China's open AI edge may become a bargaining chip, not a gift

For about two years, the best deal in AI came with a Chinese accent: strong models, cheap to run, open for anyone to download. What if that was never charity, but a strategy that is now about to change character?

Beijing's draft puts open weights on the table

China has proposed export controls on 23 AI technologies, according to The Wall Street Journal. The list spans large language models, AI agents, chip design and computing clusters. Bloomberg reports that the proposal would restrict exports of core technology for models trained on "important data" and containing at least 2 trillion parameters.

The developer-facing details matter most. Developers Digest notes that the draft includes model weights, while API use and already-public weights are exempt. It's a draft, not a law, and we should read it that way.

The exemptions are the tell, and also the trap

The carve-outs are reassuring: what you've already downloaded stays downloaded, and calling a hosted API isn't touched. But we think that's the wrong comfort to take. The draft doesn't need to close the door on the past to change the future. It only has to establish that the next frontier release, the one you'll want in six months, is something Beijing can license, delay or withhold.

That is the real shift. Open weights were soft power: a way to get developers worldwide to build on Chinese models and tooling. A control regime turns the same asset into a lever. Release when it helps, hold back when it doesn't. Anyone who built a product on the assumption that each new version would simply appear on a download page has been carrying a risk they never priced in.

Note the criteria, too. Coverage of how they would work together is thin: some summaries describe the "important data" test and the model-size test as if they could be separate, so we wouldn't assume that smaller models are safe. The "important data" language is vague, and vague rules in a trade-control context tend to be applied at the discretion of the regulator. That is our reading, not a certainty.

The chip gap is closing where it hurts least, and that's the point

The second half of the story explains why Beijing might feel comfortable tightening. According to the South China Morning Post, Huawei and DeepSeek are narrowing the AI chip ecosystem gap with Nvidia. Reuters and The New York Times both reported at the end of September that the two are building software tools for Huawei's chips.

Let's not overclaim. An analysis of DeepSeek V4 argues Huawei hasn't replaced Nvidia for frontier training and that CANN hasn't matched CUDA; the realistic win is domestic inference. Another tracker, citing Epoch, puts Huawei's per-chip lag at roughly three to four years, while noting that China is stronger in networking and system scale. Both are caveated estimates.

But inference is where models meet customers. If Chinese open models run well on Chinese silicon, China needs Nvidia less, and has less reason to give away the thing that trains the next generation of its own ecosystem. The loop of model, chip and software gets tighter and more domestic. In our view, that's what makes an export lever credible rather than symbolic.

Treat Chinese weights as a single-source dependency

Don't panic, and don't rip anything out. Do treat Chinese weights like any single-source dependency:

• Keep local copies of every checkpoint and tokenizer you rely on, since public weights are exempt today.
• Design your stack so swapping the underlying model is a config change, not a rewrite.
• Keep a tested fallback from a different region's open models, even if it's a notch weaker.
• Be careful with plans that depend on future releases or on fine-tuning pipelines tied to one vendor's tooling.

The winners from this, in our view, are Huawei and DeepSeek, who could gain a protected home market, and anyone who has already diversified. The losers are teams who mistook a pricing strategy for a public utility. We expect the draft to be negotiated and softened, but the lesson survives either way: open weights from a state-aligned industry are open on terms that can change. Build like you know it.