Will AI tokens become the new global commodities and currencies?

Source: Digital Economy Technology

Written by: Fan Wenzhong

On March 23, the Director of the National Data Bureau, Liu Liehong, announced a startling set of data at the China Development High-Level Forum: China’s daily AI Token call volume has surged from 100 billion at the start of 2024 to 100 quadrillion by the end of 2025, and surpassed 140 quadrillion in March 2026, representing over a thousandfold growth in just two years. Meanwhile, data from OpenRouter, the world’s largest AI model API aggregation platform, shows that China’s large model weekly call volume has consecutively exceeded that of the United States for several weeks, with the top three global call volumes all dominated by Chinese models. An industry revolution driven by Tokens is rapidly reshaping the global technological competition, business models, and even national core competitiveness at an unprecedented speed.

In early 2026, multiple industry developments in Silicon Valley also drew worldwide attention. OpenAI is gradually abandoning the core internet metric DAU (Daily Active Users), used for nearly 20 years, and shifting focus to TPD (Tokens Per Day) as the primary operational indicator. This shift is no coincidence. At GTC 2026, NVIDIA CEO Jensen Huang redefined data centers as “Token factories,” emphasizing that future competition will revolve around “Tokens per Watt.” This is not an isolated phenomenon but signals the arrival of a new paradigm of intelligent economy centered on Tokens, fundamentally transforming how value is measured and traded.

1. The Value and Measurement of AI Tokens

2. AI Tokens as a Value Standard in the Intelligent Era

From a computer science perspective, a Token is the fundamental unit through which AI models process various types of information. When a piece of text is input into a model, it is broken down into words or subwords; an image is divided into pixel patches; an audio clip is segmented into time slices. These indivisible basic units are all called Tokens.

In practical applications, Token measurement follows certain rules. For English text, a short word may count as one Token, while longer words are split into multiple Tokens; a common rule of thumb is approximately 1 Token equals 4 English characters. For Chinese text, typically one Chinese character corresponds to 1 or 2 Tokens. Whether during data processing in model training or during functional output in model service calls, every core action of AI is measured in Tokens. The scale of Token consumption directly reflects the workload and value output of the model, aligning with Marx’s labor theory of value.

Tokens are especially significant because they provide a quantifiable and comparable measure of value for the development of the intelligent economy. As AI technology evolves from text modalities to multimodal applications—covering programming, video, scientific research, and more—the strategic positioning of Tokens as a “unified measurement” becomes increasingly prominent. This is not arbitrary but an inevitable result of industry development: the industrial age used “kilowatt-hours” to measure electricity consumption; the internet age used “GB” to measure data traffic; and the AI era naturally requires Tokens to measure intelligent output. Economically and commercially, Tokens have become the core units of value that can be measured, priced, and traded in the intelligent age. They connect underlying energy, computing power, and data with top-level intelligent services, serving as a universal metric for AI productivity, cost accounting, and service settlement.

The value chain of Tokens encompasses hardware manufacturing, infrastructure construction, computing power provision, platform operation, and application development. In its cost structure, electricity and depreciation of hardware account for 70-80%, making them the key factors influencing international competitiveness. “Tokens per Watt” (Tokens per Watt) has become a core indicator of AI enterprise competitiveness. This means that under a fixed energy budget, those who can produce more Tokens with higher energy efficiency will have lower production costs and stronger market competitiveness.

Factors Affecting AI Token Measurement

As application scenarios become increasingly diverse, Token measurement methods have evolved from simple counts to complex, multi-dimensional, dynamically weighted systems.

(1) Input and output differentiation. The most basic measurement still follows a binary structure of “Input Tokens” and “Output Tokens.” Input Tokens represent the information provided by users to the model (including prompts, uploaded documents, historical dialogues, etc.), while Output Tokens are the responses generated by the model. In commercial billing, because generation consumes significant memory bandwidth and computational cycles, the cost of Output Tokens is usually 3 to 5 times that of Input Tokens. This price difference reflects the fundamental distinction between “creative labor” and “information reading” in computational resource consumption.

