On page 10 of the November 9 issue of People's Daily Overseas Edition, there is a sentence worth copying down: even a shared code repository will face vastly different audit and access permission rules in different jurisdictions. The context is the 2026 World Internet Conference Wuzhen Summit, which just concluded from November 2 to 5—"Open Source, Openness, Co-building and Sharing" was written into the theme. According to Ant Group's Ni Xingjun in his November 1 remarks, the term "open source" entered the summit theme for the first time, and during the four-day session about 1,800 guests from 138 countries and regions attended (People's Daily summit recap, November 7).
Many people's first reaction is, 'Good news—open source is now easier to use.' My judgment is exactly the opposite: when open source is elevated from a choice within the tech community to a national-level and even global governance agenda, the bill companies have to pay does not get smaller; it shifts from licensing fees to somewhere else—and becomes harder to avoid.
I. First, Get the Data Straight: This Is Not a Meeting for Empty Talk
| Indicator | Numerical value | Scope and Sources |
|---|---|---|
| 2025 China Enterprise AI Agent Market Size | 21.2 billion yuan | The Blue Book of the China Internet Development Report 2026 was released on November 4. |
| As of June 2026, generative AI service filing / application function registration | 924 styles / 573 styles | Same as above, official filing statistics |
| Daily average token call volume in March 2026 | 140 trillion | Same as above, growing more than a thousandfold in two years |
| AI's contribution to the global economy in 2030 | 22.3 trillion US dollars | The projected value in the World Internet Development Report 2026, accounting for 3.7% of global GDP, is not an actual measurement. |
| Global cumulative downloads of China's open-source AI large models | Surpassing 10 billion times | MIIT data, cited by China Economic Net on November 4 |
| Share of downloads of Chinese open-source models | 41% | Hugging Face 2026 Spring Report, range 2025-02 to 2026-02, third-party basis |
10 billion times
Global cumulative downloads of China's open-source AI large models (MIIT data)
More than 11 million
Number of users gathered by the national-level AI open-source community (MIIT definition)
The summit itself also stopped short of making categorical claims. Wang Jian, an academician of the Chinese Academy of Engineering, put it bluntly: open source does not mean no boundaries; how to find a balance between shared collaboration and risk prevention and control is a common question facing the global open-source community. The report 'Frontier Artificial Intelligence Security and Governance 2026: Risk Perception and Collaborative Governance', released at the same time, was tasked precisely with identifying key governance gaps (People's Daily summit overview, November 7).
“In governance, enterprises need self-discipline, government regulation and evaluation, and social oversight to work together, maintaining the vitality of open innovation while consolidating the security foundation.” — Yun Xiaochun, Academician of the Chinese Academy of Engineering and Chief Scientist of Zhongguancun Laboratory, Wuzhen Summit, November 2–5
Put these two sentences together, and the logic becomes clear: what is opened up is code visibility; what is tightened is accountability. Writing “open source” into a national-level theme is equivalent to elevating “who maintains it, who audits it, who signs off” from a tacit community understanding to a public issue that must be put on the table.
II. The Three Extra Lines on the Corporate Bill
Proportion of commercial codebases containing open-source components (Unite.AI, November 2026) — supply chain verification is not optional
Line 1: The supply chain must be verifiable, not just trusted
98% of commercial codebases contain open-source components, and most of the people maintaining these critical projects are volunteers or fit it in alongside full-time work (Unite.AI Chinese edition, November 2026). AI is scaling up vulnerability discovery, so the bottleneck shifts from “can we find them” to “who will reproduce, rate, fix, and coordinate disclosure”—every step requires someone to be accountable, and automation cannot replace that.
Security industry reports use "Patchmageddon" to describe the current situation: the vulnerability exploitation timeline has been compressed by AI to within a day, and the old excuse of "fear that patching will break the business" no longer holds; the direction is to shift to "professional open source" with dedicated security teams and SLAs, and to use artifact proxies (JFrog Artifactory, Sonatype Nexus) to enforce the use of patched versions (AI Understanding Chinese site, November 2026, industry report paraphrase).
Don't translate “open source and free” directly as “zero cost”. What is free is the act of obtaining the code; what you pay for is the SBOM (Software Bill of Materials), artifact signature verification, build provenance traceability, upstream maintenance sustainability assessment, and the SLA that someone will catch you when something goes wrong.
Referenceable engineering practices are already mature: TUF, in-toto, and Sigstore in the OpenSSF and CNCF ecosystems are the public tooling foundation for verifying 'where what you install comes from' (Ostorlab technical newsletter, November 9); Python tool vendor Astral's publicly disclosed practices are more detailed—images use Sigstore to issue cryptographically verifiable release attestations, GitHub immutable releases prevent force-pushing over old tags, the release process does not enable caching to prevent poisoning, the release environment is isolated and requires approval by a second member, and tags are not allowed to be created before a release succeeds; on the dependency side, Dependabot and Renovate are used for updates, new versions are given a 'cooldown period' and are not upgraded immediately, and upstream is sponsored through OSS Fund.
Line 2: Inconsistent permission rules for cross-border collaboration
People's Daily Overseas Edition on November 9 pointed to another category of cost: different regulatory systems, divergent definitions of digital security, and fragmented supply chains create hard barriers to joint open-source development. Applied to ERP and agent projects, this becomes very concrete: in the same repository, configurations that a domestic team can read may not be readable by an overseas team; model weights can be open-sourced, but training data, evaluation sets, and log retention are often constrained by the rules of their respective jurisdictions. "We use open source, so there are no compliance issues" is the most dangerous assumption of all.
