In May 2026, China consumed an average of 140 trillion tokens per day. This number was nearly zero two years ago, and now exceeds that of any other country in the world.
The "2026 China AI Application Panorama Report" released by QuantumBit Think Tank at the 4th China AIGC Industry Summit quantified this matter. The monthly visits to domestic AI application web pages exceeded 900 million, monthly downloads of apps surpassed 240 million, and daily active users surged 223% year-on-year.
For enterprise CIOs, the signal is clear: your employees are already using AI, whether you have approved it or not.
Among the five major trends, three are directly related to ERP
The report identifies five major trends, and I will discuss the three most closely related to enterprise digitalization.
140 trillion: China's daily token call volume in 2026, growing over a thousand times in two years (Quantum Bit Think Tank)
First, agentification. AI has evolved from "answering questions" to "executing tasks." According to the report's estimates, the token consumption for a single agent action is a hundred times that of traditional AI applications. This is why token consumption has skyrocketed—it's not that people are chatting more, but that AI is actually working behind the scenes, and it is performing complex tasks involving multiple steps and tool calls.
For ERP vendors, this means the positioning of "AI assistant" is already outdated. The "Autonomous Enterprise" that SAP promoted at Sapphire 2026 follows this logic—AI is not about helping you query data, but directly executing financial closing processes. The Odoo community is also discussing a similar approach, embedding AI agents into business processes rather than hanging them in the sidebar.
Second, becoming an entry point. ByteDance, Alibaba, Tencent, and Baidu collectively spent over 4.5 billion yuan around this year's Spring Festival, competing for a very specific thing: when users need AI, which platform they open first.
This has an impact on corporate procurement decisions. In the past, CIOs chose ERP based on feature lists and price. Now they also need to consider: who controls the primary entry point for employees to use AI within this system? If the daily AI entry point for all employees is a certain large model within DingTalk or WeCom, then the AI modules of ERP vendors may never be clicked open.
Third, monetization. Within less than 20 days of Kimi K2.5's release, its revenue exceeded the entire year of 2025. After Zhipu API raised prices by 83%, call volume increased rather than decreased. This shows one thing: when AI is truly embedded into workflows, users are willing to pay, and price sensitivity is not as high as imagined.
The old approach to enterprise software—"get users hooked for free first, then figure out how to monetize"—may not work in the AI era. Enterprise customers are already willing to pay for valuable AI capabilities, provided they truly save on labor costs.
Apple is also coming in
Apple officially announces WWDC 2026 scheduled for June 9, with the core highlight being the independent app reconstruction of Siri. This is not just a mobile feature update; it has implications for enterprise digitalization.
This time, Siri uses Google's Gemini model, but all requests run on Apple's own private cloud, so Google cannot access the data. This architecture is noteworthy—it addresses one of the most troublesome issues in enterprise AI deployment: large models are very powerful, but data cannot leave.
If this Apple system is rolled out with iOS 27 in the fall, it means hundreds of millions of devices will come with a local + private cloud hybrid AI entry point. The new question facing enterprise IT departments is: should the AI on employees' phones be integrated with the AI in the company's ERP? Can it be integrated? How is data compliance calculated?
It is still too early to make a judgment, but the direction is clear: the AI entry point will shift from browsers and chat windows down to the operating system layer. Microsoft is already doing this (Copilot embedded in Windows and M365), and Apple following suit is expected.
140 trillion per day, but what does the enterprise side say about the data
The surge in token consumption indicates that AI usage has indeed increased. However, usage does not equal value. KPMG's "2026 Global Technology Report on Industrial Manufacturing" for the same period provides another set of figures: 49% of manufacturing executives have implemented AI and generated business value, but 76% frankly state that unreliable data is the primary risk.
These two pieces of data together present a clearer picture:
| Dimension | Consumer/Developer Side | Enterprise implementation side |
|---|---|---|
| Core Indicators | Token consumption, DAU, download volume | Whether it generates commercial value, data quality |
| 2026 Status | Explosive growth (140 trillion/day) | 49% implemented, but 76% stuck in data |
| bottleneck | Computing power costs, entry point competition | Data silos, chaotic standards, and weak cleaning capabilities |
| Next step | Monetization and vertical deepening | Data governance precedes AI modeling |
For those involved in enterprise digitalization, this comparison is very useful. The approach of consumer-side AI is to burn money to seize entry points, iterate quickly, and tolerate errors; the approach of enterprise-side AI must be data-first, start with small scenarios, and expand only after calculating the ROI.
The two sets of logic cannot be mixed. Applying consumer internet strategies to enterprise AI often leads to problems during project delivery.
Practical advice for CIOs
Don't make "whether employees are using AI" a primary KPI. They are likely already using it, just not telling you. The real KPI should be: which business processes have become faster due to AI, how much faster, and how many person-days have been saved.
First, take stock of data, then discuss AI strategy. KPMG's data shows that 76% of manufacturing companies are stuck on unreliable data. This data is equally applicable to service and trading companies. If 30% of customer addresses in the ERP are wrong, any AI predictions trained on this data are garbage in, garbage out.
Seize the internal AI gateway. Gateway-oriented strategies are not just for the consumer side. Within an enterprise, where is the primary gateway for employees to use AI every day? If it's a third-party mini-program within WeCom, then the money you invested in the ERP AI module might be wasted. Rather than waiting for vendors to do it for you, it's better to figure this out now.
Start with a small scenario, avoid being big and comprehensive. The Quantum Bit report mentioned that the tools with the highest retention rate are intelligent assistant tools, but those generating the deepest commercial value are vertical scenarios. These two are not the same thing. When choosing the first scenario for an AI project, it is better to select a narrow scenario where cost savings can be calculated, rather than a broad scenario that "the whole company can use" but whose value is unclear.
140 trillion tokens is a signal, not the final outcome. The rules of the game on the enterprise digitalization side are not the same as those on the consumer internet side. Understanding the difference is more important than chasing trends.
Source: QuantumBit Think Tank "2026 China AI Application Panorama Report" (2026-05-20), Apple WWDC 2026 Official Announcement (2026-05-20), KPMG "2026 Global Technology Report on Industrial Manufacturing" (2026-05-26), IBM Think 2026 (2026-05-07)
