Kunlunxin
一家快速放量的中国本土 Nvidia 替代方案,但约 16x 的估值跃升仍未解释,IPO 结构也带有循环融资信号
昆仑芯在快速扩张的国产芯片市场里拿到了真实且经独立佐证的商业牵引力(China Mobile、Tencent);但据报道约 $50B IPO 目标仍建立在收入口径未调和、代工厂未确认,以及被分析师点名存在循环融资风险的购芯绑定结构之上。
封面要素
公司概况
Kunlunxin (Beijing) Technology Co., Ltd. 是 Baidu, Inc. 旗下半导体设计子公司,开发 AI 训练和推理芯片(Kunlun 系列)以及 CUDA 替代软件栈。它的脉络可追溯到 2011 年 Baidu 内部芯片项目,2021 年 4 月正式拆分为独立公司。Baidu 仍持有约 59.45% 控股权益,同时也是 Kunlunxin 最大客户;借助 China Mobile、China Merchants Bank、China Southern Power Grid、Geely,以及 2026 年以来 Baidu 直接竞争对手 Tencent 等具名客户,外部收入占比从总收入约 40%(2024)升至超过 50%(2025)。Kunlunxin 正推进香港 / 上海 STAR Market 双重上市,2026 年 1 月已保密提交 HKEX Form A1;据报道,路演指引到 2026 年中从约 $14.7 billion 抬升至约 $50 billion——鉴于收入、制造和 IPO 结构证据缺口未解,本报告对这一估值变化保持高度谨慎。
- 成立时间
- 2021-04-01
- 创始人
- Ouyang Jian
- 创立地点
- Beijing, China
- 总部
- Beijing, China
- 产品
- AI 训练和推理加速芯片(Kunlun I、Kunlun II、P800/Kunlun III,以及已宣布面向 2026-2027 年的 M100 和 M300)以硬件加 CUDA 替代软件栈(XTCL/XBLAS/XDNN)销售,既部署在 Baidu 自有 AI 基础设施内,也卖给外部企业、电信、金融、能源、汽车和超大规模云竞争者客户。
- 客户
- Baidu 自有 AI/cloud 基础设施是最大客户;此外覆盖电信和金融国企、能源和汽车企业、政府 / 主权计算采购渠道,以及自 2026 年起的直接超大规模云竞争者(Tencent)。
- 商业模式
- 向 Baidu 内部和外部客户直接销售芯片与集群硬件(捆绑软件栈),主要通过集中采购招标(国企 / 政府买方)和直接企业合同。
- 阶段
- late-stage private, pre-IPO (confidential HKEX filing plus STAR Market tutoring)
- 融资情况
- 最近确认私募轮次:~$283M Series D(2025 年 7 月),投后估值 ~$3.0B。正在推进香港 / 上海 STAR Market 双重 IPO;据报道路演指引到 2026 年 6 月达 ~$50B——截至报告生成日,任何公开招股书均未确认。
执行摘要
主要优势
- 真实且带对抗性的第三方验证:Baidu 的直接竞争对手 Tencent 在 2026 年开始规模化采购 P800 芯片,多家媒体将其视为当前最强的产品质量证据。
- 外部(非 Baidu)收入占比上升,从 2024 年约 40% 提高到 2025 年超过 50%;China Mobile 具名招标金额超过 $139 million,且由竞争中标支撑。
- 自 2011 年以来十多年芯片研发积累,加上深度 PaddlePaddle 原生软件集成,使昆仑芯的生态成熟度高于若干更新的国产芯片创业公司。
主要风险
- 据报道昆仑芯 IPO 目标约 $50B,不到十二个月就是 2025 年 7 月 Series D 估值(约 $3.0B)的约 16 倍;基石投资人购芯承诺结构(股权投资额的 3-7 倍)已被多位分析师明确质疑存在循环融资风险。
- 不同来源对 2025 年收入的估计相差 5-6 倍(约 $500M 到约 $2.9B),任何针对 IPO 目标的 PS 分析都高度不确定(隐含倍数约 17x 到 100x)。
- 据信但未经确认的唯一代工厂 SMIC 产能受限,且据报道把 Kunlunxin 的晶圆分配排在 Huawei 之后;这是结构性制造风险,公司未披露缓释措施。
- Baidu 同时是昆仑芯约 59.45% 控股股东和最大客户,关联方集中度较高,限制了外部对增长叙事的独立验证。
未决问题
- 在 HKEX 招股书发布前,昆仑芯经审计的 FY2024/FY2025 收入、毛利率与研发费用仍未披露。
- 生产 P800 芯片的具体代工厂(据信为 SMIC 7nm DUV)尚未得到昆仑芯官方确认。
- 尚无可独立复现的基准测试,将昆仑芯芯片与 Huawei Ascend、Cambricon 或 Nvidia 芯片直接比较。
- 昆仑芯员工数、债务 / 项目融资义务,以及完整 IPO 承销商团队仍未披露。
- 昆仑芯是否面临或正被考虑直接列入 US Entity List 尚未确认。
目录
01公司概况
1.1 身份、产品与公司结构
Kunlunxin (Beijing) Technology Co., Ltd.(昆仑芯)是 Baidu, Inc. 的半导体设计子公司,以 Kunlun 品牌开发 AI 训练和推理加速器。脉络始于 2011 年:Baidu 智能芯片与架构部开始研究多小核 “XPU” 微架构,先用 FPGA 仿真,再流片为专用硅。Baidu 2018 年 7 月发布第一代 Kunlun 芯片(K200 / Kunlun I),采用 Samsung 14nm 制程、16GB HBM2 内存;2021 年 4 月将芯片团队拆分为如今名为 Kunlunxin 的独立主体,同年完成第一轮外部融资。Kunlun II 于 2021 年 8 月跟进,采用 TSMC 7nm 制程,对外定位为 Nvidia A100 的竞争者。当前旗舰 P800(Kunlun III)2025 年面向大集群训练和推理推出,2025 年 2 月获得 DeepSeek R1/V3 全兼容;Baidu 2025 年 4 月启用 30,000 颗 P800 集群,用于训练 “DeepSeek-like” 模型并支持企业客户并行微调。Baidu 已披露 P800 之后的两颗芯片路线图:推理优化的 M100(目标 2026 年,2026 年 1 月起商业爬坡,实物产品 2026 年 7 月亮相中国央视)和训练导向的 M300(目标 2027 年)。Kunlunxin 官网现在以中文 “股份有限公司”(joint-stock company)而非早期公开记录中的 “有限公司”(limited company)称呼该主体,符合中国公司上市前通常完成的企业形式转换。[CO001, CO002, CO003, CO004, CO005, CO006]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 缺口 |
|---|---|---|---|---|
| 创立(内部 R&D 启动) | 2011(Baidu 智能芯片与架构部) | 2011 | 高 | |
| 分拆为独立公司 | April 2021 | 2021-04-01 | 高 | |
| 总部 | 北京,中国 | 2026-08-19 | 高 | |
| Baidu 持股比例 | 59.45% | 2026-01-02 | 中 | 据报道为 2026 年 1 月 HKEX 申报日口径;准确当前持股比例仍待招股书以及 IPO 发行带来的进一步稀释确认。 |
| 最新披露私募估值 | Series D 估值:RMB 21B(~$3.0B) | 2025-07-01 | 中 | IPO 阶段指引($14.7B 至 ~$50B)来自路演报道,不是已锁定的招股书估值。 |
