Chinese government authorities sent urgent requests to various experts, including professors from Tsinghua University, to provide briefings on the implications of generative AI technology, according to people with knowledge of the matter.
China’s Big Tech firms and ambitious start-ups rushed to roll out their own versions of AI chatbots and large language models (LLMs), as well as register them with the government, as part of efforts to keep American AI services away from the country’s more than 1 billion internet users.
Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.
In the first few months after ChatGPT’s release, maintaining a walled garden approach on AI services was thought to be China’s best strategy until domestic tech firms could develop products that effectively competed against those built by Western AI providers.
Even over a year after ChatGPT’s launch, venture capitalist Allen Zhu Xiaohu said he had no interest in funding Chinese start-ups building LLMs – the technology behind generative AI services like ChatGPT – because they had neither a clear path to monetisation nor data for such a business to prosper.
Zhu, known for his early investment in ride-hailing giant Didi Chuxing, asked rhetorically: “How do you make money out of just developing an LLM?”
Fast-forward to the second half of 2025, and expectations about the technical capabilities of Chinese AI companies and the LLM business have changed.
ChatGPT was released by OpenAI on November 30, 2022. Photo: Shutterstock alt=ChatGPT was released by OpenAI on November 30, 2022. Photo: Shutterstock>
The report attributed this year’s surge in open LLM usage around the world to the growing adoption of Chinese-developed systems, including Alibaba Cloud‘s Qwen family of models, DeepSeek‘s V3 and Moonshot AI‘s Kimi K2. Alibaba Cloud is the AI and cloud computing services unit of Alibaba Group Holding, owner of the Post.
That showed how China’s open-source models were winning the trust of developers around the world. US businesses such as Airbnb and even US tech giant Meta Platforms are now using Qwen.
The turning point for China’s AI sector came when Hangzhou-based DeepSeek released its V3 and R1 models – in December 2024 and January, respectively – that were on par with OpenAI’s GPT and Meta’s Llama models at the time, although they were developed at a fraction of the training costs spent by those US firms.
“One thing we learned from DeepSeek is that open-sourcing the best models can greatly help adoption,” said Baidu co-founder and CEO Robin Li Yanhong in February. That was a stark reversal from his previous comments that described open-source models as inferior to proprietary ones.
The China-US AI gap had narrowed to around three months from more than a year, thanks in part to the influx of AI talent, fast technology iteration and a boom in AI applications on the mainland, according to Tony Zhang, head of China technology research at CLSA, in September.
While Chinese companies faced a shortage of advanced processors for their AI projects because of geopolitical issues, the technology’s development has not been hindered by that constraint, Zhang said.
That resilience also showed the foresight of Chinese cloud services providers, who managed to accumulate enough AI chips for training purposes. The country has also seen a growing number of domestic chip designers develop AI processors for both training and inference tasks, he added.
DeepSeek launched its V3 model in December 2024. Photo: Shutterstock alt=DeepSeek launched its V3 model in December 2024. Photo: Shutterstock>
While the US was focused on supercomputing power and large models in the lofty pursuit of artificial general intelligence and artificial superintelligence, China was adopting a more practical approach of using AI to solve real-world problems, according to Ren.
Before ChatGPT, China had always seen itself as a front runner in AI after Beijing made the technology a national priority in 2017.
The plan outlined a road map for China to become an AI superpower by 2030. It was driven by the prevailing view at the time that the country’s vast trove of data – described as the fossil fuel for AI – would help drive economic and industrial transformation.
But the frenzy that followed ChatGPT’s made the capabilities of those companies obsolete in the era of generative AI.
When ChatGPT quickly grew to become the fastest-growing consumer AI app with about 100 million users in just two months, China’s tech industry faced a period of soul-searching as it reflected on how it missed that breakthrough and how far behind the country was now in AI.
A big screen shows how Megvii’s facial recognition system, Face++, was being used in China’s city streets. Photo: Simon Song alt=A big screen shows how Megvii’s facial recognition system, Face++, was being used in China’s city streets. Photo: Simon Song>
The dim outlook at the time even prompted Zhou Hongyi, founder of cybersecurity firm 360 Security Technology and an industry veteran with close ties to Chinese authorities, to tell an audience at the China Development Forum in 2023 that the country was two to three years behind the US in developing ChatGPT-like technology.
