Beijing didn’t just build a better AI. It built a new unit of economic power — and gave it a name.
The world has spent two years arguing about who makes the smartest AI model. China quietly asked a different question: who controls the raw unit of value that AI runs on?
At the 2026 China Development Forum this week, Liu Liehong, head of China’s National Data Administration, officially designated ciyuan as the Mandarin word for “token” — the fundamental computational unit that powers every AI interaction, from a ChatGPT reply to a DeepSeek research run. The naming wasn’t a translation exercise. It was a declaration.
In Mandarin, ci means “word.” Yuan means currency. Beijing essentially stamped the word “money” onto the atom of the AI economy.
Liu didn’t stop at branding. He called ciyuan a “settlement unit” — a term lifted straight from financial infrastructure — linking AI supply with commercial demand and making business models quantifiable. A new value system, he said, is taking shape around the invocation, distribution, and settlement of tokens. That’s central bank language applied to machine learning. The intention is unmistakable.
The Numbers That Make This Real
This isn’t state-media theater. The underlying data is staggering.
China’s daily AI token consumption hit 140 trillion as of March 2026. Fourteen months ago, that number sat at 100 billion. That’s a more than 1,000-fold increase — not growth, but a structural transformation of an entirely different order.
On OpenRouter, the San Francisco-based platform that functions as a global marketplace for AI models, Chinese models now account for 61% of total token consumption among the top ten most-used models. The three most-consumed models in a recent week were all Chinese. MiniMax M2.5 alone processed 2.45 trillion tokens in seven days — up 197% from the prior week. Moonshot AI’s Kimi K2.5 followed with 1.21 trillion. Zhipu AI’s GLM-5 came third.
In the first week of February 2026, Chinese models consumed 2.27 trillion tokens weekly. By mid-February, that figure had jumped to 5.16 trillion — a 127% surge in three weeks. American models consumed 2.94 trillion in the same period.
Andreessen Horowitz partner Martin Casado has estimated that roughly 80% of startups running open-source AI stacks are using Chinese models. OpenRouter’s own COO noted that Chinese open-weight models have become dominant because they are particularly suited to agentic workflows — the automated, multi-step AI pipelines that enterprise software increasingly runs on.
Why Chinese Tokens Are Winning on Price — and Physics
The competitive edge isn’t mysterious. It’s structural, and three layers deep.
Energy first. China’s electricity costs run roughly 40% below US levels. Every API call routed to a Chinese model gets processed in a Chinese data center on Chinese grid power. That electricity converts into AI inference at an estimated 22x markup over raw grid value — and it exports globally as a high-margin digital service that customs agents cannot touch, tariffs cannot reach, and trade statistics barely capture. China’s western provinces — Xinjiang, Inner Mongolia, Yunnan — supply abundant renewable power at scale, backed by a vertically integrated supply chain covering transmission, cooling, and server assembly.
Architecture second. US chip export restrictions forced Chinese labs to get lean. DeepSeek V3’s Mixture-of-Experts design activates only a fraction of its parameters during inference. Independent benchmarks put its inference cost at roughly 36 times lower than GPT-4o’s. MiniMax M2.5 carries 229 billion total parameters but activates just 10 billion at inference time. Constraint-bred efficiency.
Price third. MiniMax M2.5 charges $0.30 per million input tokens. Claude Opus 4.6 costs $15 per million input tokens — roughly 10 to 20 times more. In agentic workflows that burn through millions of tokens per hour, this price gap isn’t a nuance. It’s the decision.
Jensen Huang Agrees — Just From the Other Side
Here’s the moment that clarifies everything: Nvidia CEO Jensen Huang, at his GTC developer conference last week in San Jose, told the audience that “tokens are the new commodity” and declared Nvidia a builder of “AI factories” producing tokens at scale.
Two of the most consequential figures in global tech — one Chinese government official, one American chip industry titan — arrived at the same metaphor independently. That convergence isn’t a coincidence. It’s confirmation.
Corporate China Moves Fast
Alibaba has reorganized its entire AI operation into a new internal division: the Alibaba Token Hub, reporting directly to CEO Eddie Wu. The unit consolidates its Qwen model team, consumer apps, and enterprise AI products under one structure. Its stated purpose is creating, distributing, and applying tokens across the full AI stack. Alibaba isn’t describing itself as an AI company anymore. It’s calling itself a token factory — explicitly.
The market is tightening around that production. Alibaba Cloud, Baidu Smart Cloud, Zhipu AI, and Tencent Cloud all raised API prices within the same week in March 2026, citing surging token demand and supply-chain pressure. Prices going up while volume surges is what a seller’s market looks like.
Alibaba’s Qwen series currently ranks second globally in token call volume at 5.59 trillion. DeepSeek leads with 14.37 trillion. These are developer adoption numbers, API revenue signals, and — critically — geopolitical leverage embedded quietly inside codebases worldwide.
What This Doesn’t Mean Yet
The 61% OpenRouter figure deserves context. OpenRouter skews heavily toward independent developers and AI hobbyists. Enterprise procurement — where the big contracts live — still largely flows through American providers who offer compliance tooling, legal accountability frameworks, and deep integration support that Chinese providers haven’t fully matched.
Cheap token pricing also looks partly like a land-grab strategy. The wave of March 2026 price increases suggests the economics are already tightening. Training costs, talent, and the squeeze on advanced chips from US export restrictions are real constraints. Washington continues closing loopholes in its chip embargo, targeting additional Nvidia architectures and third-country routing arrangements. That’s a genuine ceiling on China’s ability to scale inference capacity long term.
Sovereign risk matters too. Regulated industries and allied governments face concrete legal exposure from routing sensitive workloads through Chinese infrastructure, regardless of the price per million tokens.
The Deeper Game
China’s token strategy is most coherent when you see it as infrastructure projection — the same playbook as fiber cables, container ports, and payment networks, applied to AI compute.
Tokens are intangible. They cross borders without paperwork and bypass tariffs. They don’t show up in trade statistics. You cannot sanction an API call. The data centers producing them sit inside Chinese sovereignty, beyond the reach of extraterritorial enforcement. Beijing appears to understand this clearly and is building accordingly — naming 2026 the “Year of Data Element Value Release” and assembling over 100,000 high-quality datasets totaling more than 890 petabytes as training infrastructure.
Chinese AI stocks rallied this week after state media highlighted the token surge. The market is pricing something that economists are slow to formally model: the country that most cheaply produces what the world’s software runs on will, over time, extract durable returns.
Oil made the Persian Gulf indispensable for a century. The question the ciyuan moment raises is straightforward and uncomfortable: what happens to the global technology economy when the cheapest, most scalable producer of its essential commodity operates outside your jurisdiction — and you never even see it in the trade data?
Related: OpenAI Shuts Down Sora: The $1B Disney Fallout and $15M/Day AI Problem Explained