| 据报道的 IPO 目标估值(香港,2026 年中路演) | ~$50B | 2026-06-29 | 低 | 在公开 HKEX 招股书提交前无法确认;该数字来自媒体报道,不是公司披露。 |
| 2025 年收入(低估计,CNBC) | RMB 3.5B+(~$500M),盈亏平衡 | 2026-06-29 | 低 | 与另一项 ~RMB20B(~$2.9B)的 2025 年估计严重冲突;调和尝试见财务情况章节。 |
| 2025 年收入(高估计,Tencent Tech / BigGo) | RMB 20B+(~$2.9B) | 2026-07-01 | 低 | 来自未具名内部人士,而非经审计披露。 |
| 具名外部客户(已披露) | 7 | 2026-08-19 | 中 | 数量反映媒体报道中具名实体(China Mobile、China Merchants Bank、China Southern Power Grid、Geely、Vivo、Tencent、BYD);不是公司确认的客户数。 |
| 员工数 | 2026-08-19 | 低 | 已审阅来源没有披露具体员工数;需向管理层或 HKEX 招股书索取。 | |
| IPO 状态 | 已提交保密 HKEX Form A1;并行启动 STAR Market 辅导 | 2026-05-08 | 高 | 截至本报告运行日,公司仍为私营;最终定价和上市日期尚未公开。 |
收入和员工数行列出明确区间 / 缺口,因为 Kunlunxin 财务数据尚未经过独立审计,也未在公开招股书披露;两条收入记录应视为相互竞争的媒体估计,而不是已经调和的单一数字。
[CO002, CO003, CO015, CO021, CO028, CO031]Kunlunxin 的身份、产品线、资本基础和客户组合都回到 Baidu;Baidu 同时是控股股东、最大客户和 IPO 发起方。
[CO001, CO015, CO021, CO037, CO040, CO044]1.2 领导层、治理与关键人物依赖
Kunlunxin 并不像许多风险投资支持的初创公司那样由独立创始人团队主导;它的领导层是专职芯片运营高管与 Baidu 公司高管的混合。Ouyang Jian 作为 Baidu 首席芯片架构师,在 2021 年 4 月拆分时出任 Kunlunxin CEO,并一直负责日常芯片战略。在他之上,Baidu 董事长兼 CEO Robin Li (Li Yanhong) 公开把 Kunlun 项目称作 14 年战略押注,并在 2026 年 1 月采访中亲自确认 M100 芯片的 2026 年发布时间。Baidu 执行副总裁、AI Cloud Group 总裁 Shen Dou 是公司芯片路线图的指定发言人,2025 年 11 月在 Baidu World 发布 M100/M300 时间表和 Tianchi256/512 集群计划。Baidu CFO Haijian He 在 Baidu 自身 2025 年 Q4 业绩稿中,把 Kunlunxin 拆分直接与释放股东价值挂钩。治理和控制权由 Baidu 贯穿:据 2023 年 DigiTimes 报道,母公司当时持股约 70.87%;随着外部投资者自 2021 年以来约每年一轮进入,到 Kunlunxin 2026 年 1 月保密 HKEX 申报时,持股被稀释至约 59.45%;同一申报显示股东名册已扩至 44 个不同主体。该结构带来的关键人物依赖,不是经典创始人叙事,而是持续 Baidu 管理层支持;同时也让 Baidu 既是 Kunlunxin 控股股东,又是单一最大客户——这一关联方动态需要后续章节进一步做财务穿透。[CO011, CO012, CO013, CO014, CO015, CO016]
| 人物 | 角色 | 背景 | 创始人-市场匹配或职能覆盖 | 关键人依赖 |
|---|---|---|---|---|
| Ouyang Jian | CEO,Kunlunxin | 2021 年分拆前是 Baidu 首席芯片架构师;分拆后出任独立实体 CEO。 | 深厚半导体架构背景,直接支撑芯片路线图执行的技术可信度。 | 高 |
| Robin Li (Li Yanhong) | Baidu 联合创始人兼 CEO(Kunlunxin 控股母公司) | 多年来推动 Kunlun 芯片项目作为战略押注,并在 2026 年 1 月采访中亲自确认 M100 发布时间表。 | 提供资本市场背书和公共可信度,但不是传统意义上的 Kunlunxin 员工或创始人。 | 高 |
| Shen Dou | Baidu 执行副总裁;Baidu AI Cloud Group 总裁 | 2025 年 11 月在 Baidu World 公开发布 M100/M300 路线图和 Tianchi256/Tianchi512 集群计划。 | 承担芯片项目对外沟通和云集成负责人角色。 | 中 |
| Haijian He | Baidu 首席财务官 | 在 Baidu 2025 年 Q4 业绩公告中,明确把 Kunlunxin 分拆和上市与释放股东价值联系起来。 | 负责分拆和双重上市背后的资本市场与财务披露策略。 | 中 |
除 Ouyang Jian 外,公开记录中没有看到 Kunlunxin 独立于 Baidu 高管体系的董事会或管理层名单;在招股书披露完整管理团队前,应把本表视为部分列举。
[CO011, CO012, CO013, CO014, CO018]Kunlunxin 公开记录从 2011 年 Baidu 内部芯片研发起步,经历 2021 年拆分、2025 年 DeepSeek 级集群上线,并进入 2026 年香港 / 科创板双重 IPO 进程;该进程附带少见的购芯条件。
[CO003, CO007, CO008, CO020, CO024, CO026]1.3 融资历史、估值与资本结构
Kunlunxin 的资本历史呈现陡峭且越来越压缩的估值曲线。2021 年 4 月拆分轮由 CPE 领投,IDG Capital、Legend Capital 和 Yuanhe Puhua 参与,给新公司估值约 RMB13 billion (~$1.9 billion)。2025 年 7 月 Series D 完成,融资约 $283 million (~RMB2.1 billion),投资方包括 BYD、CICC Guanbo、China Mobile 旗下 Zhongyi Hechuang 投资平台,以及 China Reform Holdings 旗下国资背景 High-Level Talent Fund,投后估值升至约 RMB21 billion (~$3.0 billion)。2026 年 1 月 1 日,Kunlunxin 向 Hong Kong Stock Exchange 主板保密提交 Form A1 上市申请;披露后 Baidu 港股最高上涨 9.35%。此后市场流传数字快速上移:在 Kunlunxin 2026 年 5 月启动并行 Shanghai STAR Market “辅导”流程前几周,曾有中间值 ~$14.7 billion 传出(当时指引接近 HK$100 billion,~$13 billion);到 2026 年 6 月 29 日投资者路演报道,香港目标已达约 $50 billion,消息推动 Baidu ADR 和港股双双上涨。按该水平,Kunlunxin 隐含估值(约 1.32x)超过 Baidu 同一交易日约 $38 billion 市值——多个媒体标记出醒目的“子公司估值高过母公司”动态。这些 IPO 阶段数字尚未写入公开招股书;截至本报告 2026 年 8 月生成日,Kunlunxin 仍是私营公司,最终香港和 STAR Market 定价、股份分配和承销团仍未披露。[CO019, CO020, CO021, CO022, CO023, CO024]
| 利益相关方 | 角色 | 控制权或经济重要性 | 尽调问题 |
