Net neutrality, the principle that network operators treat lawful traffic equally, has no formal home in Chinese law. The closest real-world test is a different one: which AI infrastructure can Chinese developers reach, and who decides? Rest of World reported in late September 2026 that Alibaba's ModelScope and OSChina's MoArk are competing to become China's Hugging Face, years after regulators blocked the US-based hub in 2023. This is not a net neutrality ruling. It is a case study in what happens when access to core developer infrastructure is set by discretion instead of by neutral, published rules.
What the report establishes
According to Rest of World, ModelScope, launched in 2022, claims about 170,000 models. MoArk, launched in 2023, claims more than 20,000 commonly used models, per OSChina chief executive Xu Yong. Hugging Face hosts more than 3 million open models. Beijing blocked Hugging Face in 2023 and gave no public reason, and Xu Yong argues that China needs its own AI ecosystem serving Chinese users.
The report also describes a quiet exception. Authorities tolerate AI labs that bypass the block so they can share Chinese models with the world, because total isolation would starve the domestic industry of global connections. In practice, the rule is a block that is enforced selectively.
The strongest case for the state's approach
The case for Beijing's position deserves a fair hearing. Dependence on a foreign-controlled platform is a real supply-chain risk. Rest of World notes that Washington has reportedly discussed banning Chinese models from Hugging Face. A developer community that depended entirely on a US-hosted hub could lose access overnight through a foreign policy decision. Mercator Institute researcher Rebecca Arcesati told the outlet that Chinese policymakers worry access to US-based platforms could be disrupted at any time. Seen this way, building domestic alternatives is ordinary resilience planning, and many governments are making similar moves toward sovereign infrastructure.
The developers' own testimony supports part of this. Rest of World reports that the main draw of domestic platforms is faster downloads, which suggests real value beyond mandate.
Why discretion is still the wrong tool
Resilience is a reason to build alternatives. It is not a reason to remove the choice. A neutral access regime lets a platform win on speed, quality and trust. A blocking regime lets it win by default. Chen Yunfei, a developer in Kunming quoted by Rest of World, put the competitive point plainly: domestic platforms need to convince developers why they should use them. A market where the incumbent is blocked does not force that argument.
The legal architecture shows how access in China runs on permission. The Ministry of Industry and Information Technology's January 2017 notice on cleaning up the internet access service market states that, without approval from the telecom authority, companies may not set up or lease dedicated lines, including VPNs, to conduct cross-border business. That is a licensing gate on cross-border connectivity. It is the opposite of a non-discrimination rule: the default is prohibition, and access is by approval.
The 2023 Interim Measures for the Management of Generative AI Services, issued by the Cyberspace Administration of China and six other agencies, took effect on 15 August 2023. Article 6 encourages independent innovation in foundational technologies such as algorithms, frameworks, chips and supporting software platforms, and promotes infrastructure and public training-data platforms. Article 17 requires services with public opinion attributes or social mobilisation capacity to complete security assessments and algorithm filing. Nothing in these texts names Hugging Face or sets criteria for blocking a foreign model hub. That gap is the problem. Neither developers nor foreign platforms can predict or contest the decision.
The costs of selective enforcement
Tolerating workarounds may look like pragmatism, but it carries costs that a pro-innovation analysis should name.
- Legal uncertainty for individual developers. Rest of World notes that VPN use is common in the tech industry, yet authorities have occasionally penalised citizens for unauthorised circumvention. Tolerance is a favour that can be withdrawn.
- Uneven access. Large labs with resources and quiet official tolerance can reach global models. Smaller developers face a harder choice between slower domestic mirrors and legal risk.
- Weaker competitive pressure. Domestic platforms scale faster when the global benchmark is absent. A 170,000-model catalogue is impressive, but it is still about 5% of the roughly 3 million models on the blocked incumbent, a gap a neutral-access regime would test directly.
What proportionate policy would look like
A proportionate approach to platform security would start from published criteria. If the concern is data security or the integrity of training datasets, regulators can set requirements for hosting, filing and auditing that apply equally to domestic and foreign platforms, and let any platform that meets them operate. That is how the Interim Measures already treat generative AI services in principle. Extending the same logic to model hubs would replace an opaque block with a testable standard.
The lesson also reaches beyond China. The same logic applies if the US moves to restrict Chinese models on American platforms: infrastructure access decided by executive discretion invites retaliation, fragments the open-source ecosystem, and hurts the developers who benefit most from shared tooling. The open internet works because the layers beneath applications do not pick winners.
Net neutrality, narrowly defined, is not on the Chinese policy agenda. The principle underneath it is: lawful, comparable services should be reachable on comparable terms. The ModelScope and MoArk race will produce better domestic platforms. Whether Chinese developers will be free to choose among them and the global alternatives is a policy decision, and Beijing has so far made it case by case.