一篇文章告诉你：国产 GPU 背后的技术和商业路线异同

Core Insights - The narrative of domestic GPUs has shifted from media hype to financial scrutiny, indicating a new phase where companies must demonstrate financial viability and operational capabilities through IPOs [1] - Companies in the domestic GPU sector are now competing not just on technology but also on their ability to sustain operations and deliver products [1] Company Summaries Moer Technology - Moer Technology is positioned as a potential leader in the domestic GPU market, with its IPO approved by the China Securities Regulatory Commission, aiming to raise approximately 8 billion RMB for core projects [2] - The company is pursuing a "fully functional GPU" strategy, targeting various applications including AI training and general computing, with a founding team that includes members from NVIDIA [2][3] - Despite significant revenue growth projections, the company faces challenges with substantial net losses and a high-pressure business model focused on rapid expansion and ecosystem development [5] Muxi Co., Ltd. - Muxi Co., Ltd. aims to raise around 3.9 billion RMB for the development of high-performance GPUs and AI inference GPUs, with its IPO application recently approved [6][8] - The company is focused on creating a general-purpose GPU system that encompasses graphics rendering and AI applications, with existing orders amounting to several billion RMB [6] - Muxi's strategy emphasizes serving high-end computing scenarios rather than consumer graphics, which may position it well in the domestic market [8] Birun Technology - Birun Technology focuses on general-purpose GPUs and data center computing, with a clear strategy to target cloud training and inference rather than traditional desktop graphics [9] - The company has developed its first general-purpose GPU chip, BR100, utilizing innovative packaging technologies to enhance performance in AI training scenarios [9][10] - Birun has completed multiple rounds of financing, raising over 5 billion RMB, and is preparing for an A-share listing [10] Suiyuan Technology - Suiyuan Technology specializes in AI training acceleration cards, focusing solely on this segment without diversifying into consumer graphics [13] - The company has developed its second-generation training card, T20, optimized for large model training, and is working on a new inference acceleration card [13][14] - Suiyuan's business model emphasizes real workload testing and a self-developed software stack tailored to local AI engineering needs [14] Hanbo Semiconductor - Hanbo Semiconductor aims to establish a foothold in data center inference and specific AI acceleration scenarios before expanding to broader applications [16] - The company is developing a matrix of inference cards with a focus on low power consumption and compatibility across various industry frameworks [16] - Hanbo is in the process of initiating an A-share listing, positioning itself as a versatile player in the domestic GPU landscape [16] Glandfi - Glandfi takes a unique approach by focusing on building bridges between system software and GPU microstructures, rather than following conventional GPU narratives [17] - The company prioritizes foundational work in drivers and compatibility layers, aiming to create versatile computing units capable of running various applications [17] - Glandfi's strategy involves gradual development and real-world testing, with plans to transition into AI applications in the future [17] Industry Outlook - The upcoming wave of IPOs for domestic GPU companies signifies a critical juncture where these firms must prove their capabilities in real-world applications and financial sustainability [19]