|---|---|---|---|
| Baidu, Inc.(NASDAQ:BIDU;HKEX:9888,控股母公司) | 控股股东和最大客户 | ~59.45% 股权;也是 Kunlunxin AI 基础设施的主要内部客户 | 在 IPO 招股书前,量化 Baidu 与 Kunlunxin 之间的关联方定价和收入条款。 |
| CPE (CPE Yuanfeng) | 2021 年分拆轮领投方 | 初始估值 ~$1.9B 时的早期锚定投资者 | 确认当前稀释后持股比例以及任何董事会 / 治理权利。 |
| IDG Capital | 早期 / 成长期投资者 | 参与 2021 年分拆轮 | 确认后续轮次参与情况和当前持股规模。 |
| Legend Capital | 早期 / 成长期投资者 | 参与 2021 年分拆轮 | 确认后续轮次参与情况和当前持股规模。 |
| BYD | Series D 战略投资者,据报道也是早期小额投资者 | 持有少数股权,并在 AI/ADAS 算力上有汽车行业战略利益 | 判断 BYD 是否也是 Kunlunxin 芯片客户,还是仅为财务 / 战略投资者。 |
| China Mobile(通过 Zhongyi Hechuang 投资载体) | Series D 投资者和最大已披露外部客户 | 少数股权,加上 2025 年 8 月中标的 RMB1B+(~$139M)集采合同 | 量化合同期限、续约条款和收入确认处理。 |
| China Reform Holdings(High-Level Talent Fund,国资背景基金) | 国家背景 Series D 投资者 | 少数股权;显示国家产业政策背书 | 确认治理权利以及与国有资本参与绑定的任何条件。 |
| CICC Guanbo | Series D 投资者 | 少数股权 | 确认投资规模以及是否有战略使命(而非纯财务投资)。 |
投资者名册并未完全公开;据报道,Kunlunxin 2026 年 1 月 HKEX 文件共列出 44 名不同股东,因此本表只列出已审阅媒体报道中具名的实体。
[CO015, CO017, CO019, CO020, CO022, CO037]Kunlunxin 兼具快速上升的私人估值和增长中的外部商业验证,但收入、员工数和招股书细节仍未解决。
[CO003, CO015, CO021, CO028, CO037, CO040]1.4 里程碑、商业进展与证据缺口
Kunlunxin 的商业记录在几年内从只供 Baidu 的芯片供应商,转为拥有具名外部客户的公司。2025 年 8 月,其 AI 服务器在 China Mobile 集中采购中标,三个通用计算标包合计中标金额超过 RMB1 billion (~$139 million);Baidu 另行点名 China Merchants Bank、China Southern Power Grid、Geely Automobile Holdings、Vivo 与自用场景一起部署外部 P800。BYD 在记录中出现两次:据 2023 年 DigiTimes 报道,它是与自动驾驶芯片规划相关的小额早期投资者;2025 年又作为 Series D 股权投资者进入。据报道,外部(非 Baidu)收入从 2024 年约占总收入 40% 升至 2025 年超过一半;Tencent——Baidu 的直接竞争对手——2026 年开始规模采购 P800 芯片,多家媒体将其视为有别于 Baidu 内部需求的市场验证信号。各来源对 Kunlunxin 收入数字冲突很大:CNBC 称 2024 年约 RMB2 billion(亏损约 RMB200 million),2025 年升至超过 RMB3.5 billion(~$500 million,盈亏平衡);而 BigGo Finance 援引 Tencent Tech 的数字称 2024 年收入超过 RMB10 billion (~$1.4 billion),2025 年收入“远超” RMB20 billion (~$2.9 billion)——二者相差约 5-6x,尚未调和,本报告将其视为未解决冲突,而非选取单一数字。Kunlunxin 香港 IPO 结构据报道还要求基石投资者承诺购买其股权投资额 3 到 7 倍的芯片,分析师已指出,这会引发关于外部需求是否有机的循环融资疑虑。公开证据尚不能确认 Kunlunxin 员工数、经调和的 2025 年收入数字、P800 制造代工方(SMIC 还是 TSMC),或其是否面临任何直接美国出口管制清单指定——本报告将这些缺口作为明确尽调问题带入后续,而不是编造数字。[CO031, CO032, CO033, CO034, CO035, CO036]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 影响 |
|---|---|---|---|---|---|
| 2011 | Baidu 设立智能芯片与架构部,启动内部 Kunlun 芯片 R&D | 创立 | 内部 R&D 项目(FPGA 前、流片前) | Baidu | 奠定 Kunlunxin 芯片项目历史长于多数 2019 年后成立的中国 AI 芯片竞品。 |
| 2018-07 | Kunlun I(K200)芯片发布,并部署于 Baidu 搜索引擎 | 产品 | Samsung 14nm,16GB HBM2,512GB/s 带宽 | Baidu、Samsung Foundry | 多年 FPGA 仿真后,第一个专用硅片验证点。 |
| 2021-04 | 芯片团队分拆为独立公司 Kunlunxin;首轮外部融资完成 | 创立 | ~RMB13B(~$1.9B)估值 | 投资方:CPE、IDG Capital、Legend Capital、Yuanhe Puhua | 首次把 Baidu 内部项目转成可融资的独立实体。 |
| 2021-08 | Kunlun II 芯片进入量产 | 产品 | TSMC 7nm;定位为 Nvidia A100 级性能 | Kunlunxin、TSMC | 让 Kunlunxin 成为可信的训练 / 推理芯片供应商,而不只是搜索推理 ASIC 厂商。 |
| 2025-02 | P800(Kunlun III)获得 DeepSeek R1/V3 训练和推理全兼容 | 产品 | 一键部署;单台 8 卡服务器以 8-bit 运行完整 671B 参数 DeepSeek R1 | Kunlunxin | 把 Kunlunxin 旗舰芯片直接绑定到中国最受关注的开源模型家族。 |
| 2025-04 | Baidu 在开发者大会启用 30,000 颗芯片的 P800 集群 | 规模化 | 集群支持「DeepSeek-like」模型训练和 1,000+ 企业微调工作负载 | Baidu、Kunlunxin | 展示集群级部署,而不只是单芯片基准测试。 |
| 2025-07 | Series D 融资完成 | 融资 | ~$283M 融资,投后估值 ~RMB21B(~$3.0B) | BYD、CICC Guanbo、China Mobile(Zhongyi Hechuang)、China Reform Holdings 基金 | 在四年内使 Kunlunxin 估值较 2021 年分拆水平约翻倍。 |
| 2025-08 | 中标 China Mobile AI 算力集采 | 合作 | 覆盖全部三个标包的 RMB1B+(~$139M)订单 | Kunlunxin、China Mobile | 最大已披露外部商业订单;证明非 Baidu 需求的关键证据。 |
| 2025-11 | M100/M300 路线图和 Tianchi256/512 集群计划在 Baidu World 发布 | 产品 | M100(2026,推理);M300(2027,训练) | Baidu、Kunlunxin(Shen Dou) | 传递至少五年年度芯片迭代节奏的信号。 |
