June 2, Beijing. A list called "2026 China AI Agent Pioneers" was released.
I went through this list three times. Not just because it includes AI agent cases from over 100 companies, but because a few things caught my attention for a long time: Hubei Bank uses AI for code security auditing, BOE uses AI for database operations management, Shougang uses AI for hot rolling production decisions, and Muyuan uses AI for monitoring pig herd health. Across more than 20 industries, from banking to pig farms, from highway inspections to satellite testing—AI agents are no longer in the "pilot" phase; they are running in the core business processes of enterprises.
But on the same day, another figure was also striking: Zhipu AI issued an announcement that it would list on the STAR Market, raising 15 billion yuan while incurring a loss of 4.7 billion yuan.
Looking at these two events on the same day, the landscape of China's AI in June 2026 becomes clear: the application layer is burning hot, while the capital layer is still bleeding.
What the four tracks of the Top 100 actually reveal
This list is jointly initiated by the Internet Society of China and the China Artificial Intelligence Industry Development Alliance, and hosted by BCS 2026. The biggest difference in the selection criteria compared to the past is that "application value" and "security and controllability" each account for half of the weight. Technical parameters are no longer the only yardstick; how long the intelligent agent has been running in real business, whether it is safe, and whether it can be replicated—these are the key factors.
The list is divided into four tracks and is well worth a look:
I noticed a detail: the industry application track has the broadest coverage — all eight sub-fields of government affairs, energy, finance, manufacturing, healthcare, agriculture and animal husbandry, aerospace, and education have entries. This is not a proof of concept for "what AI can do," but rather "AI has generated quantifiable value in actual business operations."
Shougang's hot rolling production AI agent runs on real production lines, BOE uses agents to manage database operations, and Zhongshan Municipal Government Services and Data Administration uses agents for secure operations on the government external network—these are not PPT projects.
If the theme of enterprise AI in 2024 was "can it be used," and in 2025 it was "should it be used," then by June 2026, the theme has shifted: how to use it safely, how to scale it, and how to continuously iterate.
IDC's Cold Water: In ERP Selection, Look at Data First, Then AI
Just before and after the release of the ranking, IDC published the report "China AI-Enhanced Enterprise ERP Market Share, 2025".
Yonyou has taken the top spot in the AI-ERP market. Over 60% of mid-sized manufacturing enterprises have already incorporated AI capabilities into their core criteria for ERP selection. The three scenarios of finance, supply chain, and production scheduling have already yielded quantifiable return data.
An electronic components company in Shenzhen reduced its month-end closing from five working days to half a day. Aier Eye Hospital deployed 17 digital intelligence employees, improving financial processing efficiency by nearly 8 times. A hardware factory with an annual revenue of 120 million yuan used AI demand forecasting, reducing raw material inventory turnover from 68 days to 34 days and releasing 2 million yuan in working capital. After Shuangliang Group implemented AI scheduling, the accuracy of production plan execution increased by 50%.
These data look exciting. However, IDC said something in the report that many people might overlook: First assess the solidity of your own data foundation, then determine the AI-ERP entry module based on business pain points, and during selection, let the system run real business data for verification.
This sentence translates to: If your master data is not managed clearly, no matter how powerful the AI-ERP is, it cannot be implemented.
I think this is the most worthwhile sentence for enterprise CIOs to read three times among all the materials this week. The Top 100 list says "AI has entered core processes," but IDC says "Whether you can enter depends first on your data foundation." One talks about the ceiling, the other about the floor. Both are correct.
The "Fire" of AI Implementation
· BCS Top 100: 20+ industries, full coverage across four tracks
· BOE/Shougang/Muyuan etc. have entered the production stage
· AIER Eye Hospital's 17 digital employees improve efficiency by 8 times
· Hardware factory inventory turnover halved, releasing 2 million
· Huawei Cloud 8.5 million developers, 50,000+ partners
The "Ice" of AI Implementation
· IDC Reminder: Look at the data foundation first, then AI
· Zhipu loses 4.7 billion, market value 630 billion
· Large model companies are generally not profitable
· Enterprise data governance remains the biggest bottleneck
· 60%+ of manufacturing enterprises choose AI-ERP, but there is still a gap in POC validation
Zhipu loses 4.7 billion in IPO, what is happening in the large model track
On June 1, Zhipu AI issued an announcement on the Hong Kong Stock Exchange: the board of directors has passed the A-share issuance proposal, planning to list on the Shanghai Stock Exchange's STAR Market. This company, which just listed on the Hong Kong stock market in January, is set to pursue an A+H listing in less than half a year. The fundraising scale is 15 billion yuan.