(2) Context measurement and memory costs. From 2024 to 2025, large models’ context windows expanded from 8K and 32K to 128K and even 1 million tokens. By 2026, handling ultra-long contexts has become routine. However, long contexts are not free. The attention mechanism based on Transformer architecture causes the computational complexity of processing long sequences to grow quadratically or linearly. Therefore, modern measurement systems introduce “context weighting coefficients.” When a user asks questions within a session with a 1 million Token context, even generating only 10 Tokens in response requires scanning or retrieving extensive historical memory, which is implicitly included in the “Active Context Tokens” cost. This makes measurement more precise in reflecting the resource costs of maintaining long-term memory.

(3) Tokenization of multimodal data. As multimodal large models (LMMs) mature, images, videos, and audio are also incorporated into the Token measurement system. A high-resolution image is no longer viewed as a single file but is sliced into hundreds of visual patches, each encoded as one or more visual Tokens. A one-minute video may convert into tens of thousands of visual Tokens. This unified measurement approach breaks down modality barriers, enabling tasks like image captioning, video understanding, and speech interaction to be calculated under the same economic model. For example, generating a 10-second high-definition video may consume a number of Tokens equivalent to writing a thousand-word article, visually illustrating the differences in information density across modalities.

(4) The implicit valuation of Tokens. With the proliferation of AI Agents, models no longer just produce single responses but engage in complex autonomous planning, code execution, self-reflection, and multi-round searches. These processes generate大量中间思维 Token，这些 Token 并不直接展示给用户，但却是高质量输出的基础。新的计量标准开始区分“表面输出 Token”和“内部推理 Token”。对于高难度的科学计算或复杂逻辑推理，内部推理 Token 的数量可能是最终输出的几十倍。一些先进平台已开始尝试按有效推理步数或思维链深度进行差异化计费，标志着计量体系从“计字数”向“计智力”的根本转变。

AI Token的发展趋势

近年来，AI Token的发展呈现出三大核心趋势：总量的指数级爆发、单位的极致压缩以及价值的分层固化。

趋势一：消耗量的核爆式增长。据统计，2024年全球日均Token消耗量约为1000亿级别，而到2026年第一季度，这一数字已飙升至180万亿，增长近1800倍。这一增长不是线性叠加，而源于应用范式的质变。早期的Token消耗主要来自人机对话（Chatbot），属于低频、浅层交互；而2026年的主流应用是自主智能体（Autonomous Agents）。一个Agent在执行任务时，会自主拆解目标、调用工具、编写并调试代码、验证结果，这一闭环过程可能产生数万甚至数十万Token的消耗。未来，随着具身智能（Embodied AI）的落地，机器人每秒的感知与决策都将转化为海量的实时Token流，预计2030年全球日均Token消耗将达到京（10^16）级别。

趋势二：单位成本的摩尔定律式下降。得益于硬件架构的迭代（如NVIDIA Blackwell及后续Rubin架构的量产）、软件算法的优化（如混合专家模型MoE、量化技术、投机采样）以及集群调度效率的提升，2026年生成一个高质量Token的算力成本相比2023年下降了约两个数量级。这种“杰文斯悖论”效应在AI领域表现得淋漓尽致：效率的提升并未减少总资源消耗，反而激发了前所未有的需求。未来，随着光子计算、神经形态芯片等颠覆性技术的引入，单位Token的能耗有望进一步降低，使得“无限智能”在理论上成为可能。

趋势三：价值分层与专用化。未来的Token市场将出现明显的“价值分层”。通用大模型产生的“标准Token”将像电力一样廉价且同质化，主要用于日常问答、基础翻译和简单分类；而经过垂直领域微调（Fine-tuning）、拥有独家私有数据加持、具备深度推理能力的“高阶Token”将变得昂贵且稀缺。例如，由顶级医疗模型生成的诊断建议Token，其价值远高于普通聊天机器人的闲聊Token。这种分层将催生“Token期货市场”和“质量认证体系”，用户将为特定质量等级（Quality-of-Service, QoS）的Token支付溢价。

2. The AI Token Industry Comparison Between China and the US

3. Production and Consumption Scale: China Surpasses the US in Total Volume

The US’s core advantages in AI lie in chip design and model capabilities. NVIDIA, as the dominant GPU supplier globally, saw its market value soar from about $300 billion at the end of 2022 to over $4 trillion today, a 14-fold increase. Behind this is the US’s sustained leadership in advanced process chip design. Meanwhile, closed-source models like Claude and GPT are still considered the most capable, maintaining high prices above $5 per million tokens. This pricing strategy reflects both the US’s technological lead and its pricing power in high-end markets.