| 2026-01-01 | 向 HKEX 主板提交保密 Form A1 上市申请 | 监管 | IPO 申请保密;披露后 Baidu 股价 +9.35% | Kunlunxin、Baidu、HKEX | 迈向独立上市的第一个正式、市场可见步骤。 |
| 2026-01 | M100 芯片进入商业爬坡 | 规模化 | 商业爬坡开始;实物产品后于 CCTV 展示(2026 年 7 月) | Kunlunxin | 标志路线图从发布转向硬件出货。 |
| 2026-05-08 | 上海 STAR Market IPO「辅导」流程启动 | 监管 | 双重上市策略确认;香港指引随后接近 HK$100B(~$13B) | Kunlunxin、STAR Market 监管机构 | 在香港申请之外建立第二条境内上市路径。 |
| 2026-06-29 | 香港 IPO 估值目标据报接近 $50B;与芯片购买挂钩的投资者结构披露 | 反向 | 据报道,基石投资者需购买相当于其股权承诺 3-7x 的芯片 | Kunlunxin、潜在基石投资者 | 估值在不到 12 个月内较 2025 年 Series D 水平约抬升 16x,少见的购买条件也引发分析师对循环融资的质疑。 |
这是第 1 章的唯一记录时间线;后续章节应引用这些带日期事件,而不是重新推导一条单独时间线。本章其他位置引用的收入数字被有意排除在本表之外,因为不同来源之间存在冲突。
[CO002, CO003, CO004, CO005, CO006, CO007]1.5 图表
02市场分析
2.1 市场边界与相邻领域
本章把“市场”定义为中国本土 AI 加速器(训练和推理芯片)市场——即中国超大规模云厂商、国企和政府关联数据中心运营方为 AI 工作负载采购 GPU/ASIC 级芯片的计算支出部分。它不包括通用 CPU 服务器、非 AI 网络设备和非中国芯片需求(Kunlunxin 未披露中国以外销售)。相邻但不同的市场包括高带宽内存(HBM)供应、先进制程代工产能(SMIC、TSMC)和 AI 软件 / 工具体系(CUDA 及国产替代),这些因素约束加速器需求,但不构成加速器需求本身。中国买方的现状替代方案,是进口或租用 Nvidia/AMD 算力(直接采购、通过缩小后的 H20 渠道,或在 Southeast Asia 等司法辖区租用离岸云);第二个由政策驱动的替代方案,是干脆不采用新的 AI 基础设施。Reuters 查阅的 IDC 数据显示,2025 年中国 AI 加速卡总出货量约 4 million 块,其中 Nvidia 55%、AMD 4%、中国本土厂商 41%——这个市场边界能覆盖 Kunlunxin 的可触达机会,而不会膨胀成“全部中国 AI 支出”。[CM001, CM002, CM003, CM004, CM005]
| 细分 / 类别 | 包含支出 | 排除支出 | 买方 / 付款方 | 对 Kunlunxin 的意义 |
|---|---|---|---|---|
| AI 训练加速器(中国) | 大模型预训练集群芯片 | 非 AI CPU / 服务器算力;中国以外训练支出 | 超大规模云厂商、国家 AI 基础设施项目 | Kunlunxin P800/M300 的核心可服务细分。 |
| AI 推理加速器(中国) | 规模化模型服务 / 推理芯片 | 数据中心外的边缘 / 消费级推理芯片 | 超大规模云厂商、企业、国企 | Kunlunxin P800/M100 的核心可服务细分。 |
| HBM 内存供应 | 连接 AI 加速器的高带宽内存 | 非 AI 工作负载使用的标准 DDR/GDDR 内存 | 芯片厂商(作为成本输入,而非面向买方的市场) | 限制 Kunlunxin 集群扩张的上游约束,不是 Kunlunxin 销售进入的市场。 |
| 先进制程代工产能 | AI 芯片 7nm 及以下晶圆制造 | 非 AI 芯片成熟制程制造 | 芯片厂商(SMIC 客户) | Kunlunxin 与 Huawei、Cambricon、Biren 及其他 SMIC 客户共同面对的上游约束。 |
| AI 软件 / 工具生态 | AI 芯片编译器、运行时、框架(CUDA 及国产替代) | 与 AI 加速无关的通用软件 | 开发者、超大规模云厂商 | 决定 Kunlunxin 的 XTCL/XBLAS 栈能多容易赢下已锁定在 CUDA 上的工作负载。 |
| 现状替代:进口 / 租用 Nvidia-AMD 算力 | 直接购买 Nvidia/AMD 芯片;离岸云 GPU 租赁 | 已被出口管制排除、无法销往中国的芯片 | 中国 AI 实验室和企业 | Kunlunxin 必须替代的主要竞争选项,不是其目前能服务的细分。 |
「对 Kunlunxin 的意义」是本报告基于本章证据作出的分类,不是公司发布的市场地图。
[CM001, CM002, CM003, CM019, CM021]2.2 多重视角下的市场规模
这里没有单一 TAM 数字可靠,因此本章保留几组测算视角,而不强行选一个。Bernstein 模型(经 AI Gazine)将中国 AI 加速器市场估为 2025 年 $37 billion、2026 年 $55 billion、2027 年 $69 billion;国产芯片产值从 2025 年 $16 billion 升至 2028 年 $91 billion,而全国总需求为 $88 billion——若预测成立,中国可能在两到三年内成为 AI 加速器净出口国。第二个按出货量看的视角(IDC 经 Reuters)显示,2025 年中国约出货 4 million 块加速卡,其中国产厂商 1.65 million 块(41%);Huawei 单独出货约 812,000 块(约占全部国产出货一半、总市场约 20%),Alibaba T-Head 约 265,000 块,Kunlunxin 和 Cambricon 各约 116,000 块——Kunlunxin 仅占中国加速器总出货约 3%,在国产厂商中并列第三且距离第一梯队较远。第三个视角来自 TrendForce 的定性跟踪,预计到 2026 年底国产芯片将超过中国 AI 服务器部署的 50%。这些视角方向一致(国产份额快速上升),但绝对规模相差数十亿美元,也没有一个单独拆出 Kunlunxin 可服务的可触达市场——本章将该缺口作为明确尽调问题,而不是制造 SAM/SOM 拆分。[CM006, CM007, CM008, CM009, CM010, CM011]
| 发布方 | 年份 | 地域 | 数值 | CAGR / 趋势 | 方法 | 置信度 | 局限 |
|---|---|---|---|---|---|---|---|
| Bernstein(经 AI Gazine) | 2025 | 中国 | $37B AI 加速器市场 | 增长至 $55B(2026)、$69B(2027) | 加速器支出和产量的分析师模型 | 中 | 单一分析师预测;未获得第二家具名研究机构在同等颗粒度上的佐证。 |
| Bernstein(经 AI Gazine) | 2028 | 中国 | $91B 国内产量 vs $88B 总需求 | 意味着 2028 年成为净出口方 | 分析师模型 | 低 | 三年后预测;依赖此处未独立验证的代工 / HBM 产能假设。 |
| IDC(经 Reuters/Yahoo Finance) | 2025 | 中国 | ~4M AI 加速器卡出货;国内份额 41% | 国内份额从历史 Nvidia 主导状态上升 | 跨厂商单位出货跟踪 | 高 | 统计单位数量,不是美元价值或按性能调整后的算力份额。 |
| TrendForce | 2026 | 中国 | >50% AI 服务器部署采用国产(预测) | 从当前 ~41% 出货份额上升 | 行业研究机构跟踪 | 中 | 部署份额预测,不是具体美元 TAM 数字。 |
