In July 2025, a Chinese AI startup launched the world’s largest open-weight AI model, sending a signal heard clearly in Silicon Valley, in Brussels, and across the research institutions of Southeast Asia. The moment was not an accident. It was the product of years of deliberate investment, talent cultivation, and a strategic orientation that three men \u2014 Liang Wenfeng, Tang Jie, and Yang Zhilin \u2014 had been quietly building toward for the better part of a decade.<\/p>\n<\/p>\n
China has emerged as a structural competitor in the global AI race, not merely a fast follower. Its companies are releasing models \u2014 Kimi K3, GLM-5.2, DeepSeek V4 \u2014 that hold their own against US alternatives. Understanding what drives that competition requires understanding the people at its centre.<\/p>\n<\/p>\n
The story of DeepSeek begins not in a tech campus but in the world of quantitative finance. Liang Wenfeng, born in 1984 in Zhanjiang, Guangdong province, graduated from Zhejiang University and co-founded quantitative hedge fund High-Flyer in 2015. Under his direction, the firm’s assets grew more than tenfold in four years.<\/p>\n<\/p>\n
Liang did not spend those profits on conventional expansion. Before US AI export controls on China tightened, he channelled High-Flyer’s earnings into more than 10,000 Nvidia chips \u2014 a decision that would prove foundational. When DeepSeek was established in 2023 as High-Flyer’s AI arm, it inherited precisely the computing infrastructure that most Chinese startups were scrambling to secure.<\/p>\n<\/p>\n
The January 2025 release of DeepSeek’s R1 model landed with the force of a structural disruption. The company claimed R1 matched the performance of leading US models at a fraction of the cost. Silicon Valley did not dismiss the claim \u2014 it debated it urgently. Weeks later, DeepSeek’s latest V4 model followed, with the company citing “drastically reduced” operational costs.<\/p>\n<\/p>\n
The financial consequences for Liang were significant. His net worth rose from US$16.7 billion to US$36 billion following DeepSeek’s first external funding round, and now stands at approximately US$37.9 billion \u2014 placing him ahead of Anthropic co-founder Dario Amodei (US$7.98 billion) and OpenAI president Greg Brockman (US$25.5 billion) on Bloomberg’s Billionaire Index, which focuses on founders whose primary revenue derives directly from AI models.<\/p>\n<\/p>\n
That funding round \u2014 raising around US$7.4 billion from venture capitalists \u2014 marked a significant departure. Liang had deliberately avoided external financing, bankrolling DeepSeek through High-Flyer’s profits alone. Rising operational costs and reported efforts to develop a proprietary inference chip appear to have shifted that calculus.<\/p>\n<\/p>\n
Liang rarely speaks publicly. In a 2024 interview with Chinese publication Anyong, he offered one of his clearest articulations of purpose: China, he argued, had grown too comfortable leveraging overseas technological innovations and commercialising them through applications, while contributing little to core invention. DeepSeek’s mission, he said, was to push the technological frontier forward \u2014 not to ride on others’ achievements, but to become a genuine contributor to them.<\/p>\n<\/p>\n
Where Liang came to AI through finance, Tang Jie arrived through decades of academic computer science. Born in 1977, Tang joined Tsinghua University’s Computer Science department in 2006, building a research career that produced more than 100 published articles and the invention of AMiner \u2014 a search system that mines academic publication data to map relationships between researchers, conferences, and institutions.<\/p>\n<\/p>\n
In 2019, Tang co-founded Zhipu AI \u2014 now rebranded as Z.ai \u2014 alongside fellow Tsinghua faculty and alumni. The company grew into a significant provider of chatbot infrastructure for Chinese enterprises, with its GLM series of models earning genuine respect in developer communities. Its latest release, GLM-5.2, continues that trajectory.<\/p>\n<\/p>\n
Tang’s academic lineage runs directly into the next generation of Chinese AI leadership: one of his past students at Tsinghua was Yang Zhilin, who would go on to found Moonshot AI. In 2021, Tang led the development of Wu Dao, China’s first superscale natural-language AI model. His team also released what they described as China’s first “virtual student” \u2014 a personification of Wu Dao capable of composing music, reasoning, responding to emotional cues, answering questions, and writing code.<\/p>\n<\/p>\n
His current ambition is broader still. On his academic website, Tang states he is putting “all my efforts” into artificial general intelligence \u2014 teaching machines to think like humans.<\/p>\n<\/p>\n