However, the US’s dominance faces structural challenges. On one hand, grid bottlenecks are constraining further expansion of AI compute power, with high electricity costs. On the other hand, dense model architectures lead to low utilization of compute resources, making unit-token production costs slow to decline.

In contrast, China’s competitive advantages mainly lie in cost control and open-source ecosystems. Chinese models like DeepSeek have driven prices down to $0.028 per million tokens, just 1/180 of GPT’s price. This extreme cost-performance ratio is attracting global developers—during the week of February 16-22, 2026, Chinese models on the OpenRouter platform consumed 51.6 trillion tokens, a 127% increase over three weeks prior, while US models only consumed 27 trillion tokens and continued to decline. Among the top five models worldwide, four are Chinese, accounting for 85.7% of the top five. In February 2026, China’s weekly call volume first surpassed that of the US and has maintained the lead, with models like MiniMax, DeepSeek, and Kimi consistently ranking at the top. China’s share of global token consumption once exceeded 60%.

It’s important to note that China’s overtaking in token consumption mainly occurred on the inference side, not during training. Inference requires less single-card performance, and domestically produced chips with deep optimization can support massive inference demands. Training still relies on a small number of high-end cards, requiring distributed architectures and MoE techniques to produce high-quality models. This structural characteristic means China has significant advantages in AI application deployment and value realization, but still has room to catch up in the foundational model innovation.

2. China’s Energy and Engineering Cost Advantages

China’s cost advantages stem from multiple synergistic factors. Electricity costs are fundamental, typically accounting for over 30% of total compute costs. Since AI training and inference are energy-intensive, a country’s grid stability and green electricity prices directly influence its token production competitiveness. Through the “East Data, West Computation” project and a unified large grid, western China’s green electricity prices can drop to as low as 0.2 yuan per kWh (about $0.028), compared to 0.08-0.12 USD per kWh in Europe and America.

Chip costs include hardware procurement, depreciation, and maintenance. The US, leveraging NVIDIA’s leading position, has advantages in high-end chip supply, but this also means higher procurement costs. China’s strategy is to rely on a small number of high-end chips during training, while extensively using domestically produced chips during inference, optimizing to minimize per-unit compute costs. At the full-stack level, Chinese manufacturers achieve deep integration of models, cloud services, and chips, maximizing compute utilization, whereas US firms often depend on third-party cloud and chip providers, incurring higher adaptation costs.

Engineering efficiency is a key variable determining token costs. On the technical front, Chinese firms widely adopt MoE (Mixture of Experts) architectures—splitting large models into multiple specialized experts, activating only relevant ones for each task. With the same $1,000 investment in compute, different technical approaches can produce over ten times more tokens. Compared to dense models, MoE architectures can significantly boost tokens per unit of compute. Full-stack optimization is equally crucial—when model developers, cloud providers, and chip designers collaborate deeply, compute utilization often exceeds expectations.

Global AI competition has shifted from a pure “model performance race” to a comprehensive national strength contest centered on “Token production efficiency” and “cost per token.” China’s low-cost, stable energy supply, vast unified market, and efficient deployment capabilities have established a huge advantage in large-scale, low-cost token production, making it a “cost-effective hub” and “mass production factory” for AI compute. The US relies on technological innovation, high-end ecosystems, and financial capital, occupying the high-value segments of the supply chain. The core of this competition involves energy pricing, industrial organization, and digital ecosystem influence—an all-encompassing contest. Soon, China may convert its domestic energy and electricity advantages into international trade advantages, creating a new highly competitive product—AI tokens. In this rapidly growing field, China runs a trade surplus with all countries except the US, which will reshape the global economic and strategic landscape.

3. Will AI Tokens Become a New Global Currency Asset?

The Gap Between Currency Necessity and Reality

To explore whether AI Tokens can become a globally circulating currency, we must first clarify the essential attributes of money. Economics states that an asset must fulfill three core functions to qualify as currency: a measure of value, a medium of exchange, and a store of value. Additionally, it must be widely accepted, stable in value, and backed by sovereign credit. Comparing these standards, AI Tokens are unlikely to become true currency in the foreseeable future.

The biggest obstacle is volatility. Over the past two years, the price of a single Token has plummeted over 99%. Such drastic fluctuations mean merchants are unlikely to accept a “currency” that could halve in value within a week. Even if prices stabilize later, the value of AI Tokens remains highly tied to the cost of compute power, which itself is affected by chip technology updates, energy prices, geopolitical conflicts, and other factors—making long-term stability difficult.

