Why China gives away its open-source AI models

China builds more solar in a quarter than most countries have built in their whole history. It also burns more coal than anyone ever has, and it’s still approving new coal plants at a record pace.

Both true. Both sitting in the same government press releases.

The usual read is that China’s leading the world on climate. And the build-out is real. China makes most of the planet’s panels, batteries, turbines, EVs. But “they care about the climate” is a thin reason to spend on that scale.

Here’s what’s actually paying for it. Energy security. China imports most of its oil and gas through straits it doesn’t control. Hormuz, Malacca, sea lanes full of other people’s navies. A solar farm in a Chinese desert, running on Chinese panels, can’t be blockaded by anyone. The climate win is real. It’s a passenger, not the driver.

I think the same thing is happening with China and open-source AI. The popular explanations catch some of the noise but miss the structure under it. So ask the solar-panel question again. What’s the energy security here? What’s the hard interest hiding behind the obvious one?

Two stories you’ve probably heard

Story one is generosity. Chinese labs keep shipping capable models you can download for free. DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, Zhipu’s GLM, and a long tail behind them. American frontier labs mostly don’t. The contrast is real.

By early 2026, Chinese open-weight models were about 41% of large-language-model downloads on Hugging Face, against 36.5% for the US, and by mid-2026 they were running something like a third of the world’s model traffic. Qwen by itself is closing in on a billion downloads. So yes, something genuine is being given away.

A small but load-bearing distinction. These are open-weight, not fully open-source. You get the finished model to download and run, not the data and recipe behind it. It matters later.

Story two is sanctions. US export controls choked off the top-end chips, so Chinese labs couldn’t win on brute force. They competed on efficiency instead, out-iterating on architecture and inference cost. Also real. And it’s why they opened up. A second-best model behind a paywall goes nowhere, because why pay for it when the best model is a click away and the good-enough ones are free? Give the same model away, though, and it gets used everywhere. When your edge is reach instead of depth, you open the doors.

Both stories tempt you because both are half-right. The first explains the volume. The second explains the timing. What’s left over is the coherence, the puzzle of why so many Chinese companies, with very different core businesses, reached for the same move.

Generosity doesn’t coordinate an industry. And sanctions don’t explain why Alibaba, sitting on plenty of compute, plays the same hand as DeepSeek, which has none to spare. To see that, follow the money. It’s parked one layer over from the model.

Spolsky, with Chinese characteristics

Back in 2002, Joel Spolsky wrote a short essay called “Strategy Letter V.” One idea in it has aged really well. Smart companies try to commoditise their complements. If you make your money in layer A, you want layer B, the thing people buy alongside A, to be cheap as dirt. Cheap B sells more A.

IBM opened up the PC, let clones commoditise the hardware, and made its money elsewhere. Meta gave away Llama and still sells ads.

Now look at China.

The model’s the layer they’re happiest to give away, because none of them sells models for a living. The money sits above it and below it.

Alibaba’s real business is cloud, the biggest in Asia. Qwen is free, in 100-plus flavours, and it’s really sales for Alibaba Cloud. Run it on your own box and Alibaba’s out nothing. Run it on theirs and you’re a paying customer. The model is the on-ramp.

Huawei’s real business is silicon, the Ascend accelerators Beijing’s been steering domestic labs toward for 2 years. DeepSeek’s V4, previewed in April 2026, reportedly runs best on Ascend, not Nvidia. A free model that flies on Huawei’s Ascend chips is free advertising for them.

Tencent sells WeChat and games, ByteDance sells attention, Baidu sells ads and cloud. None of them needs the model to make money. It was never their business, just an ingredient.

DeepSeek is the purest case. It charges for API calls, but at a tenth or less of OpenAI or Anthropic, and hands the weights out on top. That only makes sense if the model was never the product. It’s the doorway. The bill gets run up inside, on chips and cloud.

This is what the generosity and sanctions stories miss. The same logic works on its own for Alibaba, for Huawei, for DeepSeek, with no meeting required. Each is just being rational. The incentive did the coordinating, because none of them is in the model business to start with.

The race to free

Spolsky’s move assumes you can push the model toward free. In 2026 it barely needs a push. By the Artificial Analysis numbers this month, the top frontier model scores about 60 on its intelligence index (higher is smarter) and costs roughly $3.25 to run one benchmark task. A leading Chinese open model scores in the mid-40s and runs about 6 cents. The gap in smarts is real but small. The gap in price is more than fifty to one, and it’s widening on purpose. DeepSeek made a 75% cut permanent in May, and Alibaba shipped a new Qwen tier in June at a fraction of its flagship’s price.

Most work doesn’t need the top of the curve anyway. Summarising, tagging, routing, drafting, the boring middle of the job, all run fine on a mid-tier model for a cent or two on the frontier dollar. So the free Chinese models land where the real demand is, and the American labs are left selling the thin, expensive top inch fewest buyers need.

And the cheaper it gets, the more people burn. Drop the price and they build agents that loop 100 times where someone used to ask once, and that demand runs on the layers China’s been locking down, its clouds and its own chips. Make the model free and the value just moves to the floors above and below, the ones China already owns.

Open weights, closed values

So Chinese models end up everywhere, cheap and good enough to run on almost anything. And distribution is its own kind of power. Whoever supplies the base models supplies the defaults baked into them. Chinese models come with particular defaults.

Ask a Chinese open model something touchy, Tiananmen, Xinjiang, Taiwan, whether the Party should be running things, and it gets careful. It demurs. It changes the subject. Chain-of-thought leaks from DeepSeek have shown instructions to dodge Party taboos. Put under audit, Qwen keeps its answers about China “positive and constructive,” playing up the wins and skipping the criticism. Chinese rules require exactly that. A model that won’t toe the line can’t ship at home, and a model you can’t ship is dead weight.

The strange part is that the politics travels with the file.

The open label covers the first two. The other four stay with the lab, and the values ride along with every copy.

Light fine-tuning changes plenty, a new language, a new domain, a house style of its own. What it doesn’t reliably do is strip out the behaviour set during post-training, the stage the lab never publishes. There’s no neat values file to delete. The defaults are spread through the weights.

So when an Indonesian edtech shop builds on Qwen, a Lagos engineering course teaches with DeepSeek, a Brazilian fintech ships Kimi, the defaults ride along. Most of those developers have no clue which guardrails they just inherited. They shipped a worldview with their product and paid nothing for it.

This is the part “open source” quietly hides. The phrase carries 30 years of goodwill, freedom and the right to fork and change whatever you want. Open weights give you less than that. You can read the model and run it, but rebuilding its behaviour your way means retraining at a scale hardly anyone can afford. The right to fork stops where retraining gets too expensive. What you get is a sealed box. You can inherit its politics, you just can’t open it up and read them.

Fair warning, this lands on American open models too. Llama and Mistral carry values, just different ones. The difference here is that China’s are written into law, required of any model that ships at home. Whether spreading them abroad is the plan or just a side effect, the result is the same. Defaults set by Chinese regulation ride out with the models, and the world inherits them without choosing.

The model was never the point

A free model never travels alone. It carries a business strategy, and now a worldview.

The giveaway is real. So is the driver under it, the same way energy security sat under the solar panels. Every free download feeds the cloud and chips China sells, and ships political defaults that outlast any fine-tune.

That’s how the substrate gets chosen. The cheapest option wins by default. And it never arrives empty-handed.