China's 100 quadrillion token story tells you one thing for sure: Beijing now wants AI inference measured like industrial output.
China has turned a billing unit into an economic statistic. In March, Liu Liehong, head of the National Data Administration, gave the AI token an official Chinese name, ciyuan, and described it as a settlement unit linking technology supply with commercial demand. That sounds dry. It isn't.
According to figures later published by Xinhua, China's average daily token calls had passed 140 trillion by late March 2026. The same official data put the number at 100 billion at the start of 2024 and 100 trillion at the end of 2025. You don't need to love the metric to see why Beijing likes it: it turns AI use into a number officials can track and compare - then push.
The full-year base is already large. A National Data Administration survey released at the Digital China Summit in Fuzhou said China logged about 21.1 quadrillion token calls in 2025. If daily usage stays well above March's level, 2026 clears 50 quadrillion without trying. If the largest platforms keep growing the way ByteDance's Doubao has, 100 quadrillion stops sounding wild.
Doubao is the reason the national number no longer feels abstract. At the Volcano Engine Force Conference on June 23, Volcano Engine president Tan Dai said Doubao's daily token usage had reached 180 trillion, according to Jiemian. That is more than the national daily level reported three months earlier.
ByteDance is seeking $20 billion offshore to fund an AI buildout that goes far beyond TikTok
ByteDance is seeking a $20 billion offshore loan, its largest ever, nearly doubling the $10.8 billion record it set nine months ago. The debt is earmarked for AI infrastructure as the TikTok parent positions Doubao against OpenAI and Google at frontier model scale. The dollar-denominated structure reflects both the size of ByteDance's AI ambitions... - ByteDance seeking 20 billion dollar loan - AI infrastructure investment beyond social media
Here's the thing: a bigger token count isn't the same as a smarter model.
Much of the burn now comes from video, images, agents, and other inference-heavy products that run again and again after the model has already been trained. Independent analysis by Pebblous estimated that ByteDance's Seedance video model can consume about 40,000 tokens to generate one second of 1080p video. A minute of video can swamp a very large number of chatbot replies. So when you hear 100 quadrillion, don't picture only people asking homework questions. Picture short clips, enterprise workflows, customer-service bots, search replacements, and agents quietly calling tools in the background.
The number is useful, but it flatters the wrong thing #
Tokens are a clean way to count demand. They show whether people and companies are actually using AI services rather than merely announcing pilots. They also reveal where the money goes, because every token has to be processed on servers packed with chips, memory, networking gear, and power-hungry cooling systems.
But tokens don't measure judgment, originality, or frontier capability. A token spent checking an invoice and a token spent helping a scientist reason through a protein problem both show up as one token. That is the trap. Beijing has created a statistic that is easy to scale and easy to celebrate, while the harder question sits beside it: what work did all that compute actually do?
China's own policy machine is now feeding the curve. A January action plan issued by eight departments, including the Ministry of Industry and Information Technology, set targets for deeper AI use in manufacturing by 2027: three to five general-purpose large models. It also called for 100 high-quality industrial datasets and 500 typical application scenarios. The policy does not need every factory to build a frontier model. It needs factories to use inference all day.
Chips are the hard edge #
Investors have already noticed. Cambricon reported 5.996 billion yuan in first-half 2026 revenue, up 108.13% from a year earlier, according to its interim results reported by TrendForce and MarketScreener. Hygon Information's first-half report showed 9.1 billion yuan in revenue, up 66.52%, and operating cash flow of negative 427.5 million yuan.
Those are real numbers. They are also not Nvidia numbers.
ByteDance Locks In Nearly $30 Billion Loan for Its AI Buildout ByteDance has closed a $29.6 billion loan, Asia's second-largest dollar borrowing of 2026, after investors offered more than $30 billion in orders for a facility the company originally sized at $20 billion. The money is set to fuel a data center and AI infrastructure spending plan that could reach $70 billion this year. - bytedance 30 billion dollar loan for ai infrastructure - how bytedance is funding its artificial intelligence buildout
The constraint is memory as much as processors. High-bandwidth memory still comes mostly from three firms - SK Hynix, Samsung and Micron - while Chinese chip designers depend on a supply chain that is growing fast but still uneven. Hygon's negative operating cash flow shows the pressure in plain accounting terms: inventory and procurement can run ahead of finished sales when demand is rising and supply is tight.
JPMorgan has forecast that China's AI inference token consumption could grow roughly 370-fold between 2025 and 2030, a call reported by Wallstreetcn and other financial outlets earlier this year. That forecast is aggressive, but the direction is believable. Agents don't ask one question and stop. They plan, retry, search, call tools, and check their own work. Every step adds tokens.
Frankly, the headline number is both impressive and slippery. It proves China is making AI usage cheap enough and common enough to measure at national scale. It does not prove China has closed the frontier-model gap with the United States. A country can burn through huge inference volume on video generation and customer-service agents while still depending on imported memory and struggling to match the top training clusters.
That is why 100 quadrillion tokens is worth watching. It is not a trophy. It is a power bill, a chip order, a policy target, and a bet that AI will become ordinary industrial infrastructure before anyone has agreed how to measure its real output.
Also read: Jacob Coxon Says His Former Anthropic Colleagues Are Genuinely Frightened of AI • Obama Tells Democrats AI Is Moving Fast and Turning Dangerous • PwC Merges Its US and India Arms Into a 40,000 Person Firm to Beat AI to the Punch
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.