The clause in China's 2017 AI plan that nobody translated

倡导开源共享理念. Advocate the concept of open source and sharing. That line sits in a State Council document dated 20 July 2017, seven and a half years before DeepSeek put R1 on Hugging Face under an MIT licence and much of the industry decided something shocking had just happened.

Nothing shocking happened. A document said what would be done, and then it was largely done. The question worth asking is not why Chinese labs give strong models away. It is why anyone reading Chinese in 2017 expected otherwise.

The clause the English coverage skipped

I read the 新一代人工智能发展规划 in the original, the way I read most Chinese policy, because the English summaries compress it into a headline about 2030 and drop the machinery. The machinery is the whole point. Buried in the section on international cooperation is this: 鼓励人工智能企业参与或主导制定国际标准，以技术标准走出去带动人工智能产品和服务在海外推广应用.

Encourage AI enterprises to participate in, or to lead, the setting of international standards, so that technical standards going out carries AI products and services into overseas markets.

Two words do the work. 主导 means to lead or dominate, and it is deliberately set against 参与, to take part. The plan is not asking companies to attend the committee. It is asking them to chair it. Then 走出去, going out, which is not a casual phrase in Chinese policy writing. It is the name of the outbound investment doctrine formalised at the turn of the century, the one that put Chinese contractors and telecoms kit across Africa, Latin America and Central Asia.

The 2017 plan borrows that vocabulary and points it at technical standards rather than at products. The standard is the export. The product rides behind it.

That reframes everything about open weights. A standard nobody adopts is a document; a standard everybody builds on is a tax. Releasing weights under Apache 2.0 or MIT is the cheapest known mechanism for making your architecture, your tokeniser, your chat template and your tool-calling format into the thing other people's engineering assumes.

A standard nobody adopts is a document; a standard everybody builds on is a tax.

What the trillion-yuan figure actually counts

Now the number everyone repeats. You will have seen China's AI plan described as a $150 billion programme. Look at what the figure attaches to.

The plan sets three-step targets. By 2020, core AI industry above 150 billion yuan. By 2025, above 400 billion. By 2030, 人工智能核心产业规模超过1万亿元 and 带动相关产业规模超过10万亿元: core industry past 1 trillion yuan, related industries past 10 trillion. At 2017 exchange rates that first trillion lands near $150 billion.

It is a target for how big an industry should become. It is not a budget, not an appropriation and not money any ministry undertook to spend. Every headline calling it China's $150 billion AI investment has taken a projected market size and reported it as a cheque. The error has been reproduced for nine years.

Real spending is harder to pin down and rather more revealing. Beijing sets direction; provincial and municipal governments do most of the actual funding, through industrial parks, subsidised compute and talent schemes, and those budgets are scattered across thousands of local documents rather than published in one line.

The clearest recent figure came in June 2026, when Bloomberg reported that the NDRC and other agencies were drafting a roughly $295 billion, five-year plan for a national network of interconnected computing hubs, mostly operated by China Mobile and China Telecom, with a requirement that at least 80% of the technology including AI chips be domestically supplied, and the sites stitched into one network by 2028. That total excludes private spending by Alibaba and Tencent. It also sits well below the roughly $725 billion American hyperscalers earmarked for AI in a single year.

So the shape of it is not that China outspends. It is that a large share of Chinese money goes into shared substrate: power, interconnect, domestic silicon and freely redistributable weights. Substrate is the thing you build a standard on.

Weights already downloaded cannot be switched off

The second driver is a trust problem, and it is not China's trust problem to solve. It is the buyer's.

At an Afro-Tech Forum session in Shenzhen last year, a founder over from Lagos ran his demo off a laptop with the wifi disabled. Deliberately. His point was not that the model was good, though it was adequate. His point was that his last vendor had changed regional availability with two weeks' notice and his product had gone dark in a market where he had already signed customers.

That is the calculation across a lot of the world right now. If you are billing in naira and paying for tokens in dollars, per-token pricing is a currency exposure you cannot hedge. If your government is not on anyone's allied-nations list, availability is a policy variable someone else controls. Weights sitting on a disk in Ikeja are none of those things. They are a file. Nobody revokes a file.

The concrete version of this is Sunflower, from Sunbird AI in Uganda. Fourteen researchers led by Benjamin Akera, paper posted to arXiv in October 2025, models at 14B and 32B, covering 31 Ugandan languages and since extended further across the continent. It is built on Qwen 3. Not because Qwen 3 is the strongest model in the world, but because it was the strongest model they were allowed to open up, retrain and keep.

Read the 2017 clause again with that in mind. 以技术标准走出去. The standard went out. It arrived in Kampala wearing a Ugandan name.

Be sceptical of the numbers the labs quote about themselves

Two figures get thrown around as proof that Chinese labs have broken the economics of frontier AI. Both are real. Neither means what it is used to mean.

DeepSeek's R1 paper made the cover of Nature on 18 September 2025, the first major large language model to go through independent peer review, and it stated a training cost of $294,000. That number covers the reinforcement learning stage: roughly 512 H800s for about 198 hours on R1-Zero plus some eighty hours more, and around 5,000 GPU hours generating fine-tuning data. It does not cover the V3 base model underneath it, which took about 2.79 million GPU hours on 2,048 H800s and was costed at $5.576 million. And that $5.576 million is itself only the final pre-training run. It excludes the prior research, the failed architectures, the ablations and the hardware.

