In the past two weeks, three events occurred almost simultaneously. Individually, they appear to be industry news, but when viewed together, they point to the same conclusion: the underlying paradigm of enterprise digitalization is shifting to a new set of rules.
Signal 1 · Capital Market: The fifth set of standards for the STAR Market expands to AI large models—"Liang Wenfengs" can go public. Signal 2 · Technical Architecture: Microsoft Copilot Cowork integrates DeepSeek—the first time a major U.S. company incorporates a domestic model into a core B-end product. Signal 3 · Development Model: IDC predicts 75% of new applications will be implemented via AI low-code—traditional development enters structural elimination.
Three lines, from the three dimensions of capital access, technical architecture, and development efficiency, all point to the same thing: the human-driven model of enterprise digitalization can no longer hold up, and the AI-driven model is taking over. Let me break this down one by one.
Signal 1: The STAR Market Opens the Door for AI Large Models
On June 17, CSRC Chairman Wu Qing clearly announced at the Lujiazui Forum that the scope of application of the fifth set of standards for the STAR Market will be expanded to the AI large model industry. That same afternoon, the Shanghai Stock Exchange issued a supporting review guidance draft for public comment.
Let me translate the weight of this matter: The fifth set of standards on the STAR Market was originally an IPO channel tailored for unprofitable biomedical companies — no profits required, but as long as your technology has prospects and phased achievements, you can go public and raise funds. Now this channel is officially open to AI large model companies.
35 trillion "Two Innovation Boards" total market value 45% of trillion market value, technology enterprises account for 54 newly accepted IPOs on the STAR Market
The review guidelines define "phased results" as: at least one large model product must have been launched and achieved large-scale application at the time of application. This means you can't just submit a random PPT; you need to have a real product that has been deployed and is being used by users.
The guidelines also particularly support the simultaneous application of general-purpose large models and industry-specific models. This is good news for companies developing industry-specific large models (such as financial large models, manufacturing large models, and legal large models)—it means they don't necessarily have to build a general-purpose foundational model; they can also succeed by deeply cultivating a single industry.
Qiu Yong, chairman of the Shanghai Stock Exchange, made a very practical remark: strictly prevent "submitting applications with problems" and "rushing in all at once." The approval channel is open, but the review standards will not be lowered. AI companies with real technology, practical implementation, and commercialization will benefit, while those riding the hype will be filtered out.
What is the impact of this on enterprise digitalization? AI large model companies now have formal financing channels in the capital market, enabling sustained R&D investment, faster product iteration, and accelerated commercialization of industry-specific large models. When you are selecting a supplier, you no longer need to worry about them going out of business due to a broken capital chain.
Signal 2: Microsoft chooses DeepSeek to reduce costs
News from June 22: After the commercial launch of Microsoft Copilot Cowork, computing power costs have continued to rise, and the company plans to introduce DeepSeek V4 as a tiered alternative model.
After thinking about the industry significance of this matter repeatedly, I believe there are two layers:
The first layer is cost logic. After Copilot Cowork launched, over half of Fortune 500 companies deployed it early, with high-frequency multi-tool calls generating massive token consumption. The fixed subscription model could not recover costs. Microsoft then adjusted its billing system, adding pay-as-you-go elastic pricing, while building an intelligent tiered routing system—simple tasks route to DeepSeek (low cost), complex reasoning routes to high-end closed-source models (high capability). The entire system includes a quality verification mechanism, ensuring that switching to lower-cost models does not lower delivery standards.
The second layer is strategic logic. This marks the first time a leading U.S. tech giant has incorporated a domestically developed large model into its core B-end AI product. In the past, overseas manufacturers relied solely on their own high-end models, but now they are beginning to adopt a multi-model hybrid architecture to hedge against cost pressures. Leveraging its MoE hybrid architecture, DeepSeek achieves significantly lower computational overhead for equivalent tasks compared to overseas flagship models—this advantage has evolved from a "cost-effective choice for the Chinese market" into a "practical tool for global enterprises to reduce AI costs."
