<p>百度和华为技术在 2025 年上半年共同占据了中国基于自主研发图形处理单元（GPU）的云服务市场超过 70% 的份额，随着对计算能力的需求上升和一波上市计划的涌现，自主研发的人工智能芯片逐渐获得 traction。\n 根据周日发布的 Frost &amp; Sullivan 报告，这些数据将 “GPU 云” 提供商定义为控制从专有 AI 芯片到计算集群和云服务的完整芯片到云的价值链的公司。\n 报告指出，百度占据了 40.4% 的市场份额，华为紧随其后，占 30.1%。\n 这些发现反映了中国领先的 AI 芯片开发商（包括百度、华为和阿里巴巴集团控股）之间的更广泛转变，因为它们不仅限于单一芯片设计，而是致力于构建完整的硬件和软件能力，以期与 Nvidia 的 CUDA 生态系统竞争，并满足对 AI 训练和推理的激增需求。阿里巴巴拥有《南华早报》。\n 报告称：“通过系统创新，他们将成千上万或数万个自研芯片整合到一个庞大的虚拟计算资源池中。”\nFrost &amp; Sullivan 警告称，中国的国内 GPU 市场仍面临 “硬件性能、软件生态系统、系统集成和商业运营” 等限制，这意味着目前只有有限数量的供应商实现了大规模部署。\n 然而，报告指出，未来三到五年，国内 GPU 将从 “战略备份” 转变为国家 AI 产业的 “核心支柱”，并补充说，这些芯片在实际工作负载中表现出相对于通用芯片的竞争性性价比。\n 百度的领先地位得益于其 AI 计算平台百戈与其芯片部门昆仑芯开发的 GPU 的整合，市场研究公司 IDC 表示，昆仑芯在 2024 年出货近 70,000 个单位。\n\n 昆仑芯成立于 2012 年，最初是百度的一个内部部门，随着时间的推移逐渐独立。\n 在其最新的市场份额扩展尝试中，昆仑芯去年 11 月推出了两款新的 AI 加速器。百度表示，M100 将显著提高混合专家模型的推理效率，而 M300 则旨在训练具有万亿参数的超大多模态模型。\n 该报告发布之际，国内 AI 芯片设计师（从百度到 Moore Threads Technology 和 Biren Technology）受到北京推动半导体自给自足和对计算能力需求激增的驱动。\n 百度上周加入了这一浪潮，表示已秘密提交了昆仑芯的香港 IPO 申请，同时指出发行的规模和结构尚未最终确定。\n 总部位于上海的 GPU 制造商 Biren 上周五在香港开始交易，筹集了 58.8 亿港元（约合 7.2 亿美元）。其上市紧随 Moore Threads（被称为 “中国的 Nvidia”）和上个月在上海上市的 MetaX 集成电路等同行的强劲表现。\n 另一家总部位于上海的 Biren 竞争对手 Enflame 上周也完成了其 IPO 辅导过程，距离在上海的纳斯达克风格的科创板上市更进一步。\n</p>

K3SD

华为

<p>根据 Frost & Sullivan 的报告，百度和华为主导了中国的 GPU 云市场，在 2025 年上半年占据超过 70% 的市场份额。百度以 40.4% 领先，其次是华为的 30.1%。报告强调，AI 芯片开发者正在向全栈能力转型，以与 Nvidia 竞争。尽管国内 GPU 市场面临挑战，这些芯片预计将成为中国 AI 产业的关键。百度已为其芯片部门昆仑芯申请了香港 IPO，加入了如比仁科技等其他公司，参与由对 AI 计算能力需求驱动的 IPO 浪潮</p>

