Last week, news that Kingdee Cloud·Galaxy V9.1 had stopped new purchases spread across the industry. Starting July 21, the official purchase channel for this public cloud ERP version, which has served tens of thousands of small and medium-sized enterprises, was officially closed, with only one alternative—Kingdee AI Suite Public Cloud. Want to keep buying the old version? That's possible, but only through a special approval under the "Product Beyond Lifecycle Sales Control Application." However, those in the know understand that the timeline and outcome of this special channel are both uncertain.
Almost at the same time, Inspur Data released its self-developed AI data operating system in Beijing, announcing its upgrade from a cloud computing and distributed storage service provider to an AI infrastructure service provider. A few days earlier, IDC released its annual AI rankings, and Kingdee was selected for the "IDC AI 50".
这几件事单独看各有各的叙事，但放在一起，信号就很清楚了：国产 ERP 的 AI 化，已经从一个宣传口号，变成了切切实实的产品动作和商业决策。这篇文章想聊聊这三件事背后的共同逻辑，以及它对正在选型、正在用国产 ERP 的企业意味着什么。
1. Discontinuing old versions is the most straightforward statement from manufacturers
Let me first talk about Kingdee's suspension of sales this time. Many people's first reaction is "the vendor wants to force us to upgrade," which is correct in direction, but only half right.
Kingdee's product moves in 2026 are actually a combination of punches: at the beginning of the year, multiple historical versions of the K/3 Cloud series were discontinued; in March, the Galaxy V7.x series was discontinued; and in July, it was V9.1's turn. Each step narrows the old product lines, guiding incremental customers toward new directions. And the new direction is very clear—AI suite. In the first half of the year, Kingdee released the enterprise AI-native operating system "Lingji," and management set an annual target of "1 billion yuan in AI-native product revenue," even proposing to "rebuild Kingdee with AI by 2030."
说白了，停售不是目的，是手段。厂商在用生命周期管理这种最硬核的方式，告诉市场：AI 不是 ERP 的插件，而是 ERP 的新底座。你可以在外面慢慢观望，但增量市场已经没有旧版本可以买了。
For existing users, systems already deployed are unaffected, and renewals and services remain normal—Kingdee has made this clear. What really needs attention are two groups: first, enterprises planning to purchase a new ERP system, who now essentially only have AI suites to choose from, so their selection logic must adapt; second, existing customers planning to expand capacity or switch modules within the next one to two years, as decisions made during this window will affect their technology roadmap for the following three to five years.
This trend is not unique to Kingdee. Yonyou's YonClaw, Digiwin's Athena, Inspur's industrial large model, plus IDC's repeated emphasis on "the AI-enhanced ERP market growing 96.1% year-over-year in 2025," domestic vendors have collectively placed AI at the core of their product strategies. The difference is only that some are fast, some are slow, some are aggressive, and some are moderate.
二、AI 要落地，先得有"底座"
Beyond the product actions on the vendor side, there was another easily overlooked but significant piece of news this week: Inspur Data released its self-developed AI data operating system.
Many people find the term "data operating system" abstract, but it actually points directly to the real pain points of enterprise AI implementation. Inspur Data summarized four major industry challenges at the launch event: first, the difficulty of continuous evolution—enterprises cannot afford to waste their existing cloud and data infrastructure, and new AI computing power must be integrated into the current environment; second, the difficulty of data readiness—data silos are widespread, and multi-source, multi-modal data cannot be unified; third, the difficulty of ecosystem collaboration—chips, storage, models, and applications each operate independently, lacking a unified foundation; fourth, the difficulty of application deployment—how data becomes knowledge and how computing power translates into business value, with return on investment being difficult to assess.
These four points will resonate deeply with anyone who has worked on enterprise AI projects. Especially the second one—data readiness. We have served many enterprises, and many bosses assumed that deploying a large model would let them "have AI analyze business data." But once the project kicked off, they realized that merely cleaning and aligning the inconsistent data standards across ERP, CRM, and MES would consume over half the budget. No matter how smart the AI is, if you feed it dirty data, it will spit out garbage conclusions.
There are two points worth noting in the design logic of Inspur's system:
- "Xuantong" is responsible for the data foundation—high-performance all-flash converged storage, addressing how data is stored and how to make it faster;
- "万象"负责智能体运行环境——解决 AI 代理在什么环境里跑、怎么调度算力的问题。
In plain terms: build the data "granary" on one side, set up the AI "chef's" workbench on the other, and package both into standardized infrastructure to sell to you.
For enterprises, this means an important cognitive update: The biggest cost of AI implementation has never been the model invocation fee, but data readiness. Whether choosing the AI suite of a domestic ERP or using Odoo to build your own AI capabilities, data governance is an unavoidable upfront investment. Whoever sorts out their data foundation first will be able to reap the benefits of AI faster.
III. Practical Inflection Point: From "Can Chat" to "Can Work"
If Kingdee's suspension of sales is the accelerator on the "vendor side," and Inspur's release of the foundation is the accelerator on the "infrastructure side," then this week there is another piece of news from the "application side" worth considering together—a judgment that NVIDIA founder Jensen Huang has repeatedly emphasized recently: AI has entered the inflection point of practical application.
His core point is: over the past two years, everyone has been concerned about whether AI "can chat," but now the focus should shift to whether AI "can get work done." The new generation of agent systems can autonomously call upon search engines, databases, and APIs, iterating repeatedly until tasks are completed, achieving "end-to-end delivery"—AI no longer just outputs a block of text for you to organize, but directly helps you finish the job. He even asserts that the traditional IT asset-light era is coming to an end, and data centers will transform from "storage warehouses" into "factories that produce tokens."
Domestic companies have already provided quantifiable evidence. At its 25th anniversary conference, Shanghai Lanmeng disclosed that the company has built a "six-agent system" covering HR, finance, procurement, marketing, business, and middle-office operations, achieving "one person + AI completing in 30 minutes what a traditional team would take a week to do"; in desktop operations and maintenance scenarios, the large-model Agent can automatically handle 45%–90% of common issues; customer service report generation time has been reduced from 90 minutes to under 2 minutes.
Note: Numbers like 45x or 90% are not about how impressive AI is, but rather they give businesses a reference framework that is assessable and replicable. The hallmark of practical AI adoption is not how stunning the technology demonstration is, but whether it can consistently reduce time and costs within real business processes. If it cannot achieve this, no matter how large the model, it is just a toy.
Conclusion
Looking at this week's three events together, the AI transformation of domestic ERP has essentially completed its "puzzle": application vendors like Kingdee are accelerating product upgrades, infrastructure vendors like Inspur are filling in the data foundation, while Jensen Huang and numerous real-world cases are validating the premise that "AI can actually do the work." With these three forces converging, the main theme of enterprise digitalization over the next two years is basically set—whoever prepares their data and business processes first will run ahead in the AI era.
对企业而言，现在的当务之急不是纠结"用哪家的 AI"，而是回答三个更基础的问题：我的数据就绪了吗？我的哪些流程最适合 AI 先跑起来？我选的技术路线在五年内会不会被厂商停售、被市场淘汰？想清楚这三件事，比追任何热点都重要。