| ByteDance 内部支出(经 Fresh From China) | 2025-2026 | 中国(单一买方) | $12.1B(2025)升至 $14.3B(2026) | 同比增长 ~18% | 报道的单一超大规模云厂商 AI 芯片预算 | 中 | 单一买方预算,不是全市场数字;仅用于说明买方层面的规模。 |
| Council on Foreign Relations(智库) | 2025-2027 | 中国与美国(按算力调整) | 2025 年美国算力领先约 ~5x,估计到 2027 年扩大至 ~17x | 差距扩大,而非缩小 | 政策分析师算力性能模型 | 中 | 与出货份额增长叙事相矛盾;衡量原始算力,不是单位数或美元。 |
本表没有任何一个数字应被解读为中国 AI 芯片 TAM 的「唯一」口径;这些视角使用不同单位(美元、单位、按算力调整的性能)和不同时间跨度,且多项为单一来源的分析师估计。
[CM006, CM007, CM008, CM009, CM010, CM014]中国 AI 加速器市场从宽口径总需求估计,层层收窄到 Kunlunxin 当前参与竞争的国内供应商出货份额。
美元和出货量数字来自不同来源(Bernstein 模型与 IDC 出货数据),放在一起只是为了展示相对规模,不是单一调和数字。
[CM006, CM008, CM009, CM010, CM026]2026 年中国 AI 加速器市场规模估计因方法不同而跨度很大,从 TrendForce 的部署份额预测到 Bernstein 的美元口径预测不等。
2026 年美元规模区间(39-69)沿用 Bernstein 自己的 2025-2027 趋势线,并保守应用;国内份额区间(41-60%)反映 IDC 已确认的 2025 年数据(41%)与 TrendForce “exceed 50%” 2026 年预测之间的差距,其中 60% 只是上行情景,不是已发布预测。
[CM007, CM011, CM012]2.3 买方分层、增长驱动与采用约束
中国本土 AI 加速器买方分为四类,预算负责人和采用触发点不同。超大规模云厂商(Baidu、Alibaba、Tencent、ByteDance)是最大且最成熟的买方,通常有意采用多供应商策略——例如,据报道 ByteDance 2025 年在 AI 芯片上花费约 $12.1 billion(2026 年升向 $14.3 billion),在 Huawei(重训练)、Cambricon(中端推理)和 Iluvatar CoreX(Kunlunxin 仍在评估)之间分配聊天机器人推理,明确用来对冲供应风险并优化每个工作负载的成本。国企和政府关联买方(China Mobile、State Grid 系统实体、China Merchants Bank)构成第二个政策驱动分层,采购受“Xinchuang”(信创)国产替代要求和集中招标流程塑造,Kunlunxin 自身超过 RMB1 billion 的 China Mobile 中标就是例子。汽车 OEM(BYD、Geely)构成规模更小但战略可见度高的第三类,为 ADAS 和车内 AI 采购算力。全国“East Data West Computing”(东数西算)基础设施枢纽和聚焦主权计算的部委,是第四类补贴浓度高的买方。主导增长驱动是美国出口管制:2025 年 H20 禁令(后来以 15% Treasury fee 部分逆转)和后续 Entity List 行动,反复把可触达需求从 Nvidia 转向国产厂商,并由“Buy China”采购偏好和约 $47.5 billion Big Fund III 国家半导体融资强化。DeepSeek 级开源模型形成第二个需求侧驱动,给每家国产芯片厂商一个具体、高关注度的推理 / 训练工作负载可争取。主要约束在供给侧:SMIC 可用的最先进国产节点(7nm,基于 DUV)成本估计比 TSMC EUV 等效节点高 25-50%,良率也明显更低,而且产能受限,Huawei、Cambricon、Biren 等国家冠军客户争抢同一批晶圆;CXMT 的 HBM 供应相较 Samsung/SK Hynix 仍不成熟;即便硬件规格纸面有竞争力,CUDA 软件生态迁移到国产芯片仍需要显著更多工程投入。[CM013, CM014, CM015, CM016, CM017, CM018]
| 细分 | 买方 | 用户 | 付款方 | 预算负责人 | 采用触发点 |
|---|---|---|---|---|---|
| 超大规模云厂商 | Baidu、Alibaba、Tencent、ByteDance | 内部 AI 产品 / 云团队 | 超大规模云厂商自身(内部资本开支) | 云 / AI 基础设施预算负责人 | 模型训练规模需求,以及多供应商供应风险对冲。 |
| 国企 / 金融、电信国企 | 客户:China Mobile、China Merchants Bank | 企业 IT 和 AI 应用团队 | 国企自身,通过集中招标采购 | 信创政策下的采购 / IT 预算负责人 | 集中采购招标要求国产替代份额。 |
| 汽车 OEM | BYD, Geely | ADAS / 自动驾驶工程团队 | OEM 自身(研发 / 产品预算) | 车载电子 / ADAS 预算负责人 | 车载 AI 计算和自动驾驶路线图里程碑。 |
| 能源 / 基础设施国企 | China Southern Power Grid | 电网运营和 AI 规划团队 | 国企自身 | 数字基础设施预算负责人 | 智能电网和工业 AI 现代化要求。 |
| 主权 / 政府算力项目 | 东数西算枢纽运营方 | 公共部门和托管第三方 AI 工作负载 | 中央 / 地方政府补贴,加上枢纽运营方共同投资 | 国家基础设施投资预算 | 国家算力自主可控和区域发展政策。 |
| 金融服务(非国企) | 评估国产芯片的银行和保险公司 | 风控、反欺诈和 AI 应用团队 | 机构自身 | IT / 安全预算负责人 | 数据主权和供应商多元化要求。 |
预算负责人和采用触发因素分类来自本报告对本章买方策略和采购报道的综合判断;它们不是公司发布的客群划分。
[CM015, CM016, CM017, CM018, CM020]| 驱动因素 / 约束 | 方向 | 时间 | 影响 | 尽调问题 |
|---|---|---|---|---|
| 美国对先进 Nvidia / AMD 芯片的出口管制 | 驱动因素(利好国内厂商) | 2022 年以来持续,2024-2025 年加码 | 结构性地把可触达需求推向包括 Kunlunxin 在内的国产芯片厂商。 | 跟踪 BIS 规则变化以及任何可能削弱顺风的部分回撤(如 H20 许可)。 |
| "买中国" / 信创国产替代政策 | 驱动因素(利好国内厂商) | 持续推进,2024-2026 年强化 | 为国企和政府买家设定由政策托底的采购下限。 | 确认 Kunlunxin 是否享有与 Huawei / Cambricon 相同的采购偏好。 |
| DeepSeek 级开放模型需求 | 驱动因素(利好所有拥有兼容芯片的厂商) | 2025 年初以来加速 | 让每家国产芯片厂商都有一个具体且高曝光的工作负载可争取,包括 Kunlunxin 的 P800。 | 评估 DeepSeek 相关需求的持久性,而非只看一次性兼容公告。 |
| 国家半导体融资(大基金三期,~$47.5B) | 驱动因素(整体利好国内生态) | 2024 年以来投放 | 补贴成本更高的国内制造,部分抵消 SMIC 的成本 / 良率劣势。 | 确认大基金三期资本中有多少(如有)具体流向 Kunlunxin,而不是 SMIC / Huawei。 |
| SMIC 先进制程产能约束 | 约束 | 至少持续到 2026 年 | 限制包括 Kunlunxin 在内的任何 SMIC 客户扩大芯片产出的速度。 | 厘清 Kunlunxin 相对 Huawei 和 Cambricon 的具体晶圆分配优先级。 |