Renewed optimism came in 2024 when mainland media hailed the emergence of China’s new group of “AI tigers”. These included Moonshot AI, Baichuan, MiniMax and Zhipu AI, which had all generated hefty funding from various investors.
The open-source development approach of DeepSeek and Alibaba Cloud’s Qwen has encouraged Chinese AI start-ups to pursue further innovation, despite not having access to advanced graphics processing units from leading US suppliers Nvidia and Advanced Micro Devices.
“Open source is the tool of the challenger,” said Chinese AI expert Jeffrey Ding, an assistant professor of Political Science at George Washington University. “If you look at why Google decided to open source its Android operating system, it was because Apple‘s operating system was dominant at the time.”
Closed AI systems are like proprietary software, as their providers control access, decide what features to offer, and prevent inspection or modification of the underlying source code or model weights, the variables encoding their “intelligence”. Open model developers release their weights publicly, usually along with technical papers outlining their training process, allowing users to deploy and fine-tune them based on their requirements.
Still, the success of DeepSeek did not guarantee smooth sailing for other AI start-ups. Baichuan, along with 01.AI, dropped out of the AI model market since they no longer wanted to charge users to access their inferior products.
There was barely any exciting news from the other AI tigers until Moonshot AI launched its Kimi K2 model in July. It was followed by Zhipu AI, known internationally as Z.ai, when its GLM-4.5 model was announced and made a significant impression among programmers for its coding abilities. In October, Z.ai released a new flagship model, the GLM-4.6, with even stronger coding abilities.
“We had to wait so long between the launch of DeepSeek R1 until GLM 4.5,” said Li Zixuan, Zhipu AI’s head of global operations. “We were underrated, especially in China, while we were just totally ignored in the US”.
Moonshot AI released last month Kimi K2 Thinking, a reasoning variant of its open-source K2 model. Photo: Shutterstock alt=Moonshot AI released last month Kimi K2 Thinking, a reasoning variant of its open-source K2 model. Photo: Shutterstock>
MiniMax pivoted to an open-source approach in June with its M1 model, followed by its M2 model in October, which put the firm in the international spotlight as a frontier open model company.
Moonshot AI and MiniMax’s new models helped put them back on the board, which they never fully left, according to James Wang, general partner at San Francisco-based venture capital firm Creative Ventures.
Wang, who expected DeepSeek to retain the mantle of China’s leading AI start-up when its next-generation flagship model is released, said the rush to release open-source models showed that “the Chinese ecosystem now has its own competitive landscape of model companies vying for supremacy.”
Other companies in adjacent industries now looked to create their own open-source models. These included social media platform RedNote, on-demand delivery giant Meituan and smartphone and electric vehicle maker Xiaomi.
According to Chinese open AI experts Nathan Lambert and Florian Brand, Alibaba Cloud’s strategy with its Qwen family was the closest analogue to US Big Tech companies than any other Chinese AI lab. The firm was releasing many open models across a range of sizes, covering the entire stack from visual and coding models to image and video generation models.
Sebastian Raschka, an AI researcher and author, said Alibaba Cloud’s Qwen3 series of models was the most noteworthy open model release this year, alongside DeepSeek’s R1, because of its strong performance and the practical usefulness of having a variety of sizes.
In an apparent sign that Chinese AI firms have made great strides over the past three years, OpenAI and Anthropic have both publicly admonished Chinese AI firms from DeepSeek to Zhipu AI for alleged security risks.
The administration of US President Donald Trump in June announced an AI Action Plan that placed emphasis on globally diffusing US AI technologies, partly in response to perceived Chinese dominance in the open-source landscape.
So in Silicon Valley and Washington, the question now being asked is the same as the one discussed in China three years ago: Are we losing the AI race?
“Chinese AI engineers no longer envy those abroad,” Huawei’s Ren Zhengfei said at a Shanghai event last month. He noted, however, that China still lagged behind the US in attracting global talent, which would require the mainland to be more open.
Long-standing US restrictions on advanced semiconductor exports have also started to be relaxed, following Trump’s decision to greenlight Nvidia’s shipment of H200 chips to China.