But a glance at its prospectus reveals: revenue grew rapidly in the first half of the year, yet it posted a loss of 4.7 billion yuan. Its market value is approximately 630 billion yuan. This combination of "losing 4.7 billion, valued at 630 billion" is unimaginable in traditional industries. However, the large model track now follows a different valuation logic: it focuses not on current profits, but on technological generational gaps and ecological niches.
Zhipu is not the only one burning through cash. DeepSeek, Moonshot AI, MiniMax, Baichuan — none of the "AI Six Tigers" have truly achieved profitability yet. They are all betting on the same thing: as model capabilities continue to improve and economies of scale emerge, once inference costs drop to a certain threshold, their business models will become viable.
What does this mean for enterprise users? Two judgments:
First, there’s no need to worry too much about suppliers going bankrupt in the short term. Zhipu’s ability to list in Hong Kong and then push for the STAR Market shows that the capital market is still buying into large models. A market cap of 630 billion yuan corresponding to a 15 billion yuan fundraising provides strong credit backing.
Second, but you need to be prepared for model switching.The large model landscape is still being reshuffled. This week, NVIDIA released Nemotron 3 Ultra, and Huawei Cloud ModelArts Next supports on-demand routing for 15 SOTA models. In the future, enterprise AI architectures should be "not tied to any single model." Choose the platform, not the model—this is a more prudent strategy.
Three Suggestions for Heads of Enterprise Digitalization
This week has been packed with information, but the theme is actually just one: AI has fully entered the core processes of enterprises, just not in the way many people imagined.
First: Don't wait for the "perfect AI solution" to appear; start by addressing specific pain points.None of the successful cases on the Top 100 list involved "implementing a company-wide AI strategy." They all started from a specific scenario—Hubei Bank began with code security auditing, Muyuan with pig herd health monitoring, and Shougang with the hot rolling production line. Once they achieved success, they expanded horizontally. This approach is far more reliable than "first making a three-year AI plan."
Rule 2: Every dollar spent on data governance is more valuable than every dollar spent on AI models. The reminder from IDC is not just a courtesy. Among the manufacturing companies I have worked with, the difference in effectiveness of deploying the same AI-ERP system between those with good data foundations and those without can be as much as 5 to 10 times. It's not that AI is inadequate; it's that the data is inadequate. When your material codes are still inconsistent, BOM data still requires extensive manual maintenance, and financial account mappings still have numerous exceptions—under such circumstances, deploying AI for production scheduling or AI for forecasting will inevitably result in diminished effectiveness.
Article 3: Pay attention to things like NVIDIA OpenShell and Huawei Cloud AgentSphere. Once AI agents start running in enterprises, security governance will become the most troublesome issue. It's not about whether you "want to manage" it, but whether you "can manage" it — whether you can see in real-time what the Agent is doing, whether you can shut it down with one click, and whether you can set permissions by role. This week, NVIDIA and Huawei Cloud simultaneously launched Agent security runtimes. This is no coincidence. It's the market telling you through products that this problem has reached a point where it must be solved.
100+ BCS selected enterprises 20+ covered industries 60%+ manufacturing enterprises AI included in ERP selection 4.7 billion Zhipu semi-annual loss (yuan)
The first full week of June 2026 gave me the feeling that AI has moved past the "believe it or not" stage, and also past the "want it or not" stage. Now it has entered the "how to use, how to manage, how to calculate the costs" stage. What is being competed in this stage is no longer technical capability, but organizational capability and data capability.
This is the moment when the enterprise digital leader truly steps onto the stage.
