Hong Kong
—
When China’s biggest artificial intelligence players gathered for a landmark meeting in Beijing in January, one question was in the spotlight: What are the chances of a Chinese AI firm overtaking US frontrunners in the next three to five years?
The answer from a top AI scientist present at the gathering was surprisingly blunt: “Below 20 percent,” said Justin Lin, technical lead for Chinese tech giant Alibaba’s Qwen AI models. “And I think 20 percent is already very optimistic.”
The sobering assessment stood in stark contrast to a year of headlines celebrating China’s AI boom.
Since little-known startup DeepSeek shocked the world with a powerful AI model it said was built at a fraction of the cost of American equivalents, Chinese companies have topped global downloads for freely-available-to-use models and raised huge sums in market debuts.
Yet despite the fanfare, some leading Chinese AI developers have warned that China may have fallen further behind in developing frontier models. Experts point to restricted access to advanced chips and limited capital as lingering constraints.
Lin was not alone. Tang Jie, founder of one of the prominent Chinese AI startups Z.ai, also known as Zhipu, said the performance gap between Chinese and US models “may be widening.”
“In some areas we may be doing fairly well, but we also need to acknowledge the challenges and gaps we still face,” he said at the same Beijing meeting.
But that assessment does not mean China’s AI industry is stagnating.
Constraints in accessing high-performance chips and capital as well as the country’s unique tech ecosystem have fueled a divergent strategy from the US – making AI models available for public use, or open source.
The strategy, which Beijing and developers see as a way to accelerate progress and compete with US rivals, has helped Chinese companies make notable gains. Companies are aggressively rolling out AI applications based on these models for real world uses. Industries are integrating the technology in manufacturing, e-commerce and robotics.
In a televised new year address, Chinese leader Xi Jinping hailed the country’s fast rising innovative capabilities, citing AI models that are “racing ahead” and what he called “breakthroughs” in homegrown chips, as Beijing pushes for technological self-reliance.
Qwen, for example, overtook Meta’s Llama last September as the most-downloaded open model on Hugging Face, a major platform for AI models and tools. Even American firms like Airbnb have embraced it to power AI customer service.
An unprecedented wave of Chinese AI listings has also unfolded in Hong Kong. In January, unicorn startups Z.ai and MiniMax, both developers of competitive open models, went public, raising $560 million and $620 million respectively, with their share prices surging.
Global tech companies are taking notice. In December, Meta announced it would acquire Manus, an AI agent firm founded in China and later relocated to Singapore. While the deal has triggered Beijing’s regulatory review process – raising the possibility it could be unwound, it is a testament to the advances in Chinese AI technology.
DeepSeek, which emerged as a poster child for China’s AI boom after releasing an open model a year ago that delivered near industry-leading performance with far fewer resources, is also reportedly set to unveil a new model with enhanced coding capabilities later this month.
A key driver of China’s AI momentum stems from its aggressive embrace of open models, a shift sparked by DeepSeek’s breakout success. Since then, Chinese AI labs have jumped on the open-model bandwagon, in contrast to their US rivals who have largely eschewed that approach in order to protect their intellectual property.
The turn to open models in China “has dramatically reduced costs for developers and enterprises,” said Poe Zhao, analyst and founder of Hello China Tech, a newsletter focused on the country’s tech industry.
“Cloud providers like Alibaba use open models to drive cloud adoption, startups use openness to build developer ecosystems quickly.”
The boom of open models in China prompted their global usage to skyrocket from a miniscule 1.2% in late 2024 to as much as nearly 30% last year, according to a late 2025 study by OpenRouter, a marketplace for AI models.
Alibaba alone, for example, has released more than 400 open Qwen models, yielding over one billion downloads by earlier this month, according to the company.
“Open sourcing has to a certain extent become an industry consensus in China,” said Lian Jye Su, chief analyst at Omdia, a technology research firm.
The strategy not only reduces costs but also lessens Chinese companies’ exposure to geopolitical risks.
“If one day the entire company were to be sanctioned (by the US), then at least the products based on its open-sourced models could still be used by others,” Su said.
There is also a hard economic reality behind the strategy.
“There’s no other choice. The reason is that Chinese people, like consumers, as well as enterprises, do not pay for software,” said Jenny Xiao, a partner at Leonis Capital, an AI-focused venture capital firm.
But while Chinese models reign supreme in the open model space, closed models developed by US heavyweights – including OpenAI’s GPT, Google’s Gemini and Anthropic’s Claude – continue to dominate overall performance benchmarks. Closed models still account for roughly 70% of total downloads, according to OpenRouter.
Bottlenecks remain
The performance gap between Chinese and US models, while not enormous, persists largely due to constraints on computing power and capital.
“OpenAI, Anthropic and other American companies are putting massive compute into next-gen research, whereas we are relatively stretched thin,” Alibaba’s Lin said.
Washington’s export controls have barred Chinese companies from obtaining cutting-edge chips, such as Nvidia’s Blackwell and the recently launched Rubin series, as well as the equipment needed to manufacture them. Turning to domestic chipmakers for less advanced semiconductors also doesn’t fix the problem for these AI model developers.
“Their problem is they can’t get the volume,” said Paul Triolo, a China and tech expert at consulting firm Albright Stonebridge, as US restrictions also hamper Chinese chipmakers’ ability to scale output.
While Trump has approved the export of Nvidia’s H200 chips, two generations behind Rubin, China has yet to officially greenlight their import. Triolo said this puts Beijing in a difficult bind, as it weighs the near-term need for advanced chips against its push for self-reliance.
DeepSeek, Alibaba, and other Chinese tech giants ByteDance and Tencent have been granted conditional approvals by Beijing to purchase a certain amount of H200s, Reuters reported last month, citing anonymous sources.
Unlike US startups that can raise multiple rounds of venture capital funding, Chinese AI companies face a thinner investor base and growing pressure to demonstrate commercial viability quickly. That is forcing companies like Z.ai and Minimax into public listings ahead of their American competitors.
“A lot of these Chinese companies are trying to get out as soon as possible, because they’re burning a lot of cash and getting into public market is the easiest way to do that,” Xiao of Leonis Capital said.
Adding to the challenge is China’s domestic market, which is smaller and often demands heavy customization, making it harder for AI model developers to turn a profit, she added.
Still, experts caution against underestimating China’s long-term prospects.
Chinese companies have excelled at rapid deployment in consumer-facing applications and integrating AI into industrial use, said Deepika Giri, head of AI research at market research firm IDC.
Even with limitations in China’s AI models, the technology is “becoming more ubiquitous,” she said. “It may not be the best (model) available globally, but still they’re integrating, so that the industrialization of AI is progressing extremely rapidly.”
Beijing has also made AI application a focus, unveiling an action plan to deepen the use of AI in manufacturing last month, part of a national strategy to upgrade its industrial sector.
From open models to real-world applications, China has shown that constraints have not stopped it from gaining ground.
Yao Shunyu, a former OpenAI researcher who recently joined Tencent as their chief AI scientist, said China has repeatedly demonstrated an ability to “very quickly catch up or replicate” western technological developments and in some areas “go on to do it even better.”
He pointed to the country’s powerful manufacturing sector and electric vehicle production as examples in his remarks at the same summit with Tang and Lin in Beijing.
For Yao, however, the bigger challenge for China seeking to pull ahead of the US is cultural: a shortage of risk-takers, despite an abundance of top-tier talent.
“Can we actually lead the creation of new paradigms? I think this is, in some sense, the one key problem China still needs to solve,” he said.
