Published on Feb. 21, 2026

The weeks before this year’s China's Lunar New Year holiday became an intense showcase for the country's AI industry, with a cluster of significant model launches arriving in quick succession.

ByteDance, owner of TikTok, released Seedance 2.0, a text-to-video model that can produce a multi-shot film sequence in roughly 60 seconds from relatively simple prompts, while displaying fluid motion and real-world lighting that make it look like a release by one of the major production houses.

The model sparked widespread excitement from Shanghai to Hollywood. It also drew the attention of Feng Ji, creator of the globally acclaimed game “Black Myth: Wukong”. 'The childhood era of AI-generated content is over,' Feng said.

Zhipu AI, listed on the Hong Kong Stock Exchange in January 2026, simultaneously revealed that a mysterious model that had topped rankings on LLM marketplace OpenRouter under the name 'Pony Alpha' was in fact GLM-5, the latest version of its flagship model.

The new open-source model prioritises coding and agentic engineering tasks, integrates DeepSeek “Sparse Attention” to lower deployment costs and features an asynchronous reinforcement learning algorithm enabling it to learn continuously from long-range interactions.

Its performance in practical programming tests is said to be approaching that of Anthropic's Claude Opus 4.5. GLM-5 took the top spot among open-source models on benchmarking site Artificial Analysis within hours of its official reveal, and Zhipu's shares surged by 29 per cent on the day of the announcement.

“If Claude Opus represents the pinnacle of closed-source models, then the launch of GLM-5 undoubtedly marks the 'Opus moment' of domestic open-source models'," wrote Tencent News (in Chinese) .

On Chinese New Year's Eve, Alibaba released Qwen 3.5-Plus, a new open-source foundation model with 397 billion total parameters but only 17 billion activated, achieving performance that surpasses its own previous trillion-parameter Qwen3-Max model.

In benchmark tests, it outperformed Gemini 3 Pro across multiple agent and reasoning evaluations. Notably, the inference cost comes in at just 0.8 yuan per million tokens, about US$0.12, or one eighteenth of the price of Gemini 3 Pro.

Alibaba added to this momentum by launching Qwen-Image 2.0, an open-source image generation model, on February 10. The Qwen model family is also powering AI tools for the International Olympic Committee at the 2026 Milan Winter Olympics.

The back-to-back launches reflect a broader race. China had 602 million generative AI users by the end of 2025, a 141.7 per cent increase in a single year, according to the China Internet Network Information Center.

“It is a Spring Festival for Chinese models before the real Spring Festival for Chinese people,” Bloomberg quoted the fund manager Bao Xiaodong. “The performance gap between Chinese LLMs and US ones is shrinking rapidly.”

Arguably, China's frontier AI companies are no longer merely catching up. They are competing with their US rivals in real time. At the same time, the cost efficiency of Chinese AI computing, which has been hotly debated since DeepSeek’s surprise release of R1 shortly before Chinese New Year 2025, is making new headlines this holiday season again. Put somewhat pointedly, one could say that Chinese AI is becoming better and cheaper.

Galaxea AI, a Beijing-based startup founded in September 2023, completed a 1 billion yuan Series B round in February 2026, lifting its total funding to nearly 3 billion yuan or US$430 million and its valuation above 10 billion yuan or close to US$ 1,5 billion, reported the tech portal Leifeng Wang (in Chinese) .

The milestone makes Galaxea AI the fourth Chinese embodied AI company to reach unicorn status, joining Unitree, Agibot and Galbot. Of those four, it is by far the youngest, having reached the 10 billion yuan threshold in just over two years from founding.

Galaxea AI does not describe itself as a traditional robot manufacturer. Instead, the company positions itself as an 'embodied AI foundation company', focused on building what it calls full-stack capabilities spanning 'brain, body and toolchain'.

In August 2025 it released its G0 model, followed by an upgraded G0 Plus in January 2026, an out-of-the-box Vision-Language-Action model. Its wheeled robot, the R1 Pro, provides hardware support for LingBot-VLA, an open-source embodied AI model developed by Ant Group's embodied AI unit, Robbyant. The company says it holds thousands of orders from universities, research institutes and corporate clients worldwide.

Galaxea AI's rapid ascent illustrates that the AI investment landscape in China is increasingly backing companies that frame themselves as platform providers rather than hardware sellers. This is mirroring the logic that made software-defined businesses so valuable in earler technology cycles. The 10 billion yuan or “unicorn” valuation, achieved in roughly the time it takes most startups to complete a seed round, means that the market is pricing in extraordinary future growth.

When Han's CNC Technology listed on the main board of the Hong Kong Stock Exchange on 6 February 2026, it did so at the top of its offer range, raising HK$4.83 billion (US$619–620 million).

The listing drew ten cornerstone investors committing a combined total of US$309.8 million, among them Singapore's sovereign wealth fund GIC, asset manager Schroders, and Victory Giant Technology, a leading PCB manufacturer. Han's CNC's shares rose more than 7 per cent in the days following the debut.

The company, based in Shenzhen, is China's largest specialised manufacturer of PCB (printed circuit board) production equipment by revenue, with a domestic market share of 10.1 per cent in 2024, according to its prospectus.

Its core business is high-precision drilling and processing machinery used in the manufacture of printed circuit boards, the layered electronic substrates found in virtually every computing device.

