Two Sovereign AI Stacks, One Beijing Decision
The Download, Not The API
China's Ministry of Commerce, known as MOFCOM, is discussing a rule that draws a line most coverage has missed. Under the proposal, hosted access to a Chinese AI model would stay open to foreign customers. Downloading the model's weights, the trained file that makes the system run, might not. The Financial Times reported July 21, independently corroborated by Reuters, that MOFCOM is consulting three of China's top AI developers, Alibaba, ByteDance and Zhipu AI, on export controls covering model weights, training data and chip designs. The proposed regime is tiered: a simple filing for weaker open models, a security review for stronger ones, and a possible ban on public release for the most capable systems.
The hosted-versus-download distinction sounds technical. An API call rents a model one query at a time, and the file never leaves China. A weight download hands over the model itself, a file a government can run on its own chips indefinitely, no ongoing dependency on the vendor. Alibaba's Qwen passed 1 billion cumulative downloads on Hugging Face by March 2026, per index.dev. By May, Chinese open-weight models made up roughly 61 percent of the tokens routed through OpenRouter, a marketplace tracking which models developers actually run in production, per research from understandingai.org. Free downloads built that share.
Two Governments, Same Bet
Malaysia's national AI system was built by a company called Skyvast Cloud, running on Huawei Ascend chips. Skyvast deployed DeepSeek, the free Chinese model, as the core of the country's Strategic AI Infrastructure, the first national-scale DeepSeek deployment outside China, per Developing Telecoms. Singapore's government-backed program, AI Singapore, built its regional model, SEA-LION, on Alibaba's Qwen rather than Meta's Llama, for the same reason Malaysia picked DeepSeek: Qwen was free, competitive, and downloadable without a licensing negotiation, per CRN Asia. Neither government paid for a hosted API. Both downloaded the weights and built government infrastructure directly on top of the file.
That is the choice MOFCOM's consultation now calls into question. Washington's October 2023 controls decide which Nvidia chips reach China; a MOFCOM export rule would decide which Chinese models can leave China as a downloadable file at all. 'Sovereign,' in both programs' branding, was meant to signal independence from any foreign vendor. Both systems now depend on a MOFCOM filing, review or ban decision instead. China's Commerce Ministry, ByteDance, Alibaba, Zhipu, Huawei, Qualcomm and TSMC all declined to comment, per Reuters. For Skyvast Cloud's engineers and the AI Singapore team maintaining SEA-LION, the operational question sits inside MOFCOM's tiered filing, review and ban regime: which tier their vendor's model lands in, and whether that vendor keeps clearing it.
Nothing in the Financial Times or Reuters reporting says whether a MOFCOM rule, if it lands, reaches backward. Malaysia's national stack and Singapore's SEA-LION are already running on DeepSeek and Qwen weights sitting on servers in Kuala Lumpur and Singapore, not rented from a Chinese cloud each time they run a query, beyond MOFCOM's current reach.
Sources
- China is considering export controls on AI technologies (tomshardware.com)
- China considers tighter export controls on AI models and chips (finance.yahoo.com)
- News: APAC China AI model export controls (techrepublic.com)
- Malaysia launches sovereign AI stack with DeepSeek and Huawei GPUs (developingtelecoms.com)