The Rise of China’s Hottest New Commodity: AI Tokens

Imagine a new global commodity traded not in barrels or bushels, but in trillions of invisible computational units — weightless, borderless, and already reshaping the architecture of economic power. In the summer of 1858, a copper-core cable crossed the Atlantic seabed and rewired who controlled the flow of value across empires. In the spring of 2026, something structurally similar is happening, only the cable is digital, the commodity is China’s AI tokens, and the empire building is happening in plain sight.

The numbers are now difficult to ignore. China’s daily consumption of tokens — the tiny data units processed by AI models — has surpassed 140 trillion as of March 2026, a more than 1,000-fold increase from the 100 billion recorded at the beginning of 2024, and over 40 percent higher than the 100 trillion logged at the end of last year. China.org.cn Liu Liehong, administrator of China’s National Data Administration, announced the figure publicly and framed it not as a technical milestone but as a strategic one. The surge, he said, signals China’s AI industry “evolving from basic chat functions to more sophisticated systems capable of decision-making and task execution.” This is bureaucratic language with a geopolitical subtext: China is no longer catching up in artificial intelligence. It is setting the pace in the metric that matters most — actual usage, at scale, in the real economy.

From OpenRouter to the World: How China’s AI Tokens Surpassed the US

The clearest empirical signal of this shift has come from an unexpected source: OpenRouter, a San Francisco-based API aggregation platform that functions as a kind of global stock exchange for large language models. OpenRouter data published on February 24, 2026, shows that models built in China account for 61% of total token consumption among the platform’s top ten most-used models, with aggregate consumption reaching 5.3 trillion tokens out of a combined 8.7 trillion. Dataconomy The three most-consumed models that week were all Chinese. MiniMax M2.5 claimed the top position with 2.45 trillion tokens consumed in a single week — a 197% increase from the prior week. Moonshot AI’s Kimi K2.5 followed with 1.21 trillion tokens, and Zhipu AI’s GLM-5 placed third with 780 billion tokens, itself up 158%. TechBriefly

The historical reversal was swift and decisive. In the first week of February 2026, the weekly call volume of Chinese models had jumped to 2.27 trillion tokens, sending a strong signal of pursuit. Just one week later, Chinese models officially surpassed their US counterparts with 4.12 trillion tokens versus 2.94 trillion. By the week of February 16th, Chinese models had soared to 5.16 trillion tokens — a 127% increase in three weeks. 36Kr The growth is structural, not episodic, and it has been observed at the highest levels of the American venture capital industry. Andreessen Horowitz partner Martin Casado estimated that roughly 80% of startups using open-source AI stacks are running Chinese models. TechBriefly OpenRouter COO Chris Clark put the dynamic plainly: Chinese open-weight models have gained large market share because they are “disproportionately heavy in agentic flows run by U.S. firms.”

Ciyuan: When a Nation Brands Its Commodity

Beijing has never been content to let economic transformations arrive without a conceptual framework to accompany them. At the 2026 China Development Forum, Liu Liehong used the term ciyuan as the official Chinese translation for “token” during a speech on AI development, effectively resolving a debate within China over how the term should be rendered. South China Morning Post The naming is deliberate and worth examining. In Chinese, ci translates to “word,” while yuan carries double meaning: it is the basic unit of Chinese currency, and the suffix used when naming most foreign currencies in Mandarin. Liu said the token, or ciyuan, was not only a value anchor for the intelligent era but also a “settlement unit” linking technological supply with commercial demand, thereby allowing business models to be quantified. South China Morning Post

The People’s Daily had introduced the concept in January, describing ciyuans as the smallest unit of information processed by large models — possessing characteristics “emergent in the intelligent era” of being quantifiable, priceable, and tradable, with a new value system centered on their invocation, distribution, and settlement rapidly taking shape. TechFlow The semantic move is not accidental. China is not simply producing more AI tokens than the United States. It is trying to name, define, and ultimately govern the unit of account for the next phase of the global technology economy. Jensen Huang arrived at the same conceptual destination independently. At Nvidia’s GTC developer conference last week in San Jose, clad in his trademark leather jacket, Huang told the audience that “tokens are the new commodity,” declaring that Nvidia should no longer be seen mainly as a chip maker but as a builder of what he calls “AI factories” that produce tokens in large numbers. South China Morning Post Two of the world’s most consequential technology figures, one American and one Chinese, are now converging on the same metaphor — which suggests the metaphor is correct.