Limited acceptance is another critical constraint. Currently, AI Tokens are only accepted for API calls and AI applications, not for purchasing everyday goods and services. Money’s fundamental role is as a universal exchange medium for society’s commodities, but AI Tokens’ network is confined to AI services. To achieve widespread acceptance, a global transaction network covering goods and services would need to be built—a massive infrastructure project requiring long-term market development.

Compared to becoming a currency, AI Tokens are more likely to evolve into a new type of commodity asset, similar to oil, gold, or copper. This judgment is based on several observations:

First, AI Tokens possess the core features of commodities. They are standardized, tradable, and in broad demand, fitting the characteristics of commodities. Jensen Huang explicitly stated, “Future data centers will become nonstop, roaring factories, producing not traditional products but the most valuable commodity in the digital world: Tokens.” Just as the industrial age relied on oil as fuel, the intelligent age will rely on Tokens as “smart fuel.”

Second, the pricing mechanism of Tokens is increasingly similar to commodities. Currently, AI model API prices are market-driven: prices rise when supply is tight, fall when demand weakens. This mechanism closely resembles traditional commodity markets. As Token trading becomes more standardized and scaled, derivative markets like Token futures and options may emerge, providing risk management tools for producers, consumers, and investors.

Third, the supply-demand structure of Tokens exhibits typical commodity features. Supply is constrained by chip capacity, energy supply, and other physical limits, with long expansion cycles and limited flexibility. Demand, driven by AI application proliferation, grows rapidly and exhibits cyclical patterns. This supply-demand dynamic causes Token prices to fluctuate periodically rather than decline smoothly. The early 2026 price surge confirms this: despite a long-term downward trend, short-term imbalances can trigger sharp price spikes.

Fourth, Tokens are increasingly considered as strategic reserves by nations. As AI capabilities penetrate defense, finance, energy, and other critical sectors, compute power security rises to a national security level. Some countries may begin to strategically stockpile compute resources, with Tokens serving as the measurement unit for compute reserves. This trend could lead to a “compute-based monetary system”—a new reserve system anchored to compute power.

2. Stablecoins as a New Solution for AI Tokens

Given that AI Tokens are unlikely to become a currency, a promising development is the rise of stablecoins as an innovative monetary form within the AI Agent economy. When AI Agents need to make autonomous decisions and transactions, traditional financial systems reveal many incompatibilities: banks do not open accounts for AI, credit cards are not designed for algorithms, and credit systems are built for humans. For AI, money is not wealth but an interface; not a store of value but a pathway for execution logic. Against this backdrop, blockchain-based stablecoins offer unique advantages—permissionless global transactions, instant settlement, and low-cost collaboration—perfectly aligning with AI Agent economic needs.

Data shows that stablecoins are rapidly growing within the AI Agent economy. By March 2026, the total transaction count on the x402 ecosystem exceeded 163 million, with total transaction volume surpassing $45 million. The number of buyer AI Agents exceeded 435,000, and seller AI Agents over 90,000. Among them, USDC dominates the x402 protocol layer, accounting for 98.6% of transaction volume on EVM chains and 99.7% on Solana.

3. Three Possible Future Paths for AI Token Evolution

Based on the above analysis, the future of AI Tokens may follow three main trajectories:

Path 1: Maintain their role as measurement units, not becoming independent assets. In this scenario, AI Tokens remain as the unit of valuation for AI services but do not acquire independent asset attributes. Users pay for AI capabilities, not for the Token itself; Tokens are merely a billing method, not an investment target. This is the most conservative forecast and reflects the current situation.

Path 2: Evolve into commodities, forming a compute futures market. As trading volume and standardization increase, Tokens could become tradable commodities like oil or copper. Exchanges might launch Token futures and options, providing price discovery and risk management tools. Under this path, Token prices could become more volatile but also more financialized.

Path 3: Serve as a measurement standard for a compute-based monetary system. This is the most revolutionary path: compute power becomes the monetary anchor, similar to gold in the gold standard era. In this system, sovereign digital currencies (CBDCs) would be backed by compute power, with each unit of currency corresponding to a standardized amount of Tokens. Although technically and institutionally challenging, if realized, this would fundamentally reshape the global monetary system.