The AI boom has proved a powerful tailwind for Han’s CNC. As large AI models and smart hardware applications have proliferated, demand for high-performance servers and GPUs has surged, in turn driving substantial growth in orders for advanced, high-layer-count PCBs capable of handling high-frequency and high-speed signal transmission.

Han's CNC has positioned itself at the centre of this shift. On the day of the listing, the finance portal Sina Caijing reported (in Chinese) that the company has declared AI servers and automotive intelligence as its two primary target segments.

The company has built joint research and development laboratories with downstream customers and says it provides support spanning process validation through to mass production. It aims to become a leading global core service provider of PCB equipment.
Its IPO proceeds are intended to fund accelerated deployment into next-generation AI computing power solutions and deepen its footprint in high-end manufacturing value chains.

This is a prime example of two simultaneous trends. First, the explosive growth of large AI models is reshaping the global electronics industry. AI is not only increasing overall demand for printed circuit boards but also raising the complexity of the PCBs, as the specific requirements of AI computing grow. This in turn is resulting in the rapid embedding of AI in the processes for their manufacturing.

Second, from an investor’s perspective, Han’s CNC’s oversubscribed Hong Kong listing illustrates how AI infrastructure investment is reshaping China's capital markets well beyond the model developers and cloud providers that usually attract attention.

Han’s CNC is perceived as positioned in the sweet spot of AI integration in the PCB industry, which is expected to grow rapidly as automation and AI deployment cascade through China’s factories. This means that the market is now pricing AI exposure into the entire manufacturing supply chain, not just the headline technology companies.

Li Auto will soon launch something new, but it does not want us to call it a car. At least not only that. “The new Li Auto L9 is not just a great car, it is the pioneering work of embodied intelligent robots,” CEO Li Xiang wrote on his social media account this February 2026 (in Chinese).

He wants the forthcoming Li Auto L9 SUV to be seen as far more than a new vehicle, describing a tech stack designed to give the car eyes, a brain, a heart, nerves, hands, and feet.

Li stated that he spends 70 per cent of his time focused on the car itself, arguing that “embodied intelligence must be built on a solid vehicle to truly create value for users.”
The new L9 is priced at 559,800 yuan (around US$81,000), substantially above the existing L9's range of 409,800 to 439,800 yuan, and is equipped with in-house M100 chips and “advanced embodied AI.”

Li Auto called the vehicle an “intelligent agent capable of proactive service,” able to recognise, understand, and anticipate the needs of its occupants.

“Every family deserves such an intelligent partner in the future,” the CEO said, presenting his new SUV like some sort of humanoid companion robot on wheels.

Li Auto shares rose 3.6 per cent in Hong Kong following the announcement, just in time, one is tempted to add. The launch comes at a difficult moment for Li Auto.

The company delivered 406,343 vehicles in 2025, an 18.81 per cent year-on-year decline, making it one of the few new energy vehicle startups to experience a sales contraction. Li Auto has now set a target of roughly 550,000 units for 2026, implying growth of approximately 40 per cent.

Li Auto is late to this particular “all in with AI” rebranding. XPeng and Geely have been making the same move. But to see it purely as a scramble onto the AI bandwagon to impress investors would be to miss that there is a real business case to be made.

A genuine convergence is under way between the automotive and robotics industries, driven by a simple technical reality. The world models that teach a car to navigate a city autonomously are the same ones that can teach a humanoid to walk around a factory floor.

China Southern Power Grid, one of the country’s two main state-owned electricity networks, has released a dedicated “AI+” strategy and begun implementing it.

Working with Huawei, Baidu, and SenseTime, the company has built an intelligent computing cluster with a total capacity exceeding 3,000P in total scale, the business daily Meiri Jingji Xinwen reported (in Chinese).

The grid operator has developed a proprietary model system called "MegaWatt," comprising three foundational large models, 31 domain-specific models and more than 3,000 scenario models. Its natural language processing model is called MegaWatt NLP 2.0.

Southern Power Grid’s strategy illustrates how the “AI+” mandate handed down by Beijing to central SOEs is translating into large-scale infrastructure investment and operational deployment. For foreign technology and energy companies, this is a signal that China's vast power sector is becoming a significant and fast-moving AI application market in its own right.

A team of researchers and engineers in Tianjin is experimenting with the integration of AI, brain computer interfaces (BCI) and wearable devices. The aim is to help stroke patients recover more quickly from post-stroke hemiplegia.

Zhou Peng at the Haihe Laboratory, in partnership with several hospitals and Tianzong Yimai Technology Company, is developing a glove-like device that is guided by brain waves to stimulate pressure points. They call it a “Traditional Chinese Medicine Brain Computer Interface,” reports Cailianshe (in Chinese).

Clinical trials in several large hospitals are underway and will be critical to convincing critics who question Traditional Chinese Medicine’s (TCM’s) scientific foundations.
The Chinese government is not among those critics, however. It has even facilitated experiments with earlier versions of the device on board the Shenzhou spacecraft.

TCM is booming in China. There are around 700,000 TCM practitioners in the country seeing over 1 billion patients every year, as The Economist recently reported. This is also a huge and growing market for Western medical device makers.

Large AI models can also help to better organise and disseminate proven knowledge of certain treatment methods used to augment Western medicine, and may also help to discard some poor examples.

TCM’s holistic approach to human health, for example its use of food as medicine, acupuncture and breathing techniques, is rooted in thousands of years of practical experience treating patients, but the medical knowledge is very scattered and arguably in need of an intelligence upgrade.