HANGZHOU, China — In what is sometimes dubbed China’s Silicon Valley, tech giants and startups in Hangzhou are racing to build cutting-edge chips, robotics and brain-computer interfaces. Across town, aspiring founders are creating AI pets and fortune-telling apps. The city on China’s southeastern coast is reinventing itself as an artificial intelligence hub. A year after the rise of DeepSeek drew global attention, Hangzhou’s fast-moving AI scene reflects both global ambitions and intensifying competition at home. The sprint toward physical AI China and the U.S. are moving in lockstep to develop what many see as AI’s next frontier: physical applications. From Meta to Tencent , companies are racing to build “world models” that use AI to help robots move, steer driverless cars, or even simulate real-world events such as climate change. Beijing identified “embodied intelligence” as one of China’s future priorities in its upcoming Five-Year Plan. In November, the U.S.-China Economic and Security Review Commission, which provides recommendations to Congress, urged Washington to step up investment and regulatory approvals for autonomous systems and robotics , warning that China is making rapid advances in physical AI. In Hangzhou, spatial intelligence firm Manycore, as well as robot makers Unitree and Deep Robotics – among the city’s tech startups known as the “six little dragons” – are preparing to go public in Hong Kong or mainland China, joining a flurry of AI-related listings. While large language models are trained mainly on internet data to generate text and images, training AI for the physical world requires additional inputs such as weight, texture, structure, temperature, and how objects interact, including the grip or force needed to handle them. Instead of manually collecting all data points, using AI methods to calculate them will make the process much easier and faster, Manycore cofounder Victor Huang told CNBC. Huang, a former Nvidia software engineer, says his company uses the California-based company’s chips, saying they deliver superior computing power per watt – performing best for the energy they consume. But he believes China’s cheaper energy provides an edge , potentially offsetting the use of less advanced, more power-hungry chips. For instance, a three-nanometer chip can reduce power use by about 30% compared with five- or seven-nanometer chips, but companies can stay competitive if their electricity costs are 40%-50% lower, he said. Computing power, he added, cannot be viewed in isolation. It depends on data quality, energy supply and operating conditions. Locating data centers in colder regions can save energy, while cleaner data reduces the computing power needed for training. Manycore also open-sourced its spatial AI model – an approach China has embraced over many of the pay-to-use models common in the U.S., such as those offered by OpenAI and Anthropic. This approach allows the company to gather feedback, Huang said, but limits direct revenue because users are not required to pay for access. “So you’ll get pressure from investors,” he said. An indie AI scene takes shape Compared with Silicon Valley’s fascination with artificial general intelligence or superintelligence, China’s AI push has focused heavily on practical applications, from Baidu Map’s personalized recommendations to ByteDance’s chatbot and assistant Doubao. In December, Doubao was China’s top AI app with 155 million weekly active users, nearly double that of its closest competitor, DeepSeek’s chatbot, according to business intelligence service provider QuestMobile. Doubao’s popularity suggests that user experience and utility may trump technical sophistication. A looser, experimental culture is emerging alongside these efforts. While Hangzhou heavyweights like Alibaba and DeepSeek are hard at work on cutting-edge AI, a leafy, hipster-meets-hacker suburb called Liangzhu has become ground zero for AI’s quirkier side. Residents of Liangzhu call themselves “villagers,” building everything from gamified fitness trackers to ADHD-friendly calendar tools. Alex Wei, an entrepreneur who moved to Liangzhu in 2025, is developing an AI app based on traditional Chinese divination tools. He’s interested in how AI can meet people’s emotional needs. Liangzhu’s low commercial pressure is part of its appeal. “You can come to Liangzhu with 1,000 renminbi ($143) and leave with your product demo,” Wei said. “It’s a very inclusive place. You don’t have to have a unicorn product — you’ll find support even for a small app serving a thousand people,” he added. The groundswell of innovation in Liangzhu has caught the attention of investors. A monthly “Demo Day” that started in a living room now draws founders and backers from across the region and beyond. That exposure is shaping how startups think about scale. Many are eyeing overseas users, with some founders hoping to tap into China’s hardware supply chains to offer competitive prices worldwide. Intense domestic competition and consumers’ reluctance to pay for apps have also pushed startups to look abroad, observers say. Afra Wang, who writes the Concurrent newsletter on China and Silicon Valley, said some developers are using AI to escape traditional career paths in changing job markets, working instead to become “superindividuals” who build profitable businesses independently or with tiny teams. But Wang warned that some businesses are simply patching AI features for marketing appeal, from air conditioners to mirrors that check sunscreen application. She calls some of these “physical AI slop” – playing on the term for low-quality content cranked out by generative AI. For now, Hangzhou’s entrepreneurs are testing nearly every idea, from efficiency to comfort, serious to the frivolous, in a rapidly evolving market.