4. Strategies for the AI Token Era

National Strategies: Strengthening Compute Sovereignty and Strategic Infrastructure

In response to the rise of the Token economy, countries need to incorporate compute resources into strategic infrastructure planning and proactively address governance issues. Specific measures include:

• Building a compute infrastructure system. Drawing on the successful experience of the “East Data, West Computation” project, plan nationwide compute networks to optimize resource allocation. This includes establishing large intelligent compute centers in energy-rich western regions utilizing green electricity, constructing edge computing nodes in demand-dense eastern areas to ensure low-latency services, and creating a unified national compute scheduling platform for on-demand and elastic resource allocation.

• Promoting standardization of Token measurement. Currently, various platforms use different measurement methods, complicating developer choices and enterprise cost accounting, and constraining large-scale development of the Token economy. The government should guide industry associations and leading enterprises to develop unified standards, clarify conversion rules for different modalities (text, images, audio), and establish transparent, fair cost accounting mechanisms. This will facilitate efficient domestic markets and enhance China’s influence in the global Token economy.

• Improving governance frameworks for the Token economy. Rapid development brings new regulatory challenges: how to define the legal nature of Tokens (service measurement units, digital assets, or securities)? How to supervise cross-border Token transactions? How to prevent financial risks from price volatility? How to balance user protection and innovation? Addressing these questions requires close collaboration among policymakers, technical experts, industry players, and academia to develop governance systems suited to the characteristics of the smart economy.

• Participating actively in international rule-making. China should engage in shaping global AI governance rules, including: promoting international standards for Token measurement within multilateral frameworks; incorporating compute cooperation clauses into bilateral trade agreements; proposing fair taxation schemes for Token transactions in digital tax negotiations. Gaining influence in rule-setting will help China secure a strategic advantage in the future global Token economy.

Corporate Strategies: Rebuilding Token Efficiency Mindset and Business Models

• Establish a Token efficiency mindset. When selecting AI technologies, companies should prioritize Token efficiency as a core evaluation criterion, focusing on the match between compute resources and Token consumption. From prompt design and model invocation to result optimization, every step should balance efficiency and cost. Precise prompt engineering can reduce unnecessary Token use; optimized model invocation can improve compute utilization. These details directly impact overall AI investment costs. Borrowing from the communications industry’s “good-put” concept, companies should focus on “how many Tokens truly advance user goals,” rather than simply maximizing throughput. The core of this shift is moving from “how much compute is used” to “how much value is created.”

• Reconstruct business models and pricing strategies. The industry is shifting from “traffic subsidies” to “value-based pricing.” Early low prices attracted many trial users, leading to inefficient resource use—some estimates show 40% of free quota consumption was for non-business testing. By gradually raising prices, companies can filter out non-essential demand and ensure stable service for high-value customers. This “price-to-value” approach marks a move from internet-scale expansion to software industry value pricing.

• Develop new talent standards and incentive mechanisms. Jensen Huang proposed at GTC 2026 to allocate Token budgets to engineers, worth about half their annual salary, as a talent attraction strategy. He explicitly stated, “If you hire a software engineer earning $500,000 annually, and they do not consume at least $250,000 worth of Tokens, I would be concerned.”

Personal Strategies: Cultivating Token Literacy and New Human-AI Collaboration Skills

• Build Token literacy. Most users lack sufficient understanding of Token consumption, model capabilities, and pricing mechanisms, leading to issues like stock trading via AI agents resulting in account zeroing overnight, or instructing multiple AI agents to surrender API access, causing “tricked” agents. These cases highlight that Token literacy is becoming a fundamental skill in the digital age.

• Develop new human-AI collaboration workflows. Huang Huang predicted that in the future, computers will operate 24/7, continuously generating Tokens as AI agents execute tasks tirelessly. This shift requires individuals to move from “doing it themselves” to “command AI to do it,” and from “executor” to “supervisor.”

• Embrace lifelong learning and skill iteration. The rapid growth of the Token economy shortens skill half-lives. Today’s popular models will soon be replaced by more efficient architectures; current hot models will be surpassed by advanced designs. In this environment, maintaining learning and adaptation capabilities is more important than mastering specific skills. Individuals should develop continuous learning habits, stay updated on AI and Token economy developments, proactively experiment with new tools and methods, and build interdisciplinary knowledge to understand the economic logic and social impact behind the technology. Only through such efforts can one remain resilient amid the Token wave.