Intelligence is No Free Lunch: How the Commodification of Cognitive Abilities Reshapes Fundamental Economic Assumptions

Almost every AI launch repeats the narrative that “intelligence is becoming as cheap as tap water,” and data seems to support this assertion: the price of large model inference at equivalent capability levels has dropped by several orders of magnitude in just a few years. This has led “marginal costs approaching zero” to become a default assumption for entrepreneurs, inducing them to blindly stack Agents, tokens, and automation, believing that business models will naturally succeed. However, this understanding overlooks a crucial contrast: despite plummeting unit prices, enterprises' total expenditures often rise rather than fall, which is the most easily misunderstood core contradiction in the current AI transformation.

This phenomenon is not accidental but reveals the true locus of change. Essentially, it's not merely “AI making costs lower” – that would simply be capability upgrades, an optimization within the old paradigm. What is genuinely happening is that **“cognitive abilities” are, for the first time, becoming a factor of production that can be owned, reproduced, depreciated, and traded.** This implies that the determination of marginal costs is no longer exogenously given by technology but is endogenously determined by capability tiers and task structures. The key to discerning a paradigm shift lies in whether the nature of the factor of production and the method of cost determination have changed. In the old paradigm, knowledge, despite its characteristics of “high fixed costs and near-zero marginal replication costs,” was treated as a “special case + institutional patch.” The fundamental difference in this revolution is that even the “cost of productive capacity” itself has become a reproducible and endogenously determined variable.

The collapse of AI costs is not universally zeroing out but is highly layered. a16z data shows that LLM inference costs, measured by MMLU-equivalent performance, have decreased 1000-fold within three years; Epoch AI also indicates that the inference price to reach GPT-4 level decreases by approximately 40 times annually, but the rate of decrease varies significantly across different capability milestones. A review by MIT FutureTech is even more precise: cutting-edge model price-performance improves by 5-10 times annually, but the highest performance tier saw a 31-fold reduction, while the lowest tier only saw a 1.7-fold reduction. This demonstrates that “marginal costs approaching zero” are task-specific and capability-layered; what truly decreases significantly is the “price to acquire a given capability,” whereas cost changes for lower-end capabilities are less pronounced.

Concurrently, the rise in enterprises' total bills has its inherent logic. Firstly, the cost structure exhibits a “dumbbell shape”: the training costs for cutting-edge models have grown 2.4 times annually since 2016, with GPT-4's training cost reaching as high as $40 million. This means that **fixed costs are rising exponentially, while marginal costs are falling exponentially**, with economies of scale concentrated in the hands of a very few entities controlling computing power. Secondly, while generation is cheap, verification is more expensive. The MIT team found that despite the plummeting model price-performance, the costs of running benchmark evaluations (such as SWE-bench Verified, ARC-AGI) have remained flat or even increased. This is an embodiment of Baumol and Bowen's “cost disease”: when the costs of the generation sector collapse, the relative prices of the slower-improving verification, judgment, and evaluation sectors rise, shifting the value bottleneck from “production” to “verification, specification, trust, and accountability.”

Thirdly, efficiency improvements lead to demand expansion. This aligns with the classic Jevons paradox: a decrease in price is the cause of an explosion in consumption, not a guarantee of lower bills. Industry analysis points out that inference models generate a large number of “thought tokens,” Agents repeatedly resend context in loops, and task token consumption can be 3 to 50 times that of chat tasks, with some Agent's single-task cost being 50-60 times that of a chatbot. The improvement in efficiency has, paradoxically, stimulated higher-intensity AI usage, leading to soaring total expenditures. Finally, it must be emphasized that Acemoglu's research indicates a disconnect between micro-level cost collapse and macro-level productivity gains, with AI's boost to total factor productivity over a decade expected to be limited, suggesting that macroeconomic benefits have not yet been fully validated.

Cognitive abilities, as an ownable and reproducible factor of production, are profoundly rewriting traditional economic fundamentals. Firstly, they subvert the Y=F(K,L) production function, pushing “reproducible cognitive capital” to the core, making it a critical input alongside human judgment, competitive physical capital, and data-energy. Technology is no longer an exogenous total factor productivity residual but an internal asset that enterprises can produce, own, depreciate, and trade. Secondly, the definition of marginal cost also undergoes a transformation; it is no longer a fixed number but a dynamic “task cost curve,” its value determined by multiple endogenous variables such as capability tiers, token consumption, processing steps, and crucially, verification and trust costs. The continuous decline in nominal unit price, coupled with the counterintuitive increase in total costs for specific tasks, is a direct manifestation of this complex logic.

This paradigm shift also leads to a fundamental migration of scarcity. As intelligent generation becomes nearly free, the true constraint is no longer the output of information, but rather “what is worth doing, whether it is done correctly, who is responsible, and whether trust can be secured.” The core of scarcity has shifted from “bits” to credibility, judgment, and human attention, which cannot be easily replicated.

Finally, corporate boundaries are also being reshaped by this dual constraint. On one hand, the commodification of cognitive abilities significantly reduces transaction costs, enabling many cognitive activities to be efficiently outsourced, leading to flatter and more streamlined company structures. On the other hand, increasingly high fixed costs for computing power concentrate the most cutting-edge cognitive capabilities in a few super-platforms that command vast resources. This ultimately shapes a “dumbbell” organizational structure – a coexistence of a few giants with numerous highly specialized and trustworthy micro-units, while the middle layer faces accelerated compression.

Facing the new paradigm where “cognitive abilities” become a core factor of production, enterprises must cease building business models solely based on the “unit price” of AI services. A sustainable strategy involves constructing a detailed “task cost curve,” incorporating all endogenous variables such as capability tiers, processing steps, retry rates, and indispensable verification overheads. Any unit economics model based solely on token unit price will quickly become distorted under complex real-world Agent loads. Therefore, a deep understanding and precise measurement of the true cost structure of tasks is the foremost prerequisite for enterprise decision-makers to survive and thrive in the new economic environment.

Concurrently, enterprises should elevate “verification capability” to the strategic status of a core asset, rather than merely treating it as a cost expenditure. When intelligent generation becomes cheap and verification becomes expensive, mechanisms capable of automated evaluation, high-quality regression, strict auditing, and accountability will form new competitive moats. Whoever can first develop verification capability into a reusable asset will gain pricing power and influence in this era of “cost disease.” Furthermore, in a future where intelligence can be easily replicated, true competitive advantage will stem from “irreplicable” capabilities. This means enterprises need to focus their team's energy on the precise specification of ambiguous intentions, complex orchestration of cross-AI capabilities, professional endorsement of final outcomes, and unique taste and insight that cannot be replaced by AI.

Polishing these uniquely human value creation points into core competencies is key to building a moat. Finally, regarding the pace of macroeconomic transformation, we must maintain a clear understanding. While micro-level cost collapse is an undeniable fact, a large-scale productivity revolution remains an unresolved macroeconomic question. For instance, Acemoglu's research suggests that AI's boost to total factor productivity over a decade might be very limited. Therefore, when formulating long-term strategies and cash flow forecasts, it is crucial to remain cautious and conservative, avoiding confusing aggressive fundraising narratives with actual operational expectations.