The Structural Edge: Electricity, Architecture, and the Token Economy

China’s dominance in China’s AI tokens is not a speculative narrative driven by state media hype or a single viral product launch. It rests on compounding structural advantages that are difficult to reverse quickly through policy alone.

The most fundamental is energy. China’s total electricity costs are approximately 40% lower than in the United States — a physical cost advantage that competitors cannot easily replicate. China Academy When a developer anywhere in the world calls a Chinese AI model’s API, the request is processed in a Chinese data center powered by the Chinese grid. The economic value of that electricity is exported globally as a high-margin digital service — one that bypasses customs, evades tariffs, and barely registers in conventional trade statistics. Industry estimates suggest that converting raw electricity into AI processing services can increase its value by up to 22 times compared to simply exporting electricity at the grid rate. China.org.cn China’s western regions — Xinjiang, Inner Mongolia, Yunnan — provide abundant, low-cost renewable energy at scale. The country has also built a vertically integrated supply chain spanning ultra-high-voltage transmission equipment, liquid-cooled data centers, and server assembly that few rivals can match.

The second advantage is architectural. Chinese AI laboratories have pioneered efficiency-first model design under the pressure of US chip export restrictions. DeepSeek V3’s Mixture-of-Experts architecture activates only a fraction of the model’s parameters during inference, with independent tests showing its inference cost is roughly 36 times lower than GPT-4o. MiniMax M2.5, despite having 229 billion total parameters, activates only 10 billion during inference. China Academy These are not merely clever engineering choices. They are the product of operating under genuine resource constraints — constraints that have paradoxically made Chinese models leaner, cheaper, and more deployable at global scale.

The third advantage is price. MiniMax M2.5 charges $0.30 per million input tokens and $1.10 per million output tokens. By comparison, Claude Opus 4.6 costs $5 per million input tokens and $25 per million output tokens — roughly 10 to 20 times more expensive. TechBriefly In the new agentic AI era, where a single automated workflow can consume millions of tokens in a matter of hours, this price differential is not a marginal consideration. It is frequently the deciding factor. A Silicon Valley developer who once tested workflows with GPT-4 at tens of dollars a day has little rational reason not to switch when a Chinese alternative delivers comparable benchmark performance at a tenth of the cost.

Alibaba Token Hub and the Industrialization of Ciyuan

Corporate China has received the signal and reorganized accordingly. Alibaba has established a new internal division called the Alibaba Token Hub, directly overseen by Chief Executive Eddie Wu, moving the research team that develops its flagship Qwen models, the consumer-facing app division, and major AI-related products under a single unified structure. Bloomberg The unit will focus on creating, distributing, and applying tokens — the basic computing units used by AI models — while integrating several internal teams to cover the full AI stack, from foundation model development to enterprise-level AI applications. TechNode The naming of the division after the commodity it produces is itself a statement of intent. Alibaba is not building an AI company. It is building a token factory.

The reorganization lands against a backdrop of surging Chinese AI cloud pricing that reflects genuine demand pressure. Alibaba Cloud announced price increases on select services effective April 18, 2026, citing global AI demand, rising supply-chain costs, and sharp increases in token call volume. Baidu Smart Cloud made an identical announcement the same day. Zhipu launched a new agent-optimized model and simultaneously raised its API price by 20% on March 16th. Tencent Cloud adjusted billing strategies for its intelligent agent development platform starting March 13th. 36Kr When Chinese AI providers raise prices in unison, it is not a cartel behavior — it is a market clearing mechanism. The supply of ciyuans is being consumed faster than it can be provisioned, and the price signal is propagating through the ecosystem.