Line three: maintainers themselves become a risk exposure
Of the three practices Ni Xingjun offered on November 1, the second is "prioritize open-source governance and sustainability, ensuring long-term project maintenance, community governance, and contributor incentives," and the third is building a virtuous cycle of "contribute—use—contribute again"; his example was Ant Design, with more than 99,000 cumulative GitHub stars (Toutiao, November 1).
The question most easily overlooked during technology selection: if this project had no one maintaining it tomorrow, what would we do? Many enterprises can produce a complete feature comparison table, yet cannot answer how many maintainers a key dependency has, how long its most recent security response took, or whether there is a commercial support party that could take over. This is moving "single point of failure" from the data center to the community.
The community itself has begun writing commitments into verifiable terms. The PGSTY SILO manifesto, last revised on November 8, states: repositories belong to the organization rather than to individuals; the build process is documented with provenance proofs so that anyone can rebuild an equivalent artifact from the source code; maintenance cessation for more than six months is publicly declared and archived; and weakening any clause requires 90 days' prior public notice. What it provides is verifiable actions and deadlines, not attitudes.
Three: "Open does not mean boundless" — what it means for enterprises
Adding the three lines of account together reveals something counterintuitive: the more the summit promotes open source as a public good, the less enterprises can use it as a public good. Public goods are available to all, but responsibility is not automatically shared out; once something goes wrong, the only party that can step forward to explain is the enterprise that connected it to its production systems. At the Forum on the Rule of Law in Cyberspace, Wu Tao, deputy chief judge of the No. 3 Civil Adjudication Tribunal of the Supreme People's Court, offered this methodology: "increase the supply of rules and stabilize behavioral expectations" (People's Daily summit summary, November 7) — not relying on ad hoc decisions each time an incident occurs, but on rules written clearly in advance.
There is another easily overlooked signal: Wang Peng, director of the China Internet Network Information Center, said at the summit that agents autonomously completing the entire chain from requirement understanding to result delivery across systems and entities will enhance the resilience of industrial and supply chains (People's Daily summit summary, November 7). Once cross-entity invocation occurs, what you depend on is not only the components you installed yourself, but also every link in the partners' invocation chain. The boundary of the supply chain is growing longer, not shorter.
Core argument: The 2026 Wuzhen Summit elevated "open source and openness" to a national-level and global governance agenda. What this brings is not "open source is cheaper," but the explicitization of responsibility — license fees are saved, while verification costs, cross-border permission costs, and maintenance sustainability costs are explicitly priced. What enterprises should do is not use less open source, but change open source from a "cost-saving option" in procurement terms into a "verifiable chain of responsibility" in engineering terms.
Four: Three practical recommendations for CIOs and digital leaders
Put SBOMs and artifact signature verification into the acceptance form, not the security white paper. Require components and images entering the production environment to provide an SBOM, and require the release side to issue verifiable signatures and build provenance proofs; set up signature-verification gates in CI, and impose a cooling-off period for new versions. Just copy the Astral approach: immutable releases, isolated release environments, two-person approval, and tag protection.
Build a "maintainer risk table" for key dependencies. Record four items for each core project: the number of active maintainers, the security response time over the past year, whether there is a commercial or foundation support party that can provide an SLA, and the contingency plan for replacement or a self-built fork if maintenance ceases. Lay this table out together with the TCO table at selection meetings. For components that underpin core accounting, permissions, and identity, prioritize "professional open source" with a dedicated security team and an SLA.
For cross-border projects, first separate permissions and data jurisdictions, then discuss code sharing. List item by item which jurisdiction's audit rules apply to repositories, configurations, logs, model weights, and evaluation data; write clearly in the contract the log retention period, export formats, and how data will be taken away upon exit. Then turn "contribute—use—contribute again" into a budget line item: a fixed annual upstream sponsorship or commercial support fee is cheaper than finding someone to put out the fire after an incident.
Open source has never promised to bear responsibility on behalf of enterprises; what it promises is that the code is visible and you can act on it yourself. What really changed during these four days in Wuzhen is that this matter was placed on the public agenda — from now on, when someone asks "who is managing this component," the cost of being unable to answer will be higher than in the past.
Sources: People's Daily Overseas Edition, 2026-11-09, page 10, "Open Source and Openness: Promoting Global Digital Cooperation"; People's Daily, "Open Source and Openness, Co-construction and Sharing — Summary of the 2026 World Internet Conference Wuzhen Summit" (November 7); China.org.cn/CCTV, "2026 World Internet Conference Blue Book Released" (November 4); China Economic Net, "Wuzhen Summit Answers the World: AI Development Should Not Have 'Walls'" (November 4, including data cited from MIIT and Hugging Face); Toutiao (Ni Xingjun's remarks, November 1); Unite.AI Chinese edition and AI Understanding Chinese site (November 2026); Ostorlab technical newsletter (November 9); PGSTY SILO manifesto (last revised 2026-11-08).
Data scope note: USD 22.3 trillion and 3.7% are blue-book forecast values; 41% is on the Hugging Face platform statistical basis; 10 billion times and more than 11 million are on the MIIT basis; 98% and "vulnerabilities appear within one day" are paraphrases from industry reports; Ant Design's 99,000 stars is the vendor's own statement. Different sources have different statistical boundaries and are not aggregated across bases.
Zodioo · Enterprise Digitalization Observations · 2026-11-09
