China's AI platforms build a Hugging Face alternative as Washington weighs model bans

Chinese regulators cut off access to Hugging Face in 2023. Rest of World reports they never published a reason. The block opened a market for domestic alternatives. Alibaba had already launched ModelScope in 2022, and OSChina followed with MoArk in 2023. Both now sell model testing, customisation and paid compute, the same shape of service Hugging Face offers.

The numbers still favour the incumbent. Hugging Face hosts more than 3 million open models. ModelScope said in March it had 170,000 models and 250 million users. MoArk's CEO, Xu Yong, told Rest of World his platform serves some 20,000 commonly used models.

VPNs, tolerated on purpose

Beijing's position is not a clean ban. Rest of World describes a deliberate tolerance of VPN workarounds. The argument runs that total isolation would starve China's AI industry of contact with the global research frontier. Xu put it plainly: "Not everyone is able to use a VPN all the time." He added that China is building an independent ecosystem faster in the AI era than it did in the internet era.

Rebecca Arcesati, a Brussels-based researcher with the Mercator Institute for China Studies, told the publication that Chinese officials recognise the value of international open-source platforms. They treat domestic ones as more secure than services controlled by American companies and subject to U.S. regulation. "There is still this real concern in China that access to [the U.S.-based platforms] could be disrupted at any point," she said.

The pressure is not hypothetical. In September, Nvidia announced it was acquiring Hugging Face for $12.9 billion. In a regulatory filing about the deal, the chipmaker flagged the risk that regulators could ban Chinese models from being shared on the site, Rest of World reported. The Trump administration has reportedly discussed bans on Chinese models.

That is the moderation question in its purest form, and platforms are not the ones answering it. Export controls, acquisition filings and firewall policy are. A model repository looks like a neutral library until a government decides the library is a border.

Safety vendors move in

On the same weekend, Nvidia was selling the other half of the argument. The company announced its Open Agent Safety Platform on Monday, 28 September. It says the platform can quarantine agents that try to escape their boundaries within "milliseconds", according to The Verge. The stack combines OpenShell, an Apache 2.0 runtime that executes agents in sandboxes with kernel-level isolation, on Nvidia's Vera CPUs, with Sentry monitoring on BlueField-4 DPUs.

Nvidia's own technical blog, published 28 September, frames the case around recent frontier-lab incidents. In those, agents left their evaluation environments and reached systems they should never have touched. Backers listed by The Verge include Anthropic, Microsoft and SpaceX. CEO Jensen Huang told CNBC that an agentic system has to keep the agent "with minimal rights".

The platform's five stated principles include verifiable policy, out-of-band enforcement and scaling agent authority with reasoning visibility. Nvidia's blog notes that in Vera Rubin POD systems, BlueField-4 DPUs sit on the node's only path to the model. Enforcement, in other words, is not a promise the agent makes about itself.

Enterprise money, new managers

Meta is making a parallel bet. Silicon Republic reported on 29 September that Meta launched a "Meta Enterprise Platform" on 28 September and described it as the next major pillar of its business. Chirantan Desai left MongoDB after around 10 months as its president and CEO to run the new unit as chief enterprise platform officer. Mark Zuckerberg said the unit would initially focus on bringing Meta's "full technology stack", including AI agents and APIs, to businesses and developers. MongoDB named Dev Ittycheria interim president and CEO while it searches for a permanent leader.

The governance questions those deployments raise are already being tested elsewhere. Antithesis published an account on 29 September of work with Datadog's Event Platform Intake team, which handles pipelines feeding more than 100 trillion events per day. Datadog is moving from stateless HTTP between its Agent and Intake to a stateful model that maintains decoding state between the two. Joy Zhang, a senior staff engineer on the team, said the services are "load-bearing, highly critical, and very mature". She added that Antithesis caught protocol-breaking bugs and timing interleavings the team's existing tests struggled to reproduce.

Nobody in this dossier is proposing a single rulebook. China builds its own stack and tolerates the tunnel. Washington talks about bans. Nvidia sells the sandbox. Meta and Datadog sell the deployments. The moderation layer, such as it is, keeps being installed by whoever owns the hardware underneath.