| HBM 供应约束(CXMT 尚不成熟,落后 Samsung / SK Hynix) | 约束 | 持续存在 | 限制包括 Kunlunxin P800 / M 系列在内的国产 AI 芯片扩展内存带宽。 | 确认 Kunlunxin 当前 HBM 来源(国产 CXMT 还是进口)。 |
| CUDA 生态锁定 / 迁移成本 | 约束 | 持续存在;最晚至 2026 年 8 月仍有记录 | 即便硬件规格具备竞争力,也会拖慢训练工作负载采用国产芯片。 | 评估 Kunlunxin XTCL 栈相对 Huawei CANN 的真实迁移时间 / 成本数据。 |
方向和时间来自本报告对本章增长驱动因素和约束证据的综合判断;驱动因素和约束并不互斥,也可能以不同方式影响不同厂商。
[CM013, CM019, CM020, CM021, CM022, CM025]不同买方细分市场在预算所有者和采用触发因素上差异明显,这会影响 Kunlunxin 在不依赖 Baidu 内部需求时,能否直接打入各类客户。
[CM015, CM016, CM017, CM018, CM013]中国 AI 买家从评估国产芯片到搭建生产级 AI 基础设施,要经过政策筛选、采购、集成和规模化部署几个阶段。
各阶段数值只是基于定性证据(多厂商比拼、招标、迁移,以及多个来源中的规模评论)给出的相对权重,不是来自单一数据集的实测转化率漏斗。
[CM015, CM019, CM021, CM022]2.4 相互矛盾的估计与规模尽调缺口
多项被广泛引用的数字彼此冲突,或无法从已审阅来源独立调和。Bernstein 预测中国到 2028 年成为 AI 芯片净出口国;Council on Foreign Relations 的评估却给出另一面:2025 年美国最强芯片仍比中国最强芯片约强 5x,且 CFR 预计到 2027 年差距可能扩大到 17x——也就是说,出货量份额和原始算力份额可能同时朝相反方向移动。另有 2026 年 8 月报告发现,许多中国 AI 开发者仍通过 Southeast Asia 离岸云安排租用 Nvidia 算力,专门绕开 CUDA 迁移成本,这让将 41% 国产出货份额解读为实际消耗 AI 算力的 41% 变得复杂。已审阅来源没有一个能在 Baidu 自身内部需求之外单独拆出 Kunlunxin 的可触达市场,也没有来源提供 P800 对 Nvidia H100/H200 性能的独立验证基准——Morgan Stanley 引用的 DeepSeek R1 推理约 1,521 tokens/second 估计(将 P800 定位在 H20 和 A100 之间)是目前最可信的第三方数据点,但它只是单一估计,不是可复现实测套件。这些缺口很重要,因为它们意味着 Kunlunxin 市占率和利润率轨迹尚不能像 Huawei 或 Cambricon 那样有同等信心建模。[CM023, CM024, CM025, CM026, CM027]
03竞争格局
3.1 竞争格局概览
Kunlunxin 的竞争集合横跨直接国产同行、主导性在位者和受出口限制的现状方案。Huawei Ascend 系列(910B/910C,以及更新的 Atlas 350)是国产在位者,背后有 Huawei 自身 CANN 软件栈和 SMIC 生产优先级。Cambricon 和 Moore Threads 是已上市的两家国产同行(Cambricon 历史上在 Shanghai STAR Market,Moore Threads 通过 2025 年 12 月 STAR Market 首秀上市),估值均远高于 Kunlunxin 私募标记。Biren Technology 和 Iluvatar CoreX 都在 Kunlunxin 瞄准的同一 2025 年 12 月至 2026 年 1 月窗口完成香港上市;Enflame Technology 和 Alibaba 内部团队 T-Head(Pingtouge)仍处于 pre-IPO 或仅内部阶段。在这些直接同行之外,任何中国买方的现状替代方案,是在出口许可或离岸云租赁允许时继续采购 Nvidia(以及较小程度 AMD)芯片,或完全自研芯片,而不是从上述任何厂商采购。合计看,至少六家 AI 芯片公司在 2025 年 12 月至 2026 年 6 月之间上市或推进上市流程;到 2026 年中,“国产 GPU 军团”(Cambricon、Moore Threads、Biren、Enflame、Hygon 等)的合计市值已接近 RMB 2 trillion——Kunlunxin 自身 IPO 正是在乘这波估值潮。[CP001, CP002, CP003, CP004, CP005, CP006]
| 竞争对手 | 类别 | 规模 / 融资 | 目标客群 | 差异化 | 局限 |
|---|---|---|---|---|---|
| Huawei Ascend (HiSilicon) | 国内直接竞品 / 现任龙头 | ~$12B AI 芯片收入(2026 估计);Huawei 子公司,未单独估值 | 超大规模云厂商、政府、企业(国内覆盖最广) | SMIC 产能优先级;CANN 软件栈;国内出货量最大(2025 年 ~812,000 颗) | 未独立上市;按 CFR 分析,绝对算力仍落后 Nvidia。 |
| Cambricon (寒武纪) | 国内直接竞品,已上市 | RMB 630B-1T (~$88-140B) 市值(2026);2025 年收入 RMB6.5B,首次实现年度盈利 | 云端 AI 推理,超大规模云客户(ByteDance、Alibaba) | 首家实现盈利并突破 RMB1T 市值的国内纯 AI 芯片厂商。 | 按过去 12 个月 PE 200-373x 交易,相比全球半导体同行估值倍数极端。 |
| Moore Threads (摩尔线程) | 国内直接竞品,已上市 | IPO 后市值 RMB 282-310B (~$39-43B)(2025 年 12 月);IPO 融资 ~RMB6B | 全栈 GPU(AI + 图形 + 视频),面向消费者和企业 | MUSA 全栈软件生态,配 Musify CUDA 迁移工具;创始人曾任 Nvidia VP。 | 三年累计亏损近 RMB6B;高度依赖外部融资。 |
| Biren Technology (壁仞科技) | 国内直接竞品,HKEX 上市(2026 年 1 月) | ~$2.2-2.9B IPO 前估值;多轮累计融资 ~RMB5B+ | 高性能云端 AI 训练 / 推理(BR100/BR106 系列) | 首款支持 Chiplet、PCIe 5.0 和 CXL 的中国 GPU(BR100,2022);完整支持 DeepSeek-V3 全栈。 | 自 2023 年 10 月起被列入美国实体清单,供应链访问受限。 |
| Iluvatar CoreX (天数智芯) | 国内直接竞品,HKEX 上市(2026 年 1 月) | ~$5.2B IPO 后市值;2025 年收入 RMB1.03B(同比 +91.6%) | 云端 AI 推理(ByteDance 聊天机器人工作负载) | ByteDance 订购 50,000+ 颗用于推理;~54% 毛利率,在国内同行中位居前列。 | 规模小于 Huawei / Cambricon;对单一大客户(ByteDance)敞口集中。 |
| Enflame Technology (燧原科技) | 国内直接竞品,IPO 前(2026 年中科创板审核通过) | ~$2.8B IPO 前估值(2024 年 12 月 Series E);目标科创板融资 ~RMB6B | 面向超大规模云和企业客户的云端 AI 训练 / 推理 | 软硬件协同设计研发投入;多代芯片路线图(IPO 融资支持第 5 / 第 6 代研发)。 | 截至报告运行日期,尚未独立盈利或上市。 |
