Why did China stop giving away the strategy that just won?

Tencent's Hunyuan lab released Hy3 — a 295-billion-parameter model — on July 6, under the Apache 2.0 license, software's most permissive: anyone, anywhere, can use it, modify it, and resell it without paying Tencent anything. It was a reversal. April's preview version had explicitly excluded the European Union, the United Kingdom, and South Korea from using it at all. Tencent quietly dropped that clause for the official release.

The next morning, Reuters reported something that made Tencent's timing look less like generosity and more like a closing window. China's Ministry of Commerce, the news agency said, had spent the past month in private meetings with Alibaba, ByteDance, and the startup Z.ai — discussing whether models like this one, future ones, should be allowed to leave the country at all. A tiered system was floated: basic open-source tools would get a simple filing, more advanced technology would face a security review, and the most capable frontier models would be barred from public release or limited to domestic use only. Nothing has been decided. The sourcing is three people "not authorised to speak," declining to be identified. And the reported scope matters: this would apply to future models, not the ones already in the world.

The direction makes more sense once you look at how the last eighteen months actually got cheap. DeepSeek disclosed a training cost of $5.576 million for the run that produced R1 — 2,000 Nvidia H800 GPUs, chips specifically weakened to comply with US export controls, run for roughly 180,000 GPU-hours. A comparable Western model, built without that restriction, is estimated to have cost $80–100 million on 16,000 uncapped H100s. The chips America restricted to slow China down are the same chips DeepSeek was forced to use efficiently — mixture-of-experts routing, 8-bit precision, a reinforcement-learning method that skips a redundant second model — and that efficiency didn't just make R1 cheaper to train. It made it cheaper to run. Common coding tasks that cost roughly $10 on a US frontier model run under $0.50 on DeepSeek: a twentyfold gap on the single most common enterprise AI workload today.

Credit the caveat, not just the headline: $5.576 million was one training run, by DeepSeek's own account, excluding the failed experiments and architecture research that came before it. Independent analysts put the company's total GPU spending closer to $1.3 billion. The efficiency is real. The bargain-bin price tag on a single line item is not the whole story.

Still, the efficiency was real enough to change the economics of an entire industry — and here is the part that isn't really about chips at all. China doesn't need to win the business of selling AI models. It has bet that it can't, and that trying would be a waste of the advantage it does have.

"Smart companies try to commoditize their products' complements."

— Joel Spolsky, "Strategy Letter V," 2002

Spolsky was writing about software companies in 2002, describing a pattern that predates AI by decades: when the price of the thing that goes with your product falls to zero, demand for your product rises. Microsoft licensed Windows cheaply to any PC maker so hardware would become commodity and Windows would become the toll booth. IBM published the PC's technical specs so a thousand manufacturers would compete on price beneath IBM's own margins. China's AI labs, largely state-supported and explicitly measuring success at the national level rather than the balance-sheet level, are running the identical play at the scale of a country: give away the model, so demand rises for the chips, the cloud infrastructure, and the standards China is building underneath it.

That is the strategy Reuters says Beijing is now reconsidering — not because it failed, but because it worked completely enough to change the calculation. A strategy of openness makes sense exactly as long as giving things away buys you more than keeping them would. Eighteen months ago, China's models needed the world's attention. This month, Chinese models account for 45% of traffic on OpenRouter, up from under 2% a year earlier — attention it no longer needs to buy. What a strategy has to prove before it's rewarded is different from what it has to protect once it's won.

Nothing here reverses what already happened. DeepSeek's weights, Qwen's weights, Hy3's weights are downloaded, forked, and running on servers Beijing does not control, in a form no future policy can retract. Whatever gets decided about the next model, this generation already left.