A report jointly released by Andreessen Horowitz and OpenRouter shows that the total token call volume of Alibaba’s Qwen series ranks second globally at 5.59 trillion, second only to DeepSeek’s 14.37 trillion. 36Kr These are not vanity metrics: they represent real developer adoption, real API revenue, and real geopolitical influence embedded in the codebases of companies that may scale into tomorrow’s global technology infrastructure.

The Counterpoints: Profitability, Chip Constraints, and Sovereign Risk

Honest analysis demands acknowledgment of what the token volume data does not tell us. Market share on OpenRouter — a platform beloved by independent developers and AI hobbyists rather than large enterprise procurement departments — does not translate automatically into enterprise dominance. The main battleground for corporate AI workloads remains, for now, in the hands of American providers offering the accountability, compliance tooling, and integration depth that large institutions require. OpenRouter represents a thin slice of the global AI market; its developer-skewed demographics mean the 61% figure overstates Chinese penetration of the full economy.

The profitability question is equally live. Aggressive token pricing is partly a land-grab strategy — buying market share at margins that may not be sustainable. The simultaneous wave of Chinese cloud price increases in March 2026 suggests the economics are tightening. DeepSeek’s inference costs may be radically lower than GPT-4o’s, but training costs, talent costs, and the escalating expense of acquiring increasingly scarce advanced chips under US export restrictions are real. Washington’s ongoing efforts to tighten the chip embargo — extending restrictions to additional Nvidia architectures and closing loopholes used to route chips through third-country entities — represent a genuine long-run constraint on China’s ability to scale inference capacity. And sovereign risk is not zero. Developers in regulated industries and allied governments face real legal and reputational exposure from routing sensitive workloads through Chinese infrastructure, regardless of how cheap or fast those tokens may be.

Token Exports as a New Form of Digital Soft Power

Yet the strategic logic of China’s position is more durable than its critics typically concede. Tokens are intangible, bypass customs, evade tariffs, and don’t appear in official trade statistics. China exports massive compute and electricity services, yet it remains virtually invisible in trade data. China Academy This invisibility is a feature, not a bug. Token exports occupy a legal and regulatory grey zone that trade hawks find difficult to target. You cannot sanction a token. You cannot put a tariff on an API call. The infrastructure that produces the tokens — the data centers, the power grid, the model weights — sits firmly within Chinese sovereignty and beyond the reach of extraterritorial enforcement.

Beijing appears to understand this clearly. China has named 2026 the “Year of Data Element Value Release,” is building a single national data market with unified property rights, and by end of 2025 had compiled over 100,000 high-quality datasets totaling more than 890 petabytes — roughly 310 times the digital collection of the National Library of China. MEXC The scale of data assembly, combined with cheap inference, low-cost energy, and rapid model iteration cycles, constitutes a vertically integrated token economy that took China’s industrial sector decades to assemble in steel or semiconductors — and that is being assembled in AI in a matter of years.

Chinese artificial intelligence service stocks rallied this week after state media highlighted a sharp increase in domestic AI model adoption and a surge in the token usage they generate. Bloomberg The market’s reaction is rational. Investors are pricing in what economists have been slow to formally model: that the token, like oil before it, will become a commodity whose production geography matters enormously to the distribution of global wealth. The country that most cheaply produces what the world most needs will, history suggests, extract durable rents. In the oil era, that was the Persian Gulf. In the token era, the early evidence points unmistakably toward the Yangtze River Delta, the Pearl River Delta, and the data centers of Guizhou province humming with renewable hydropower.

The British Empire laid the cables. The rest, as they say, was history. The question now is who controls the flow — and at what price per million tokens.