The Qwen adoption numbers have the same character. In August 2026 Alibaba announced 3 billion downloads and more than 300,000 derivative models. Hugging Face's own count for the same period was about 2.045 billion downloads and 151,448 derivatives. Both can be true: Hugging Face measures its own hub and cannot see ModelScope, API traffic or private mirrors. But the honest version of the claim is the smaller one, and the smaller one is still remarkable. 151,448 derivatives is roughly 2.6 times Meta's entire footprint on the platform.

Use the checkable number. It wins the argument anyway.

Alibaba kept Qwen-3.7 for itself

Here is where my own thesis gets uncomfortable, and it should.

In July 2026 Alex Colville documented the turn for ASPI: Alibaba, the company whose openness is the strongest evidence for everything above, made Qwen-3.7 proprietary. API only. The company said the following iteration would be opened again, and by most accounts it was, but the precedent is set and everyone in the industry noticed. ByteDance has run a split policy for a while, releasing some models while keeping its flagship image and video generation closed.

The intellectual scaffolding for this already exists in Chinese. Tang Jie, co-founder of Z.ai, wrote in People's Daily in May 2025 that open source and closed source should be treated as 两条腿, two legs, one serving development and one serving national security. That is a very economical way of reserving the right to close anything, at any time, on grounds nobody outside the room gets to evaluate.

Which means the honest formulation is narrower than the one usually offered. Chinese labs do not open their models. Chinese labs open the models where the strategic return on adoption exceeds the strategic return on exclusivity. So far that has covered almost everything, because almost nothing has been far enough ahead to be worth hoarding. The moment something is, the calculus flips, and the policy language to justify flipping it was published a year in advance.

The strongest objection, which is that this is just what second place does

The best counter-argument is not that the policy reading is wrong. It is that the policy reading is unnecessary. Open weights are simply what any lab does when it cannot win on capability, and no State Council document is required to explain it. If you are behind, commoditising the layer you are behind on is the obvious move. Give the model away, damage the margin of the leader, build position in the layer above.

The precedent is right there and it is American. Mark Zuckerberg published Open Source AI Is the Path Forward alongside Llama 3.1 in July 2024, arguing that open models stop power concentrating in a few companies. The argument was elegant, widely believed and made for reasons of competitive position. In April 2026 Meta shipped Muse Spark, its first closed frontier model. The manifesto lasted under two years. Nobody should have been surprised by that either.

So the sceptic's case is strong: the doctrine follows the position, and when the position changes the doctrine is quietly retired. Alibaba closing Qwen-3.7 is exactly the behaviour that theory predicts.

What that account cannot explain is timing and breadth. Meta's manifesto came after it was clear Meta was not going to lead; the Chinese clause was written in 2017, before there was a frontier to be behind on, and it specified the mechanism rather than the sentiment. It named standards leadership as the objective and named open sharing as the instrument. Nor does the competitive story explain why open releases have come from labs in very different positions at once, from a hedge fund spinout to a state-adjacent university lab to the cloud arm of a company that already had a strong closed offering.

And there is the American Action Plan of 23 July 2025, which called for leading open models founded on American values, and for allies to be building on American technology. Washington did not conclude that openness was a loser's strategy. It concluded that whoever's weights everyone builds on sets the defaults. Which is the 2017 argument, arrived at eight years later, in English.

The part that is not settled

The August 2025 AI Plus directive set adoption targets that make the intent plain: new-generation smart terminals and agents past 70% penetration by 2027, past 90% by 2030. Diffusion, not capability. Being unavoidable rather than being first.

The tension nobody has resolved is what happens when a Chinese lab clearly holds the best model in the world. Every stated justification for openness assumes it does not. Tang Jie's second leg is standing there for precisely that day, and the security framing means the decision gets made by people who will never publish their reasoning.

There is a Nigerian version of this worry that I hear more often than the geopolitical one. A team in Yaba builds three years of product on an open Chinese base, then the base goes closed at the exact moment their market matters. They will have inherited the tokeniser, the fine-tuning pipeline, the eval harness and the habits. Migration cost is not a licence question.

That is what the 2017 plan called 走出去, and it is doing what it said it would. The file on the disk is genuinely yours. The shape of your thinking about what a model is, less so.

Sources

Full Translation: China's New Generation Artificial Intelligence Development Plan (2017), DigiChina, Stanford : https://digichina.stanford.edu/work/full-translation-chinas-new-generation-artificial-intelligence-development-plan-2017/

国务院关于印发新一代人工智能发展规划的通知 (original Chinese text), gov.cn, 20 July 2017 : https://www.gov.cn/zhengce/content/2017-07/20/content_5211996.htm

China Plans $295 Billion Investment to Build Nationwide AI Data Centers, Bloomberg, June 2026 : https://www.bloomberg.com/news/articles/2026-06-09/china-prepares-295-billion-plan-to-fund-nationwide-ai-buildout

Secrets of DeepSeek AI model revealed in landmark Nature paper, 17 September 2025 : https://www.nature.com/articles/d41586-025-03015-6

DeepSeek didn't really train its flagship model for $294,000, The Register, 19 September 2025 : https://www.theregister.com/2025/09/19/deepseek_cost_train/

Qwen is the world's most downloaded open model, by a smaller margin than Alibaba says, TNW, 16 August 2026 : https://thenextweb.com/news/alibaba-qwen-downloads-hugging-face-open-models

Some Chinese AI will go closed-source, but not all, Alex Colville, ASPI Strategist, 22 July 2026 : https://www.aspistrategist.org.au/some-chinese-ai-will-go-closed-source-but-not-all/

Sunflower: A New Approach To Expanding Coverage of African Languages in Large Language Models, arXiv:2510.07203 : https://arxiv.org/abs/2510.07203