天数智芯启动招股 定价144.6港元/股

Core Viewpoint - The company TianShu ZhiXin is set to launch its IPO on the Hong Kong Stock Exchange, marking a significant milestone as China's first general-purpose GPU company to achieve mass production of both training and inference GPUs [2][3]. Group 1: IPO Details - TianShu ZhiXin will begin its IPO subscription on December 30, 2025, and end on January 5, 2026, with plans to list under the stock code "9903" on January 8, 2026 [2]. - The company aims to issue 25.43 million shares at a price of HKD 144.6 per share, targeting a total fundraising amount of approximately HKD 3.7 billion, resulting in a market capitalization of HKD 35.442 billion [2]. - A cornerstone investment team of 18 top investors has been established, with a total subscription amount of HKD 1,583,195,754, including notable firms such as ZTE Corporation (Hong Kong) and UBS AM Singapore [2]. Group 2: Company Overview and Product Offering - TianShu ZhiXin is recognized as the first company in China to achieve mass production of both training and inference general-purpose GPUs, with its "TianGai" training GPU and "ZhiKai" inference GPU series already in large-scale delivery [3]. - As of June 30, 2025, the company's products have been deployed in over 290 clients with more than 900 actual deployments, and the revenue for the first half of 2025 reached CNY 324 million, reflecting a year-on-year growth of 64.2% [3]. - The company emphasizes that the valuation of domestic general-purpose GPU companies is shifting from short-term profit focus to sustainable technology assets and delivery capabilities, driven by national strategies on computing sovereignty [3]. Group 3: Strategic Focus Areas - The company identifies four core breakthrough areas necessary for advancing domestic AI chips: architecture and precision optimization, software ecosystem enhancement, deep scene adaptation, and supply chain and capacity assurance [4][5]. - The focus on architecture includes self-developed general computing architecture to enhance computing density and memory bandwidth [4]. - The software ecosystem aims to unify programming models and optimize operator libraries to shorten model migration cycles [4].