On July 20, the 2026 World Artificial Intelligence Conference concluded in Shanghai. With over 140 forums, 100,000 square meters of exhibition space, and more than 3,000 exhibits, the scale set a new record. But what truly caught my attention was not these numbers, but a signal conveyed by this year's conference: The focus of enterprises has completely shifted from "what AI can do" to "how AI can enter core business systems and create value."
In previous years, the scene of large models competing on parameters and dialogue capabilities is almost invisible this year. Instead, what we see are quietly running real machines—Huawei Ascend 950 super nodes, Super Fusion TokenBox, and over 160 intelligent agents from Dingjie running simultaneously. AI has shifted from "being able to talk" to "being able to do."
Today, I want to discuss, from the perspective of enterprise CIOs and ERP practitioners, the three most noteworthy signals at WAIC 2026 and what they mean for management software selection.
Signal 1: SuperFusion launched a "Smart Enterprise ERP" — this is not just another ERP vendor, but a computing power provider pushing into the ERP space
xFusion showcased several "big guys" at WAIC. The most eye-catching was not the product name Smart Enterprise ERP itself, but the complete value chain behind it: WATT (Energy) → FLOPS (Computing Power) → TOKENS (Token Production) → AGENTS (Agent Execution) → VALUES (Business Value Transformation).
In translation, it means: For AI to truly generate value in enterprises, it cannot rely solely on an "AI function" button. Instead, it must start from power assurance, then move to computing power supply, token production, and agent execution, ultimately landing on business value. Not a single link can be missing.
Core highlights: Super Fusion Smart Enterprise ERP covers key areas such as business analysis, operational processes, and quality management, driving AI from the "tool layer" to the "core business system" — achieving the three goals of "operable, manageable, and value-creating".
The accompanying TokenBox is even more noteworthy—it can run the full-blooded DeepSeek V4 1.6T large model on a single machine, with noise levels as low as 35dB (library-grade). This means high-performance AI can truly be deployed in an office, no longer requiring a dedicated server room. For small and medium-sized enterprises, this "deployability" is far more important than parameters.
Super Fusion is a company spun off from Huawei's x86 server business, and its foray into ERP is itself quite intriguing. Computing infrastructure manufacturers are systematically entering the enterprise management software field, indicating that ERP is no longer just a matter of "software"—in the AI era, the foundation of ERP is shifting from the "application layer" down to the "infrastructure layer."
Signal 2: Digiwin threw out 160 manufacturing intelligent agents — scenario density is ten thousand times more important than model parameters
At WAIC, Digital China showcased Athena Enterprise Intelligent Operation Space (EIOSpace), which is equipped with over 160 intelligent agents covering 100 high-frequency real-world scenarios in manufacturing. From executive data inquiries, quality anomaly handling, equipment fault prediction to supplier on-time delivery ranking — this is not a PowerPoint presentation, but a real machine in operation.
I roughly divided these 100 scenes into three categories:
| Scene type | Typical example | Input-output characteristics |
|---|---|---|
| Management Decision Category | Executive inquiry, business analysis, automatic report generation | Fast results, low barriers, direct benefits for managers |
| Quality/Production Category | Quality anomaly handling, equipment fault prediction, process parameter optimization | Highest ROI, but requires a data foundation |
| Supply chain collaboration | Supplier on-time delivery rate ranking, procurement anomaly warning | Cross-organizational collaboration is difficult, but its value is the most enduring. |
According to the latest data from IDC, the proportion of Chinese enterprises that have adopted large models and intelligent agents surged from 9.6% in 2024 to 47.5% in 2025—a fivefold increase. However, what makes Dingjue's approach interesting is that it does not build an AI platform first and then adapt it to scenarios, but rather the opposite: first make 100 real-world scenarios work, then let the scenarios drive platform development.
For friends doing ERP implementation, this is actually a more down-to-earth strategy. Rather than deploying an "enterprise AI operating system" that leaves clients unsure where to start, it's better to say directly, "I helped you automate quality anomaly handling" — the effect is immediate.
Signal 3: Token is becoming the "new currency" of enterprise AI — a new challenge for CIOs
At WAIC, there is another word that appears almost everywhere: Token.
xFusion has Token Factory (an AI production system for enterprises) and TokenBox, UCloud has launched a "Token Factory," and Infinigence's Agentic MaaS platform has seen its daily token call volume increase by about 40 times compared to the beginning of the year—Sheng Lei, Deputy Director of the National Information Center, said something at the conference that I think hits the nail on the head: "In the industrial economy, we measure by kilowatt-hours; in the internet economy, by traffic; and in the era of the intelligent economy, tokens are becoming an important yardstick for measuring the actual productivity of AI."
What does this mean for CIOs? At least three things:
First, Token consumption management will become a new function of IT operations. Just as enterprises today monitor server CPU usage, in the future they will need to monitor the Token consumption of each business scenario—which scenario is "burning money" and which scenario has a positive ROI must be clearly accounted for.
Second, when selecting an ERP, AI token efficiency will become a hard metric. For the same AI-assisted financial reconciliation, some ERPs consume 5,000 tokens per conversation, while others consume only 800 tokens—the gap is not insignificant.
Third, localized Token production capacity will become key to enterprises' AI autonomy and control. Super Fusion TokenBox can achieve noise levels of 35dB, meaning enterprises can deploy their own Token production facilities in offices, no longer fully relying on cloud APIs.
Another thing that cannot be overlooked is: China Mobile simultaneously launched the "Mobile Tiangong" industrial internet platform, collaborating with central state-owned enterprises to facilitate the implementation of 5,000+ AI projects; Lenovo introduced the FDE workshop, embedding intelligent agents into the entire process of "R&D, production, supply, sales, and service." Major telecom operators and hardware manufacturers are becoming new drivers of enterprise AI adoption—they have customers, scenarios, and computing power, but what they lack is the piece of the puzzle that brings "AI into core business systems." That piece is precisely ERP.
Three Practical Suggestions for Business Managers
After watching WAIC 2026, I have compiled three judgments that can be used in the second half of this year:
1. Redefine your "ERP AI Readiness". It's not just about whether your ERP vendor has AI features, but whether AI can enter your core business systems — whether it can directly manipulate data, automatically execute processes, and be managed and measured.
2. Scenario density matters more than model parameters. Don't be misled by big companies' large model parameters. Digiwin proved one thing with 100 scenarios: The key to AI implementation in manufacturing isn't how powerful the model is, but how many scenarios it covers. First, identify and successfully implement 3-5 of the most frequent and painful scenarios in your enterprise, then consider scaling up.
3. Token costs must be included in the ERP budget.In the second half of 2026, if you are evaluating an ERP upgrade or a new selection, it is recommended to treat "AI Token efficiency" as an independent evaluation dimension. Those AI features that consume a large number of Tokens per operation may be much more expensive to deploy than you think.
WAIC 2026 has concluded, but the signals it left behind will not fade quickly: AI is transforming from a "demonstration product" into a "production tool," from "assistance" into "execution." The next three years for the ERP industry will not be about incremental upgrades—but rather a rewrite of the underlying logic.