| Alibaba T-Head / Pingtouge (平头哥) | 自用型超大规模云芯片部门,IPO 路线 | 未披露独立估值;累计出货数十万颗 PPU 芯片 | Alibaba Cloud 自有 AI 基础设施,其次是外部云客户 | PPU 规格声称 BOM 成本比 Nvidia H20 低 ~40%;深度嵌入自有云。 | 未独立披露财务;成败与 Alibaba Cloud 自身 AI 路线图高度绑定。 |
| Nvidia(现状 / 存量基准) | 现状替代品,不是国内竞争对手 | ~$3T+ 全球市值;2025 年中国 AI 加速器份额 ~55%(下降中) | 全球超大规模云厂商,以及任何能获得出口许可或离岸租赁芯片的中国买家 | CUDA 生态仍是每家国产厂商试图迁出的软件基准。 | 在中国受到直接出口限制;H20 准入时断时续,并受政治条件影响。 |
规模 / 融资数据混合了市值、私募估值和收入,取决于各公司披露口径;跨行比较只能作方向性判断,不能当作同口径倍数。
[CP001, CP002, CP006, CP009, CP014, CP016]3.2 能力、定价与商业化路径对比
看原始能力,据独立对比,Huawei Ascend 910C 提供约 Nvidia H100 推理性能的 60-80%;Huawei 更新的 Atlas 350 则由 Huawei Ascend Computing Business 总裁声称可达到 Nvidia 中国特供 H20 芯片 2.87x 性能——若该说法成立,国产芯片竞争力就不再只是“达到 Nvidia 最强产品百分之多少”的框架。Morgan Stanley 独立估计 Kunlunxin P800 在 DeepSeek R1 推理上约 1,521 tokens/second,定位在 Nvidia H20 与 A100 之间——大体接近 Huawei 的定位,但依据是单一分析师估计,而非标准化基准。软件上,所有主要国产厂商都已构建(或正在构建)CUDA 替代栈:Huawei CANN、Moore Threads MUSA(带 “Musify” 代码迁移工具),以及 Kunlunxin 自身 XTCL/XBLAS/XDNN 栈。各厂商定价基本不透明,因为没有一家公布标价;最接近的公开参照点是 Alibaba T-Head 声称其 BOM 比 Nvidia H20 低约 40%,以及 Moore Threads 消费级 MTT S80 GPU 定价接近 Nvidia RTX 3060 档位。商业化路径上,Moore Threads 和 Biren 都采用 fabless 模式并有全栈野心(硬件 + 软件 + 系统),而 Kunlunxin 的商业动作仍高度依赖 Baidu 自身云和企业销售关系,独立渠道较弱。[CP007, CP008, CP009, CP010, CP011, CP012]
| 采购标准 | Kunlunxin (P800) | Huawei Ascend (910C/Atlas 350) | Cambricon | Moore Threads |
|---|---|---|---|---|
| 相对 Nvidia H100/H20 的推理性能 | DeepSeek R1 上 ~1,521 tok/s(Morgan Stanley 估计);介于 H20 与 A100 之间 | 约 H100 的 60-80%(910C);Atlas 350 声称为 H20 的 2.87x(Huawei 声称,未获独立验证) | 所审来源未给出独立基准测试 | 平湖架构:32T FP32,高于 A100、低于 H100 |
| 已验证训练集群规模 | 30,000 颗 P800 集群(Baidu,2025 年 4 月);此后已有 10,000+ 卡集群 | 有大规模集群报道,但所审来源未独立量化 | 所审来源未独立量化 | 获得 10,000 卡集群订单(Kua'e 集群,2026 年 3 月,RMB660M) |
| CUDA 替代软件栈 | XTCL / XBLAS / XDNN(编译器 + 数学库) | CANN | 所审来源未详述 | MUSA(带 Musify 迁移工具) |
| 晶圆代工 / 制造准入 | SMIC 7nm(推测;未获官方确认) | SMIC 7nm,据报道享有生产优先级 | 所审来源未详述(限制前历史上依赖 TSMC) | Fabless;所审来源未详述代工厂 |
| 上市状态(截至 2026 年 8 月) | 未上市;已秘密递交 HKEX 申请,并接受科创板辅导 | 未单独上市(Huawei 子公司) | 已上市(上海科创板) | 已上市(上海科创板,2025 年 12 月) |
| DeepSeek 模型家族兼容性 | 完整支持 R1 / V3(2025 年 2 月);V4 当日适配(2026 年 4 月) | V4 当日适配(2026 年 4 月) | V4 当日适配(2026 年 4 月) | V4 当日适配(2026 年 4 月) |
标记为“所审来源未详述”的单元格反映真实的公开证据缺口,并非确认缺乏该能力;若干厂商(Biren、Iluvatar CoreX、Enflame、Alibaba T-Head)未列入矩阵列,以保持可读性,而是在竞争对手画像表(TP001)中覆盖。
[CP007, CP008, CP009, CP010, CP022, CP023]| 厂商 | 单价或合同模式 | 包含能力 | 折扣或未知项 | 影响 |
|---|---|---|---|---|
| Kunlunxin P800 | 无公开标价;通过企业 / 云直签合同销售(如 China Mobile 集中招标) | 芯片 + XTCL / XBLAS 软件栈 + 集群集成支持 | 未披露单价或折扣结构 | 国内厂商普遍定价不透明;比较只能依赖 BOM 成本代理指标。 |
| Alibaba T-Head PPU | 不单独销售;嵌入 Alibaba Cloud 实例 | 云实例级访问;据报道 Alibaba Cloud 计划借国产芯片把推理价格下调 ~50% | 底层芯片单位经济性未披露 | 一旦国产芯片成本与 Nvidia 持平,可能引发激烈云价格竞争。 |
| Moore Threads MTT S80(消费级参考点) | 消费 GPU 定价,与 Nvidia RTX 3060 大致同档 | 一张卡同时提供图形和 AI 计算 | 企业 / 数据中心卡价格未披露 | 唯一有公开消费级价格参考点的国产厂商,可粗略比较 BOM / 定位。 |
| Huawei Ascend | 无公开标价;大规模政府 / 企业采购合同 | 芯片 + CANN 栈 + 系统集成 | 未披露单价 | SMIC 生产优先级可能让 Huawei 拥有结构性成本优势,但标价看不出来。 |
| Nvidia H20(现状参考) | 历史上通过出口许可销售;在中国供应时断时续 | 完整 CUDA 生态访问 | 据报道,获许可销售需向美国财政部支付 15% 费用 | 是每家国产厂商定价和营销时对标的性能 / 生态门槛。 |
本表没有任何厂商(包括 Kunlunxin)发布独立单元标价;这些数据是代理指标(BOM 成本主张、消费级 GPU 对比、云定价计划),不是已确认价目表。
[CP011, CP012, CP024, CP025]以出货规模和公开市场验证两个维度看国产 AI 芯片厂商,Kunlunxin 落在中等规模、生态绑定较强的位置;但截至本报告生成日,它不同于 Cambricon、Moore Threads、Biren 和 Iluvatar CoreX,仍未 IPO。
X 轴近似表示 2025 年中国出货规模(0-100,以 Nvidia 作为参考上限);Y 轴近似表示公开市场验证(0=完全私有,100=大型上市市值)。两轴都是本报告基于所引出货和上市证据给出的序数评分,不是来源方发布的指数。
[CP002, CP009, CP016, CP019, CP021]Kunlunxin 的软件栈和集群规模证据大体可与直接国产同行相比,但在这一组主要厂商中,它是唯一仍未上市、无法用公开市场独立验证资本状况的公司。
