作者丨骆轶琪 编辑丨倪雨晴 2025年上半年，国产AI芯片持续获得热捧，多家厂商业绩持续高增。在旺盛的AI推理需求支撑下，国 内头部AI芯片厂商左手积累产成品，右手储备关键材料，以应对未来市场发展。 21世纪经济报道记者梳理公司财报发现，总收入规模与海光信息还有一定差距的寒武纪，在上半年却与 前者的净利润规模已经十分接近；海光信息在上半年相比同业，有较大比例花费用在了销售上；龙芯中 科虽然盈利能力依然面临压力，但在市场周期回暖进程中，旗下核心产品毛利率表现已经逐渐回归，近 期其携"三剑客"进入市场，接下来有望看到进一步商用的好消息。 当然，国内AI芯片厂商的份额仍有较大提升空间。从沐曦股份披露的回复问询公告可见，政策支持对 国内市场采用国产AI芯片有一定驱动，但更大程度还是来自于内生竞争力提升和国内开源生态的密切 联动。 订单与存货"齐飞" 从收入规模看，分别有CPU和DCU（GPGPU的一种）业务的海光信息一直都在稳健增长态势中。上半 年海光信息实现营业收入54.64亿元，同比增长45.21%；归母净利润12.01亿元，同比增长40.78%。 同时，在旺盛的AI推理需求支撑下，正式摆脱"多年亏损"帽子的寒武 ...

作者丨骆轶琪 从收入规模看，分别有CPU和DCU（GPGPU的一种）业务的海光信息一直都在稳健增长态势中。上半年海光信息实现营业收入 54.64亿元，同比增长45.21%；归母净利润12.01亿元，同比增长40.78%。 同时，在旺盛的AI推理需求支撑下，正式摆脱"多年亏损"帽子的寒武纪追赶势头猛烈。上半年实现营业收入28.81亿元，同比增长 4347.82%；归属于上市公司股东的净利润10.38亿元，上年同期为亏损5.3亿元，同比增长295.85%。 寒武纪无论从总资产还是营收规模都与海光信息有一定差距，但在净利润层面实现了快速接近。 寒武纪在公告中提到，净利润大幅 增长源于营业收入增长，期内公司在持续拓展市场，助力人工智能应用落地。 编辑丨倪雨晴 2025年上半年，国产AI芯片持续获得热捧，多家厂商业绩持续高增。在旺盛的AI推理需求支撑下，国内头部AI芯片厂商左手积累产 成品，右手储备关键材料，以应对未来市场发展。 21世纪经济报道记者梳理公司财报发现， 总收入规模与海光信息还有一定差距的寒武纪，在上半年却与前者的净利润规模已经十分 接近； 海光信息在上半年相比同业，有较大比例花费用在了销售上；龙芯中科虽 ...

Group 1 - The U.S. government has imposed export controls on NVIDIA's H20 chip, requiring licenses for sales to China, which indicates a significant tightening of trade regulations in the AI chip sector [2][6] - NVIDIA's CEO Jensen Huang visited China again, expressing a desire to continue collaboration, highlighting the impact of these export restrictions on the industry [1][2] - The restrictions on H20 and similar AI chips from AMD and Intel create a substantial opportunity for domestic Chinese AI chip manufacturers to capture market share previously held by NVIDIA [6][14] Group 2 - Huawei's Ascend 910C chip is currently the most prominent domestic alternative, with its latest generation products being integral to China's AI infrastructure [7][14] - The Ascend 910C chip achieves a computing power of 800 TFLOP/s (FP16) and a memory bandwidth of 3.2 TB/s, which is approximately 80% of NVIDIA's H100 performance [7][10] - Despite its advantages, the Ascend 910C faces challenges such as increased power consumption and potential interconnect bottlenecks, which could hinder its efficiency in large-scale AI training tasks [9][11] Group 3 - The production of the Ascend 910C chip relies heavily on TSMC for manufacturing, as domestic foundries like SMIC struggle with yield rates and production capacity [12][13] - The supply chain for critical components, such as HBM memory, is also a concern, as it involves complex logistics and potential legal issues [13][14] - A diverse ecosystem of domestic AI chip companies, including Alibaba, Baidu, and Tencent, is essential for reducing risks and ensuring stability in the AI chip market [14][15] Group 4 - Companies like Cambrian and Moore Threads are emerging as significant players in the AI chip market, with innovative designs and capabilities that can compete with NVIDIA's offerings [15][16] - The development of a unified software platform by Moore Threads aims to facilitate the transition for users from NVIDIA's CUDA to their own architecture, enhancing user experience and performance [16] - The collaboration among various domestic firms is crucial for building a resilient and self-sufficient AI chip ecosystem in China, moving away from reliance on uncertain supply chains [17]