Source: Tencent Research Institute

Author: Si Xiao, Vice President of Tencent and Dean of Tencent Research Institute

August 26, 2026

In March 2023, during an on-site field assessment of the fur manufacturing cluster in Haining, Zhejiang—conducted for WeChat Channels' emerging e-commerce operations—small and mid-sized enterprise (SME) owners voiced a shared structural headache: new apparel cuts displayed at wholesale stalls were routinely copied by competitors within hours. Historically, the primary moat in this market was proprietary design talent. Yet securing elite fashion designers required prohibitive capital expenditure, effectively pricing smaller operators out of the market.

By SOLOMOAT Editorial Team

When presented with early text-to-image models capable of generating dozens of production-ready design sketches from basic descriptive prompts, these merchants realized that core design capacity could be supplied with near-zero marginal cost. The traditional design talent barrier dissolved overnight. As one merchant observed: "Competition will no longer depend on who can draft a design, but on who knows which design to pick."

That insight captures the core economic transition. When execution is commoditized, strategic discernment becomes the scarce asset. Foundation models deliver raw intellectual output rather than software tooling. Human value shifts entirely toward directing, filtering, and deploying that capacity.

This operational reality led to a foundational thesis presented at the China (Chengdu) International Science Fiction Conference later that year: the structural essence of foundation models is not "Model as a Service" (MaaS), but "Intellect as a Service" (IaaS). While the broader market focused on licensing software wrappers and models, the real shift was underway: intellect itself was being repriced as a standardized factor of production.

A Standardized Yardstick for Intellect

Pricing a new economic asset requires a standardized unit of account. In the artificial intelligence era, that unit is the Token.

Economic history demonstrates that structural market shifts rely on baseline units of measurement: bushels for grain, barrels for crude oil, kilowatt-hours for electricity, and bits for information. While predecessor metrics measured physical mass, energy, or discrete data packets, the Token measures intellectual operations directly: an instance of comprehension, logical deduction, qualitative judgment, or generative synthesis.

Unlike the bit, which liberated information from physical substrates for frictionless distribution, the Token uncouples intellectual capacity from the biological human brain, allowing cognitive work to be executed on demand. An intangible capability has gained a precise commercial meter.

When a query is dispatched to a foundation model, an invisible meter tracks consumption. Generating plain text consumes baseline tokens; rendering high-fidelity video or compiling complex 3D meshes accelerates consumption exponentially; deploying multi-agent swarms to resolve scientific research problems causes compute and token velocity to surge to entirely new orders of magnitude. As task complexity ascends, the higher-order reasoning required is reflected in the token count.

The defining economic characteristic of the Token lies in how its value is realized:

Consuming 1,000 tokens through an API endpoint costs the same whether used for casual dialogue or for auditing cross-border M&A legal contracts—yet the realized economic output diverges by multiple orders of magnitude.

While variable returns on raw materials have long existed—a kilogram of flour can yield standard street bread or a Michelin-starred pastry, and raw paint can produce amateur graffiti or a fine-art masterpiece—those historical value differentials required human craftsmanship. The Token marks the first juncture in economic history where cognitive value-add is generated synthetically at industrial scale.

Additive manufacturing offers a direct analogy. While raw polymers, nylon, and carbon fiber yield vastly different tensile strengths and operational lifespans when printed, different foundation models processing identical token volumes yield vastly different intellectual densities. Standardization does not imply homogeneity.

However, additive manufacturing manipulates physical matter through layer-by-layer extrusion, whereas foundation models ingest discrete tokens to synthesize code, structural designs, and strategic text. Combining the two democratizes the entire supply chain from ideation to physical production. Ultimately, the Token does not meter raw materials, compute hardware, or labor hours; it meters machine-generated intellect supplied at scale.

Financial markets have recognized this reality. Secondary spot and forward arrangements in aggregate token routing and compute arbitrage confirm that the market treats the Token as a fungible, tradable, and allocable resource.

Structural Frictions in an Unregulated Asset Class

Describing the Token economy within legacy frameworks remains challenging due to its cross-disciplinary profile across four distinct verticals:

The Technical Layer: Tokens represent the atomic operating units of tokenization algorithms and transformer architectures. Engineering focuses on inference velocity, context windows, attention mechanisms, and compute efficiency per token.

The Economic Layer: Tokens exhibit standardized production alongside heterogeneous consumption profiles with near-zero marginal switching costs. Downstream utility can only be priced post-consumption, requiring new analytical models to track demand-side valuation.

The Institutional Layer: Cross-border token flows involve compute sovereignty, national data security, localized energy grids, and international industrial competitiveness. Because token traffic operates outside traditional customs jurisdictions, central bank capital controls, and territorial data boundaries, legacy trade treaties and cross-border tax regimes remain ill-equipped to govern it.

