China’s Daily Token Calls Top 500 Trillion as AI Industry Tests Token Economics

A token is a discrete unit used when a model encodes and processes text and other information. In large-model API services, input and output are usually counted and billed by token. Images, audio and actions can also use their own tokenization or other representations, but tokens from different models are not naturally equivalent. Tang Jie, a Tsinghua University professor and founder of Zhipu AI, said tokens have measurable, priceable and tradable characteristics, comparable to the “barrel” of the oil age and the “kilowatt-hour” of the electricity age.

Billing changes are already visible. Shanghai Mobile introduced “compute beans,” allowing users to buy and use model services through a quota that can be included in telecom bills, the report said. For companies, token-based pricing allows model usage costs to be split by department and task. DeepSeek began exploring peak and off-peak pricing in August and adjusted Flash-series prices on Sept. 10: off-peak cached-hit input was 0.02 yuan per million tokens, and the peak price was twice that. The figure applies only to a specific model, time period and cache condition, not to all tokens. Google said in its second-quarter 2026 results that Gemini models process about 22 billion API tokens per minute. Nvidia reported $89 billion in data-center revenue for the same period and said a new platform reduced the cost of generating each token.

These indicators—call volume, revenue and per-token cost—measure different things. The report said token measurement still needs calibration because tokenization, caching, modality, output quality and billing rules differ across models; comparing “price per million tokens” alone cannot establish cost performance. Work on token measurement standards is aimed at cross-model comparison and billing transparency. The report said token economics first makes intelligence services easier to measure and price, while procurement and deployment are still decided by total cost at equivalent quality and by the actual result delivered to users.

For compute suppliers, the question is shifting from how much capacity has been built to how much effective inference can be delivered reliably. Shanghai-based Enflame Technology listed on the STAR Market on Sept. 11 with an opening market value of more than 170 billion yuan. Industry reports on the listing day said the market is moving from evaluating hardware configuration toward token-generation efficiency and delivery quality, Leiphone reported. Customers still need chips, servers and cloud resources, but they increasingly ask how many real requests a given budget and service-quality constraint can support. Enflame Chairman Zhao Lidong highlighted “chip-model synergy”: deeper cooperation across the supply chain so domestic compute adaptation, chip development and model development proceed together. In early cooperation, fixing operators and features after a customer model upgrade took one to two months; some adaptations now take two to three days, and some scenarios only hours, he said.

Chip architecture and system scheduling are also being tested for efficiency. At a Pudong AI industry event on July 18, Zhangjiang company Dongfang Suanxin showed its DF1000 near-memory computing chip as a regional AI achievement. Near-memory computing aims to reduce the time and energy spent moving data, though actual gains depend on workload and deployment conditions. Wuwen Xinqiong said its cross-cluster heterogeneous inference architecture reduced first-token waiting time by 51.5% and per-token cost by 37.5% under specific test conditions. Those results reflect optimization in its test scenario and cannot be directly applied to all businesses. Throughput, first-token latency, continuous output speed, availability, energy use, model quality and total cost per task must be read together, the report said.

In embodied AI, the grasp on the quality-inspection line illustrates the chain. A humanoid robot takes a flat panel from a bin, places it into an inspection cabinet and sorts finished products, working alongside people. In Zhiyuan’s embodied model disclosed in September 2026, one frame can be encoded as 24 visual tokens, and experiments showed low-latency generation of continuous actions. A more compact visual encoding may reduce the information later inference must process and leave room for real-time control, but fewer tokens do not necessarily mean faster actions; model architecture, hardware execution and the control loop also matter. Embodied agents are potential large-scale token users because they must continuously perceive, predict and respond. Robots cannot wait for the cloud on every action, so end-side computing, cloud-edge-device collaboration and compute-network collaboration are being developed for millisecond-level response.

On Sept. 15, an open-source end-side inference engine cut the end-to-end inference latency of an embodied model from 278 milliseconds to under 26 milliseconds in public tests, according to the report. For dynamic operations, inference latency and network dependence can affect usability more directly than the unit price of a token. Zhangjiang-based chip company VeriSilicon focused on front-end data processing. Its chief strategy officer and executive vice president, Dai Wei, proposed that front-end perception and back-end compute work together, preprocessing information on the “cerebellum” side to reduce the computing burden on the “brain” side. The idea is to layer computing tasks rather than simply add back-end capacity. The core issue for models and devices is allocating limited compute to the parts that determine task success.

Corpora and rules sit on the data side. Compute runs the system and models perform calculation, while corpora determine what experience a model learns from. In embodied intelligence, multimodal corpora include video, sensor records, action trajectories, task labels and failure feedback, far beyond text. Pudong has more than 500 public datasets, and by August 2026 China had built more than 126,000 high-quality datasets, according to relevant departments cited by the report. Scale alone is not enough: usability, authorization and coverage of real scenarios matter, and moving from datasets to high-quality datasets and from effective training volume to scene performance requires iterative optimization.

Rules are following. On Aug. 27, 2026, China released the national standardization guiding technical document “Artificial Intelligence—Embodied Intelligence Data Quality Specification—Part 1: Real Data” (GB/Z 218.1—2026). The National Data Administration said it would promote embodied intelligence data standards in due course, the report said. Pudong is trying to organize supply around nearly 1,800 AI companies, more than 150 embodied intelligence companies and pilot-testing and mass-production resources. For developers, the report said, the valuable part is not a scattered resource list but accessible compute services, adaptable models, compliant data and verifiable application scenarios.

The 2026 Global Developer Pioneer Conference and International Embodied Intelligence Skills Competition will be held at Zhangjiang Science Hall from Oct. 23 to 25. Teams will face practical tasks including grasping, assembly, inspection and delivery. Zhong Junhao, secretary-general of the Shanghai Artificial Intelligence Industry Association, said: “We are not looking at what performances robots can do, but whether they can do work well at real stations and generate benefits.” The report said a successful grasp reflects algorithm tuning, sufficient compute and reorganized corpora, while sustained “stable, fast and worthwhile” grasping is the industrial capability.

Editor's Summary

China’s daily average token calls exceeded 500 trillion by June 2026, and the AI industry is using tokens to price model services and measure inference efficiency. The Leiphone report shows that token counts must be combined with compute efficiency, end-side latency, data quality and real-task success rates, especially in embodied AI, to determine whether intelligent systems create production value.