China has proposed export controls covering AI models, chips and AI agents, according to a Bloomberg report. The draft rules would, if adopted, give Beijing formal authority to restrict the overseas sale of some of its most commercially sensitive technologies. Details on licensing thresholds, enforcement dates and agency jurisdiction were not disclosed in the report. That absence matters more than the announcement itself.

The proposal targets three layers of the AI stack at once. Models, the trained systems themselves, sit alongside the chips that run them and the agent software that increasingly acts on user behalf. A single regulatory frame spanning all three would mark a shift from China's previous, narrower approach of policing dual-use items through existing catalogues. Bloomberg did not report which ministry would administer the regime. It also did not report whether foreign-owned fabs or model developers operating in China would fall under the rules. Both questions will decide the commercial impact.

What does the proposal actually require?

On the evidence available, very little is confirmed. Bloomberg describes a proposal, not enacted law. Chinese trade measures typically pass through a public comment period before finalization, and earlier technology export rules took months to move from draft to enforcement. The report contains no timeline for implementation. It contains no list of affected chip categories or model sizes. Any claim about immediate export blockages would overstate the source.

What the proposal does signal is intent. Beijing appears to be building legal instruments that mirror, in structure though not in scope, the export control architecture the United States has deployed since October 2022. Whether China's version targets retaliatory leverage or genuine security screening is not stated in the report.

Who is exposed if the rules are enacted?

The draft, as summarized by Bloomberg, touches several constituencies at once:

• Domestic AI developers selling models or inference services abroad.
• Chinese chip designers and their overseas customers.
• Buyers of Chinese-built AI agents, a category now embedded in enterprise software rather than sold as discrete hardware.
• Multinationals that train or host models inside China.

Each group faces a different exposure. Model exports are hard to police because weights cross borders as data, not cargo. Chip shipments are physically traceable and therefore governable. Agents complicate the picture further because they bundle model access, tooling and cloud delivery in one product. The proposal's breadth may prove its enforcement weakness.

Why now, and against whom?

The timing follows years of tightening US restrictions on advanced semiconductors and chipmaking tools destined for China. Beijing has previously used export controls selectively, including curbs on gallium and germanium announced in 2023 and restrictions on rare earth processing technology. Those measures targeted materials where China holds dominant supply. AI models and agents are different. China is a supplier in a competitive global market, not a chokepoint monopolist. Restricting their export risks ceding commercial ground to US and European vendors, which is likely why the controls remain at the proposal stage.

No company was named in the report as a target or a beneficiary. No revenue figures or shipment volumes were attached. Any estimate of market impact would require precisely the numbers Bloomberg's account does not provide.

What happens next?

The proposal will move through China's regulatory comment and revision process, and the final text will likely narrow or redefine the categories now circulating in draft form. Watch for two indicators: the identity of the administering agency, and whether the final rules set technical thresholds, such as parameter counts or compute levels, that would make licensing predictable rather than discretionary. Until then, the controls remain a signal of capability, not a constraint on trade.