In an internal memo to Z.ai staff this month, Tang framed the moment in terms that left little room for ambiguity. A “great wave” of technological evolution had arrived, he wrote. The progression from artificial general intelligence to artificial superintelligence was irreversible. Z.ai’s core objective for the next two years would be to reach the next high ground in AGI technology \u2014 and the company would deliberately slow its short-term monetisation of applications to concentrate all core resources on that goal.<\/p>\n<\/p>\n
“The wave is here, the trend is irreversible. Zhipu AI must be the one who faces the wave and reaches upwards,” he reportedly wrote. “If we do not reach the summit, we have failed.”<\/p>\n<\/p>\n
The rhetoric echoes Liang’s, but the institutional context differs. Lin William Cong, president’s chair professor of Finance, Computing and Data Science at Nanyang Technological University, situates both men within a broader structural shift. “China’s tech sector has long been very practical \u2014 arguably too focused on commercialisation and application, while free-riding on foundational research done elsewhere,” he told CNA. The current generation of Chinese AI labs represents, in part, a deliberate correction \u2014 an attempt to demonstrate that Chinese firms can produce original innovation, not simply execute fast-follow strategies.<\/p>\n<\/p>\n
Professor Cong is careful, however, not to frame this as anti-commercial. “DeepSeek’s aggressive pricing and Zhipu’s enterprise business are commercial acts \u2014 they’re just subordinated to the research mission rather than driving it.” The combination of open weights, low prices, and a research-first orientation, he argues, reflects a calculated bet: that making frontier technology cheap enough for an entire ecosystem to experiment on top of it is the fastest route to both AGI and market scale.<\/p>\n<\/p>\n
Yang Zhilin is, at this particular moment, the most visible of the three. His company’s July 2025 release of Kimi K3 \u2014 the world’s largest open-weight AI model \u2014 generated a wave of international attention that extended well beyond the usual AI research circles.<\/p>\n<\/p>\n
Yang’s formation as a researcher is distinctly transnational. He completed his undergraduate degree at Tsinghua University under Tang Jie’s supervision, then pursued a PhD at Carnegie Mellon University, graduating in 2019. His doctoral supervisor, Ruslan Salakhutdinov, described him publicly as “absolutely brilliant.” He subsequently interned on the AI research teams at both Meta and Google.<\/p>\n<\/p>\n
Apple reportedly expressed interest in hiring him. Instead, Yang returned to China and founded Moonshot AI in 2023. The decision attracted considerable commentary in both Chinese and international technology media \u2014 a question of what the US had failed to offer, and what China had succeeded in making attractive.<\/p>\n<\/p>\n
Professor Cong identifies several structural factors at play. China’s top AI labs can assemble large, young, highly motivated teams from a deep STEM talent pool. Operating domestically gives founders direct access to one of the world’s largest consumer and enterprise markets, backed by strong state policy and funding. On the US side, visa precarity and the broader climate of US-China technological competition add friction for researchers weighing their options.<\/p>\n<\/p>\n
For someone of Yang’s profile, however, the calculus may also involve something less tangible. “Building China’s first frontier lab carries a kind of historic significance that a senior role at a US company simply can’t offer,” said Professor Cong.<\/p>\n<\/p>\n
Moonshot AI has grown rapidly. According to a Reuters report, the company raised more than US$2 billion in May 2025 from investors including Meituan, China Mobile, and CPE. Its total historical fundraising exceeds US$5.5 billion, and its valuation reached US$30 billion in June. Yang has said Kimi’s team numbers around 300 people \u2014 lean by the standards of frontier AI development.<\/p>\n<\/p>\n
The demand response to Kimi K3 was immediate and overwhelming. Two days after launch, Moonshot AI paused new subscriptions. “Kimi K3 has received far more love than we expected, and our GPUs are feeling it,” the company posted on X.<\/p>\n<\/p>\n
