Sovereign Chinese AI is a strategic reality

China has built Sovereign Chinese AI (SCA): a self-contained stack of silicon, memory, software frameworks, models and deployment infrastructure spanning on-device inference, local systems and cloud platforms, designed to run independent of US control and progressively outside the Nvidia and TSMC supply chains. Beijing’s term for the goal is an “independent and controllable” AI ecosystem [1], and the ecosystem now exists.

The existence of the SCA stack is not a claim of present-day equivalence or superiority. That is the wrong test. The strategic fact is that it now functions. China can train, adapt, deploy and distribute capable AI systems through a stack increasingly outside US control. SCA is on a trajectory to parity with the West, and it was built by deliberate choice.

SCA has not been developed in secret or hidden away. Chinese open models now account for roughly 41% of downloads on Hugging Face, ahead of US models, with Alibaba’s Qwen the most-used open family in the world [2]. DeepSeek’s latest frontier model was optimised first for Huawei Ascend and Cambricon hardware, with early access withheld from Nvidia and AMD [3]. Anyone can download the open weights, run them and benchmark them against US and Nvidia results. A growing share of the wider market, including US startups, now builds on Chinese models as the base layer [4].

That openness is deliberate. Chinese models are not merely competing with US models; they are beginning to dominate open-model adoption pathways, pulling developers, benchmarks, deployment architectures and procurement decisions toward Chinese technology. Adoption comes first. Dependency follows.

The hardware is sovereign too, even where it remains technically uneven. Huawei designs its own AI accelerators, with a published roadmap of Ascend parts through 2028 and an accompanying push toward Huawei-designed high-bandwidth memory (HBM) in place of imported HBM [5]. The Ascend 910C is being produced at meaningful scale as SMIC expands domestic 7nm fabrication and new capacity comes online [6]. Yields, HBM supply, advanced packaging and cluster-scale efficiency still trail the TSMC-centred frontier [6]. These are real constraints. They are also the near-term cost of sovereignty, not evidence that the strategy has failed. Design, memory and manufacture are moving in-house, closing the dependencies the stack was built to remove.

US export controls have accelerated Chinese efforts to advance SCA, but they did not create the strategy. A supply chain that Washington can close at will is itself the vulnerability Beijing set out to engineer away. When Washington reopened a conditional licensing path for some advanced semiconductors in late 2025 and early 2026 [7], Beijing did not simply return to dependence. Nvidia now says it has largely conceded the Chinese AI chip market to Huawei [8], and China’s economic planners have called for pairing domestic models with domestic chips [9]. The drive predates the controls and will outlast any round of them.

AI systems are not value-neutral. A model carries the assumptions, controls and intent of those who build and align it [10]. If the global supply of AI products and services becomes substantially Chinese, then what those systems do, how they do it and who is permitted to use them become questions of Western security. Those questions cannot be answered by assertion, reassurance or wishful thinking. They require independent technical work.

This is why SCA cannot be treated as an issue that can be wished away, or managed with another reflexive export-control lever. It is a standing reality. The question it puts to the UK, the United States and the wider Five Eyes is whether we can understand it on our own terms.

The work of understanding the SCA stack requires people, processes and technology capable of evaluating it (services, apps, tools, models and hardware) within lawful and authorised bounds. The hardware route is narrow but real. Under General Prohibition 10, the purchase or use of controlled Huawei Ascend parts is treated as a presumptive violation worldwide. But BIS has stated that it will not pursue enforcement where a party obtains a PRC 3A090 chip solely for technical analysis or evaluation of an individual chip [11]. That narrow route matters. It is one basis on which a lawful UK capability for SCA security research can be built.

The limit on Western understanding is not Chinese secrecy. It is whether we choose to build expertise and secure access.

Camulos works on AI security, including how open models can be steered [10] and how their behaviour can be probed and characterised [12]. The UK and partners need their own capacity to assess SCA on a defensible technical basis, built lawfully, with the people and access that requires.

If this is your problem, or you think it should be, get in touch.

References

1. China’s drive toward self-reliance in artificial intelligence: from chips to large language models, MERICS.
2. State of Open Source on Hugging Face: Spring 2026, Hugging Face.
3. Three reasons why DeepSeek’s new model matters, MIT Technology Review.
4. The Model Provenance Problem, S. Spillard, Camulos, 2026.
5. Huawei Connect 2025 keynote: the Ascend roadmap and Huawei-designed memory, Eric Xu, Huawei.
6. China’s chip champions ramp up production of AI accelerators at domestic fabs, Tom’s Hardware.
7. Department of Commerce Revises License Review Policy for Semiconductors Exported to China, BIS.
8. Nvidia says it has ‘largely conceded’ China’s AI chip market to Huawei, CNBC.
9. China pushes homegrown AI stack with local chips, LLMs, DIGITIMES.
10. The Père David’s Owl: Subliminal learning and political ideology, D. Martin, Camulos, 2026.
11. Guidance on Application of General Prohibition 10 (GP10) to PRC Advanced-Computing Integrated Circuits, BIS.
12. China’s AI Evolution: DeepSeek and National Security, S. Mercer, S. Spillard and D. Martin, CETaS, February 2025.