<p>Baidu and Huawei Technologies collectively accounted for more than 70 per cent of China’s market for cloud services built on home-grown graphics processing units (GPUs) in the first half of 2025, as home-grown artificial intelligence chips gain traction amid rising demand for computing power and a wave of public listing plans.\nThe figures, published in a Frost &amp; Sullivan report released on Sunday, defined “GPU cloud” providers as companies that controlled the full chip-to-cloud value chain – from proprietary AI chips to computing clusters and cloud services.\nBaidu held a 40.4 per cent share, followed by Huawei with 30.1 per cent, the report said.\nThe findings reflect a broader shift among China’s leading AI chip developers – including Baidu, Huawei and Alibaba Group Holding – as they move beyond individual chip design to build full-stack hardware and software capabilities, in a bid to rival Nvidia’s CUDA ecosystem and meet surging demand for AI training and inference. Alibaba owns the South China Morning Post.\n“Through systematic innovation, they are integrating thousands or tens of thousands of self-developed chips into a single, massive virtual computing resource pool,” the report said.\nFrost &amp; Sullivan cautioned that China’s domestic GPU market still faced constraints including “hardware performance, software ecosystems, system integration and commercial operations”, meaning only a limited number of vendors had reached large-scale deployment so far.\nHowever, the domestic GPUs would move from being “strategic backups” to “core pillars” for the country’s AI industry over the next three to five years, the report said, adding that these chips were now demonstrating competitive cost-to-performance ratios over general-purpose chips in real-world workloads.\nBaidu’s lead was underpinned by the integration of its AI computing platform Baige with GPUs developed by its chip unit Kunlunxin, which shipped nearly 70,000 units in 2024, according to market research firm IDC.\n\nKunlunxin was founded in 2012 as an internal unit of Baidu before becoming increasingly independent over the years.\nIn its latest bid to expand market share, Kunlunxin unveiled two new AI accelerators last November. Baidu said the M100 would significantly improve inference efficiency for mixture-of-experts models, while the M300 was designed to train super-large multimodal models with trillions of parameters.\nThe report comes as domestic AI chip designers – from Baidu to Moore Threads Technology and Biren Technology – are driven by Beijing’s push for semiconductor self-sufficiency and soaring demand for computing power.\nBaidu joined that wave last week, saying it had confidentially filed for a Hong Kong IPO of Kunlunxin, while noting the size and structure of the offering had yet to be finalised.\nBiren, a Shanghai-based GPU maker, began trading in Hong Kong last Friday after raising HK$5.58 billion (US$720 million). Its debut followed strong recent listings by peers including Moore Threads – dubbed “China’s Nvidia” – and MetaX Integrated Circuits in Shanghai last month.\nEnflame, another Shanghai-based rival to Biren, also completed its IPO tutoring process last week, moving closer to a listing on Shanghai’s Nasdaq-style Star Market.\n</p>
<div class="lb-trans"><p>百度和华为技术在 2025 年上半年共同占据了中国基于自主研发图形处理单元（GPU）的云服务市场超过 70% 的份额，随着对计算能力的需求上升和一波上市计划的涌现，自主研发的人工智能芯片逐渐获得 traction。\n 根据周日发布的 Frost &amp; Sullivan 报告，这些数据将 “GPU 云” 提供商定义为控制从专有 AI 芯片到计算集群和云服务的完整芯片到云的价值链的公司。\n 报告指出，百度占据了 40.4% 的市场份额，华为紧随其后，占 30.1%。\n 这些发现反映了中国领先的 AI 芯片开发商（包括百度、华为和阿里巴巴集团控股）之间的更广泛转变，因为它们不仅限于单一芯片设计，而是致力于构建完整的硬件和软件能力，以期与 Nvidia 的 CUDA 生态系统竞争，并满足对 AI 训练和推理的激增需求。阿里巴巴拥有《南华早报》。\n 报告称：“通过系统创新，他们将成千上万或数万个自研芯片整合到一个庞大的虚拟计算资源池中。”\nFrost &amp; Sullivan 警告称，中国的国内 GPU 市场仍面临 “硬件性能、软件生态系统、系统集成和商业运营” 等限制，这意味着目前只有有限数量的供应商实现了大规模部署。\n 然而，报告指出，未来三到五年，国内 GPU 将从 “战略备份” 转变为国家 AI 产业的 “核心支柱”，并补充说，这些芯片在实际工作负载中表现出相对于通用芯片的竞争性性价比。\n 百度的领先地位得益于其 AI 计算平台百戈与其芯片部门昆仑芯开发的 GPU 的整合，市场研究公司 IDC 表示，昆仑芯在 2024 年出货近 70,000 个单位。\n\n 昆仑芯成立于 2012 年，最初是百度的一个内部部门，随着时间的推移逐渐独立。\n 在其最新的市场份额扩展尝试中，昆仑芯去年 11 月推出了两款新的 AI 加速器。百度表示，M100 将显著提高混合专家模型的推理效率，而 M300 则旨在训练具有万亿参数的超大多模态模型。\n 该报告发布之际，国内 AI 芯片设计师（从百度到 Moore Threads Technology 和 Biren Technology）受到北京推动半导体自给自足和对计算能力需求激增的驱动。\n 百度上周加入了这一浪潮，表示已秘密提交了昆仑芯的香港 IPO 申请，同时指出发行的规模和结构尚未最终确定。\n 总部位于上海的 GPU 制造商 Biren 上周五在香港开始交易，筹集了 58.8 亿港元（约合 7.2 亿美元）。其上市紧随 Moore Threads（被称为 “中国的 Nvidia”）和上个月在上海上市的 MetaX 集成电路等同行的强劲表现。\n 另一家总部位于上海的 Biren 竞争对手 Enflame 上周也完成了其 IPO 辅导过程，距离在上海的纳斯达克风格的科创板上市更进一步。\n</p>
</div>

Baidu and Huawei tighten grip on China’s GPU cloud as chipmakers chase IPOs

百度和华为加紧对中国 GPU 云的掌控，而芯片制造商则在追求首次公开募股