation limitations; however, the cost growth trend (2.4x/year) is more robust than the absolute values [Verified - Author-stated limitations]. Although Acemoglu's TFP estimates are based on task models and existing literature and are considered a 'conservative lower bound,' they are sufficient to falsify the notion that a 'revolution has already been completed at the macroeconomic statistical level' [Verified - Paper abstract]. The greatest risks lie in treating 'zero marginal cost' as a universal rule, misinterpreting 'falling nominal unit prices' as 'falling total bills,' and equating 'micro-level cost reduction and efficiency gains' with a 'changed macroeconomic paradigm.' These are typical misapplications of old paradigm inertia within the new paradigm, potentially leading to significant business and strategic misjudgments.