Quick Summary of the Key Points
This article serves as a kind of "data warning" for everyone following the AI industry: In the third quarter, official figures on the daily average usage of AI tokens nationwide will be released. The data as of the end of June showed 500 trillion tokens, which is a five-fold increase from the end of last year, leading many to predict that China's AI industry will surpass that of the United States in scale. However, the token, as a new unit of measurement for the AI industry, still does not have a unified statistical standard across the industry. The apparent discrepancy between China's higher token usage and its lower total revenue compared to the US is actually the result of several hidden factors, including inconsistent measurement methods, vastly different token values, the replacement of traditional computing power by domestic solutions, and the expansion of AI services overseas. When the new data for the third quarter is released, we should not just focus on the numbers but also analyze the underlying changes in the industry.
---
Detailed Breakdown and Interpretation
1. The 500 trillion tokens you see are not measured using the same standard
Many people assume that tokens are a uniform unit of measurement, like "kilowatt-hours" or "deliveries," where the count is straightforward. In reality, the methods used by different companies to calculate tokens vary significantly:
- Some companies only count tokens used by paid external users; for example, when Volcano Engine announced the share of its large model in the public cloud, it excluded tokens used for internal testing.
- Other companies include tokens used by their app users, internal staff, and third-party partners; for instance, DouBao’s reported token volume includes all scenarios.
- There are also cases of double-counting in cloud provider statistics: If a company sells both its own developed models and third-party open-source models, the same user’s requests may be counted twice.
- A large number of tokens are not even included in the statistics: For example, private deployments of AI models on company servers or local AI processes on mobile devices are not recorded by cloud or model providers.
This is similar to counting the number of food deliveries nationwide; some platforms only count paid orders, while others include meals for company employees or even test orders, leading to vastly different figures. The 500 trillion tokens officially released are just a rough summary from the supply side and do not represent an accurate count of the entire market.
2. The claim that "China’s token usage surpasses that of the US" is not substantiated
Many online discussions compare OpenAI’s and Google’s daily token usage (22 trillion and 32 trillion, respectively) with China’s 500 trillion, concluding that China’s AI industry is ten times larger. This is incorrect:
- The US does not have a unified token measurement standard for the entire market. The numbers provided by OpenAI and Google only account for tokens generated by their APIs and do not include internal usage within their ecosystems (e.g., tokens from Google searches or Office’s built-in AI).
- Comparing based on computing power is more reliable: The total power consumption of AI in the US is about 13.7 GW, while in China, it is about 4.4 GW, which is more than twice as much. With such a difference in total computing power, it is impossible for China to generate ten times as many tokens as the US.
- The OpenRouter platform, often used for comparison, is mainly for small and medium-sized developers and does not represent the market for large companies that directly purchase OpenAI and Anthropic services, making the comparison irrelevant.
3. The value of a token varies significantly between China and the US
Even if the number of tokens is similar, the value of each token differs greatly:
- The quality of tokens varies: In China, many tokens were obtained for free by users for trivial purposes, such as writing jokes or generating memes, while in the US, tokens are used for high-value tasks like legal documents or financial risk modeling, creating much more value.
- The context in which tokens are used also affects their value: US companies are willing to sacrifice some token generation speed for lower latency and faster responses, leading to higher per-user prices. In China, models were often used to increase token volume, often for low-value tasks, resulting in lower prices.
- Labor costs differ: The hourly cost of programmers, lawyers, and consultants in the US is several hundred dollars, while in China, it is only a few hundred yuan. The savings from using AI to replace human labor are significantly higher in the US, leading to higher token prices.
4. Two major trends behind the 500 trillion tokens
Two important trends are hidden within this figure:
- Domestic computing power is rapidly becoming less dependent on NVIDIA: Before 2023, 95% of China’s AI computing power came from NVIDIA GPUs, but by 2025, NVIDIA’s share has dropped to 55%, with domestic chip manufacturers accounting for 41%. This trend will continue, and by 2030, domestic computing power could account for over 90%. This means that China will have more control over the cost of token production.
- Tokens have become a new export commodity: More than 10% of the tokens used by leading Chinese model companies come from overseas users. For example, MiniMax earned 60% of its revenue from overseas in the first half of the year, and Volcano Engine’s overseas model usage accounts for nearly half of its revenue. This indicates that China’s token growth may not be solely driven by domestic demand but also by overseas users.
In summary, while the 500 trillion tokens represent significant progress in China’s AI industry, the figures should be analyzed in context, considering differences in measurement methods, industry standards, and the value of tokens.