The Societal Layer: While AI broadens access to advanced capabilities—enabling non-specialists to write software, render media, and design physical products—it introduces a structural divide. Strategic performance is dictated by an operator’s ability to orchestrate advanced model tiers and fund enterprise token budgets.

Addressing this divide depends on Token Efficiency: the net economic value generated per unit of token consumption. The enterprise AI race has transitioned from functional validation (capability) to unit economics (return on invested compute). Enterprise architectures cannot rely exclusively on flagship frontier models; they require a tiered allocation model:

AI transitions into a dependable macroeconomic driver only when enterprise value creation exceeds the cost of compute.

Four Structural Theses on the Token Economy

Empirical research by Tencent Research Institute yields four primary conclusions:

The Core Economic Scissors Divergence: Moore’s Law in Intelligence Costs Meets Jevons’ Paradox in Enterprise Demand. Algorithmic distillation, silicon specialization, and infrastructure optimization drive the unit cost of equivalent intelligence down along a Moore’s Law trajectory. Lower prices unlock previously unviable unit economics, expanding token invocation volumes and gross spending. The spread between falling unit costs and surging gross demand defines enterprise profit pools.

Exponential Token Growth Is Driven by Autonomous Long-Horizon Task Execution. According to China's National Data Administration data from March 2026, nationwide daily token calls expanded from 100 billion in early 2024 to 140 trillion by March 2026—a nearly 1,400-fold increase over twenty-four months. This acceleration stems from higher success rates in long-horizon autonomous tasks that operate without human intervention, removing the human attention bottleneck. Similar to autonomous vehicle fleets increasing total vehicle-miles traveled as driver intervention drops, AI systems executing multi-step planning, automated tool integration, code execution, and self-validation drive token consumption into higher volume regimes.

Tiered Pricing Institutionalizes a Clear Market for Strategic Judgment. Divergent token pricing across model tiers reflects the commercial valuation of machine reasoning capability. Just as hourly billing rates for junior legal associates differ sharply from senior partners based on the quality of judgment provided, tiered token pricing establishes a transparent market mechanism for cognitive skill levels.

Institutional Classification Lags Cross-Border Flow Realities. Cross-border token transactions transcend basic API queries, manifesting as exported model weights, compute capacity leasing, or packaged autonomous agent services. Global regulatory frameworks face significant gaps regarding cross-border data compliance, tax jurisdiction, dispute resolution, and operational liability. These shortfalls persist because legacy legal taxonomies have no category for an asset that fuses compute capacity, algorithms, electrical power, and automated business workflows.

The Primacy of "Intellect" Over Tooling

Focusing on "intellect" rather than secondary concepts like compute infrastructure or model wrappers stems from a background in intellectual property law. The core subject of IP protection is defined as the product of human intellect. Legal systems draw a clear line between the production process and the final deliverable; regulatory and commercial frameworks protect and trade the intellectual output itself, indifferent to the mechanics that generated it.

Generative models function as engines for the industrial production of intellectual output. The resulting analyses, codebases, design architectures, and strategies substitute directly for biological intellectual labor. While synthetic matrix multiplication differs mechanically from biological neural activity, the commercial output is identical: functional intellectual work.

This dynamic underpins the shift from "AI as a Service" to "Intellect as a Service." The "Artificial" label is secondary. Future high-performance systems may diverge from transformer architectures or current deep learning models entirely; the enduring economic reality is the on-demand delivery of human- and superhuman-level intellect via cloud infrastructure. Model architectures are temporary middleware; the deliverable is intellect.

The term Intellect is deliberate:

Cognition describes an underlying process.

Thinking describes an operational action.

Intellect defines the core capability to solve complex problems and capture real economic value.

Enterprise buyers do not procure machine cognition or processing steps; they procure the capability to execute tasks to a defined standard. The Token serves as the baseline accounting metric for that capability.

The Ultimate Value Anchor

When expressive drafting and code compilation become on-demand computational commodities, the definition of authorship shifts away from manual execution toward the strategic ownership of core judgment, problem framing, and directional choices.

In 2017, Tencent Research Institute introduced the principle of "Tech for Good," formalizing its foundational AI thesis in 2024: Humanity is the measure of AI; AI is the extension of humanity.

The ultimate assessment of the Token economy remains simple: whether it expands access to advanced intellectual capacity, optimizes productivity, and elevates human agency. That inquiry remains the defining strategic benchmark for the ecosystem going forward.

Frequently Asked Questions

What does a token measure economically?

A token meters a unit of model-mediated intellectual processing, while the economic value of that processing depends on context, model quality, and deployment.

Where does human advantage remain?

Humans retain advantage in problem selection, taste, judgment, accountability, and converting inexpensive machine intelligence into valuable market outcomes.