My judgment: Under the challenge of AI burning cash, layered routing and multi-model mixing will become common solutions for global vendors. This is good news for CIOs handling enterprise AI selection — you don’t need to choose an "all-purpose model," but can schedule tasks by complexity in layers, reducing total costs by 30-50%.
Signal 3: 75% of New Applications Are Deployed via AI Low-Code
IDC released the "2026 Global Intelligent Application Development Trends Report" in early June, presenting a set of data that made me repeatedly verify the source:
75% New AI Low-Code Construction by End of 2026 47.3% Growth Rate of China's AI Low-Code Market 3.8% Growth Rate of Traditional Software Development 18% Share of Pure Native Development
47.3% vs 3.8%——The gap between these two numbers is no longer a trend, but a chasm. The market share of traditional development models has fallen below 20%, officially entering structural elimination.
Gartner's concurrent report also provides cross-validation: by the end of 2026, 75% of new applications globally will be built through AI low-code platforms, an increase of 45 percentage points compared to 2025. The AI low-code penetration rate among small and medium-sized enterprises will exceed 80%.
| Comparison Dimension | Traditional native development | AI low-code | Gap |
|---|---|---|---|
| Delivery cycle for small and medium-sized systems | 3-6 months | 3-15 days | Efficiency improvement of 90%+ |
| 3-year full-cycle R&D cost | 200,000-500,000/set | 50,000-120,000 per set | Cost reduction 75% |
| Demand Response Cycle | 45 days | 3 days | Speed increased by 93% |
| Code Bug Rate | 8-12% | 1-3% | Reduce by 75% |
| Xinchuang compliance adaptation | Weak, difficult to modify | Powerful, native adaptation | Structural gap |
These data are not theoretical deductions, but actual measurement results from IDC Q2.
My judgment on this matter is quite straightforward: traditional development will not completely disappear—ultra-large financial cores, military-related confidential systems, and highly complex distributed architectures still require native development as a fallback. However, for 90% of digitalization scenarios in government, manufacturing, and commerce, AI low-code comprehensively outperforms in efficiency, cost, iteration speed, and compliance adaptation. This is not an "optional upgrade" but a "necessary replacement."
Three signals point to the same conclusion
Look at the three lines together:
The capital market has opened the front door for AI—AI companies can continuously raise funds, conduct research and development, and iterate. You have more confidence in choosing AI suppliers.
Technology architecture shifts to multi-model hybrid—you don’t need to bet on the future of a single model, but instead schedule tasks by layer, with controllable costs and diversified risks.
Development model shifts to AI low-code—you don’t need to maintain a 30-person development team to build a middle-platform system; 3 people plus an AI low-code platform can handle it, reducing the delivery cycle from six months to two weeks.
The picture pieced together from these three things is as follows: capital supports technological iteration, multi-model routing reduces technical risk, and AI low-code shortens delivery cycles. When all three conditions are met simultaneously, the bottleneck for enterprise digitalization is no longer "insufficient technology," but rather "whether the organization can keep up with change."
Practical advice for enterprise CIOs
First, reassess the development team structure. If your team is still using traditional native development for general business systems, seriously consider switching to an AI low-code platform. It's not about cutting people, but about shifting from "writing code" to "designing business logic + reviewing AI output." With the same team, productivity can be increased by 3-5 times.
Second, design a multi-model routing strategy. Don't put all AI tasks on a single model. Use low-cost models (like DeepSeek) for simple tasks, high-capability models (like Claude/GPT) for complex reasoning, and balanced models for the middle layer. Microsoft is already doing this, and you can too.
Third, pay attention to the capital health of AI suppliers. The STAR Market has opened the door for AI, meaning that AI companies with implementation capabilities have sustainable financing channels. When selecting a supplier, ask: Do you have plans to go public? Do you have users with large-scale applications? If you only have technology but no implementation or capital support, the risk is significant.
Fourth, stop waiting for a "mature solution." 75% of new applications are already being built with AI low-code, and a market growth rate of 47.3% shows that early adopters are already moving forward. If you wait until the "technology is fully mature" to make a decision, you'll find that your competitors have already launched three systems using AI low-code, while your requirements document is still in the review stage.
