The most counter-intuitive business story in Chinese technology is also the simplest to state: the country's largest tech companies have become the world's most aggressive open-sourcers of core AI technology.
Alibaba has open-sourced more than 460 Qwen models, most under Apache 2.0. Huawei open-sourced CANN — its entire compute architecture for neural networks, the closest analogue to CUDA in its stack — in August 2025, and committed to publishing all CANN operators on GitCode by late September of that year, with domain libraries, the graph engine, Ascend C and the MindIE inference engine following in December.
Neither company needed to do this to sell cloud services. Both did it because, at a certain point in a platform war, ubiquity is worth more than exclusivity.
Key takeaways
- Alibaba / Qwen: 460+ open models; 2.045 billion Hugging Face downloads in 2026; 151,448 derivative models on the hub; Apache 2.0 licensing.
- Huawei / Ascend: CANN fully open-sourced from August 2025; community at 67 projects and 3,500+ monthly active developers by WAIC 2026; commitment of 1,500 PFLOPS of compute and 30,000 development boards a year to ecosystem work.
- The licensing asymmetry: among Chinese releases above 20B parameters in 2026, about 59% are Apache 2.0 and 22% MIT. Among comparable US releases, about 29% are Apache or MIT, 41% custom terms, 30% no declared licence.
- Business effect: Alibaba reported in March 2026 that Cloud Intelligence Group revenue rose 36% year on year, with AI-related product revenue posting triple-digit growth for a tenth consecutive quarter, with Qwen central to the offering.
The strategy in one sentence
If your model is the default fine-tuning substrate, you own the layer beneath the applications — and the cloud, tooling and services revenue that follows accrues to you whether or not the model is free.
This is not a new idea; Google gave away Android for the same reason. What is new is the speed and completeness with which Chinese companies have applied it to AI, and the fact that they did it while Western labs were moving toward custom licences with field-of-use restrictions.
Hugging Face's own report draws the conclusion explicitly: Qwen has become part of the "default workflow" for developers deciding what to fine-tune and deploy. Defaults, once established, are extremely expensive to displace — a competitor must persuade the entire toolchain to adapt, not merely outperform on a benchmark.
Huawei's harder version of the same bet
Huawei's case is more interesting because the asset being given away is defensive. CANN is the software layer that makes Ascend silicon usable. Open-sourcing it does not attract hobbyists; it attracts the enterprise and academic developers who decide what gets deployed in China's domestic compute build-out.
The results, one year in, are modest in absolute terms — 67 community projects, 3,500 monthly active developers — against a two-decade CUDA incumbency. But Huawei's accompanying commitments (1,500 PFLOPS of compute, 30,000 boards annually) signal that this is a decade-long programme rather than an announcement.
The strategic logic is unavoidable: if China is going to run AI on domestic silicon, someone has to make that silicon easy to program, and the company selling the silicon is the only party with both the incentive and the resources to do it.
What it means for everyone else
For developers outside China, the practical outcome is that the most permissively licensed, best-tooled open models in the world currently come from Chinese labs. That is a fact about licensing and distribution, not about geopolitics — and it is why developers in Southeast Asia, Africa and Latin America, where budgets are tight and legal departments are small, have adopted them fastest.
For Western labs, the response has already begun: Meta and NVIDIA both shipped new open models in 2026, with NVIDIA explicit about competing on open weights. Competition in openness is a genuinely good outcome for the global developer community.
For China's companies, the risk is the mirror image of the opportunity. Policymakers have weighed restrictions on overseas access to the strongest domestic models — a measure that would protect a strategic asset by giving up the reach that produced it.
Sources: Hugging Face State of Open Models (August 2026); Huawei disclosures at the August 2025 Ascend summit and HUAWEI CONNECT 2025; Alibaba Cloud earnings, March 2026.