Yang’s personal website, last updated in 2024, describes his work in a single sentence: “I am working on a startup.” The understatement is deliberate. He is a committed music fan, and Moonshot AI’s offices reflect that \u2014 conference rooms named after Radiohead, Led Zeppelin, Queen, and Nirvana; subscription tiers named after classical tempo markings (Adagio, Andante, Moderato); the company itself originally named, in Chinese, after Pink Floyd’s The Dark Side of the Moon<\/em>.<\/p>\n<\/p>\n On China’s talent infrastructure, Yang has been direct. “After decades of accumulation, China has built one of the best talent systems in the world \u2014 it can cultivate the world’s people with the strongest exploratory capabilities, and it does so in a systematic way,” he said, as reported by Beijing Daily.<\/p>\n<\/p>\n Across their public statements and institutional choices, Liang, Tang, and Yang share a notable orientation: a commitment to open-weight model releases and accessible pricing. The convergence is not coincidental.<\/p>\n<\/p>\n Professor Cong frames it as structurally rational for a challenger. “Giving the model away is the fastest way to win global developer adoption and set de facto standards \u2014 it commoditises the layer where US labs charge a premium, which pressures their business model without your having to beat them head-on.” The strategy also fits China’s constraints under US chip export controls, and leverages its enormous domestic developer population. Open weights allow that population to build on frontier systems rather than concentrating innovation inside a small number of closed labs.<\/p>\n<\/p>\n The distinction between open-weight and fully open-source, however, matters. Professor Cong notes that for most Chinese AI models, the weights are downloadable but the training data and full training recipe remain closed. Releases are also sometimes staggered \u2014 API access first, weights later \u2014 which reveals that commercial timing continues to shape how open these systems actually are.<\/p>\n<\/p>\n Kong Tuan Yuen, a research fellow at the National University of Singapore’s East Asian Institute, points to a geopolitical dimension. By providing high-performance models at significantly lower costs, China has positioned itself as a central access point for emerging economies seeking affordable AI infrastructure. That is a form of influence, not merely a pricing strategy.<\/p>\n<\/p>\n It also creates a sustainability problem. Since core products are essentially free, Chinese AI startups will need to develop fundamentally new business models to remain viable over the long term, Dr Kong argues.<\/p>\n<\/p>\n The open-weight approach carries a further risk: safety. Publicly releasing core model components allows third parties to remove or modify safety features. “This makes it easier for malicious users to misuse the models, as well as increases the difficulty of regulating global AI governance issues such as accountability and data security,” said Dr Kong.<\/p>\n<\/p>\n Professor Cong raises a concern that extends beyond any single company or country. As labs and nations race toward artificial general intelligence, the competitive structure itself creates dangerous incentives. “Everyone has a reason to move faster and cut corners on safety, and no single player wants to be the one who slows down,” he said. “For something with stakes this high, we owe it to ourselves, and frankly to future generations, to treat getting it right as more important than getting it first.”<\/p>\n<\/p>\n The three men at the centre of China’s AI rise have built institutions that are, by any serious measure, operating at the global frontier. Whether the sprint they are leading \u2014 and the broader race it is part of \u2014 leaves adequate room for that question is one that no model release, however impressive, can answer on its own.<\/p>","protected":false},"excerpt":{"rendered":" China’s AI Founders and the Race to the Frontier In July 2025, a Chinese AI startup launched the world’s largest open-weight AI model, sending a signal heard clearly in Silicon Valley, in Brussels, and across the research institutions of Southeast Asia. The moment was not an accident. It was the product of years of deliberate […]<\/p>\n","protected":false},"author":4,"featured_media":3967,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","stackable_optimized_css":"","stackable_optimized_css_raw":"a:0:{}","blocksy_post_meta_options":"a:1:{s:17:\"styles_descriptor\";a:3:{s:6:\"styles\";a:3:{s:7:\"desktop\";s:0:\"\";s:6:\"tablet\";s:0:\"\";s:6:\"mobile\";s:0:\"\";}s:12:\"google_fonts\";a:0:{}s:7:\"version\";i:8;}}"},"categories":[19,24],"tags":[],"class_list":["post-3968","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-technology","category-trending"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":8}},"custom_fields":{"stackable_optimized_css":"","stackable_optimized_css_raw":[],"blocksy_post_meta_options":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":8}}},"yoast_head":"\nOpen Weights, Strategic Calculation, and the Safety Question<\/h2>\n<\/p>\n