[CP007, CP010, CP017, CP018, CP022]3.3 护城河耐久度、切换成本与反向竞争证据
Kunlunxin 最常被引用的护城河,是与 Baidu 自身 DeepSeek 级工作负载直接集成(P800 集群用全国产芯片栈训练 Baidu Wenxin 4.5 Turbo,后续又训练 Wenxin 5.1 模型),以及正在扩大的外部客户名单,其中已包括竞争性超大规模云厂商 Tencent——多家媒体将其视为最高等级市场验证,可类比 Apple 从竞争对手 Samsung 采购 OLED 面板。但这条护城河并不独特:2026 年 4 月 DeepSeek-V4 发布当天,Huawei Ascend、Cambricon、Hygon、Moore Threads、Alibaba T-Head、Kunlunxin 等都完成了当天适配,说明 DeepSeek 兼容性本身已经是入场券,而非持久差异化。对 Kunlunxin 地位更严肃的威胁是制造准入:据报道,SMIC 在晶圆分配上优先 Huawei,排在 Cambricon 和 Kunlunxin 等其他国产客户之前;Huawei 已将芯片良率约翻倍至接近 40%,形成小厂难以复制的结构性生产优势。Biren Technology 在 2023 年美国 Entity List 指定下仍继续运营,加上 2025 年美国出口管制扩大覆盖中国本土半导体生产,说明随着 Kunlunxin 商业存在感提升,它本身也可能面临类似指定,尽管已审阅来源未确认 Kunlunxin 当前在清单上。Kunlunxin 2025 年约 116,000 块出货量(按 IDC 数据与 Cambricon 持平)远远落后于 Huawei 约 812,000 块和 Alibaba T-Head 265,000 块,意味着 Kunlunxin 当前基于规模的护城河在头部国产厂商中最弱,而不是最强。[CP014, CP015, CP016, CP017, CP018, CP019]
| 护城河主张 | 威胁 | 严重性 | 缓解措施 / 尽调问题 |
|---|---|---|---|
| Baidu 生态 DeepSeek 集成(P800 在全国产栈上训练 Wenxin 4.5/5.1) | DeepSeek 模型兼容已成入场券:Huawei、Cambricon、Moore Threads 等都在 2026 年 4 月完成 DeepSeek-V4 当日适配。 | 高 | 评估 Kunlunxin 的 DeepSeek 集成在当日兼容之外还有何差异化(如成本、延迟、大规模集群效率)。 |
| 跨竞争对手客户验证(Tencent 2026 年采购 P800 芯片) | 单个标杆客户赢单本身不足以证明相对 Huawei 更大出货基数的持久份额。 | 中 | 跟踪 Tencent 订单会发展成持续扩大的关系,还是停留在一次性验证采购。 |
| 全栈 CUDA 替代软件(XTCL/XBLAS/XDNN) | Huawei CANN 和 Moore Threads MUSA 是更成熟、公开文档更多的软件生态,并配有专门迁移工具(Musify)。 | 中 | 要求提供 Kunlunxin 软件栈相对 CANN / MUSA 的独立开发者采用数据(贡献者、迁移案例研究)。 |
| 通过 SMIC 获得制造准入 | 据报道,SMIC 将 Huawei 的晶圆分配优先级置于包括 Cambricon 和 Kunlunxin 在内的其他国内客户之上,这是小厂商难以复制的结构性劣势。 | 高 | 确认 Kunlunxin 的具体晶圆分配优先级,以及是否与 SMIC 签有长期供应协议。 |
| 规模(出货量) | Kunlunxin 2025 年 ~116,000 颗出货量远低于 Huawei(~812,000)和 Alibaba T-Head(~265,000),在国内头部厂商中规模最弱。 | 高 | 跟踪 2026 年出货数据,判断 Kunlunxin 的外部客户赢单能否转化为份额提升。 |
| 美国直接出口管制敞口 | Biren Technology 2023 年被列入美国实体清单,说明中国 AI 芯片厂商一旦知名度和涉出口能力提升,就可能面临美国直接限制。 | 中 | 监测 BIS 实体清单是否专门新增 Kunlunxin,区别于更广泛的中国半导体行业限制。 |
严重性评级是本报告基于所引证据作出的定性评估;按该表类型要求,各行按严重性降序排列。
[CP014, CP015, CP016, CP017, CP018, CP019]Kunlunxin 在公开市场验证和出货规模上落后于头部国产同行,但有一个差异化数据点领先:已确认拿下跨阵营超大规模云厂商客户 Tencent。
[CP015, CP019, CP020, CP021]04财务情况
4.1 收入来源、变现与商业化路径
Kunlunxin 的核心收入机制是直接销售芯片和集群——AI 加速器硬件加 XTCL/XBLAS/XDNN 软件栈——既卖给 Baidu 自身内部 AI 基础设施,也通过集中招标(如 China Mobile 中标)或直接合同(Tencent、China Merchants Bank、China Southern Power Grid、Geely、Vivo)卖给外部企业和政府客户。任何 Kunlunxin 芯片都没有公开价格表;最接近的公开变现信号,来自 Baidu 自身说法:随着外部客户放量,Kunlunxin 收入“远超”往年;以及 J.P. Morgan 2025 年 11 月估计,芯片相关收入可能从 2025 年约 RMB13 billion 增长六倍以上至 2026 年 RMB83 billion。商业化仍主要走 Baidu 自身云和企业销售关系,而非独立搭建渠道,这与 Kunlunxin 在 2021 年拆分前作为 Baidu 内部芯片项目的起点一致。据报道,外部(非 Baidu)收入从 2024 年约占总收入 40% 升至 2025 年超过 50%,这是现有最清晰的收入结构质量代理指标,但仍只是媒体估计,不是经审计分部披露。[CI001, CI002, CI003, CI004, CI005, CI006]
| 收入来源 | 机制 | 单位 | 当前数值 / 状态 | 质量 | 尽调问题 |
|---|---|---|---|---|---|
| Baidu 内部 AI 基础设施销售 | 向 Baidu 自有云和模型训练业务直接销售芯片 / 集群(关联方交易) | RMB(单价未披露) | 最大单一收入来源;媒体估计约占 2024 年收入的 ~50-60%,2025 年降至 ~40-50% | 关联方收入;定价条款未获独立披露 | 从 HKEX 招股书索取关联方定价 / 转让定价细节。 |
| 外部企业 / 国企招标销售 | 集中采购招标(如 China Mobile)和企业直签合同 | 每份合同 RMB(如 2025 年 8 月 China Mobile RMB1B+ 赢单) | 增长最快的收入来源;据报道外部收入占比从 ~40%(2024)升至 >50%(2025) | 有具名、有日期的合同证据;目前最强的收入质量信号 | 确认 China Mobile 订单的合同期限、续约条款和收入确认处理。 |
| 跨竞争对手超大规模云销售(Tencent) | 向竞争性超大规模云厂商直接商业销售 P800 芯片 | RMB(合同价值未披露) | 2026 年新增;合同价值未披露 | 质量最高的外部验证信号,但关系单一且刚披露 | 索取 Tencent 关系的合同规模,以及是持续性还是一次性结构。 |
| 汽车 / ADAS 销售(Geely;BYD 状态未确认) | 面向车载 AI / ADAS 计算的直接销售 | RMB(单价 / 合同价格未披露) | Geely 已确认为客户;BYD 仅确认为股权投资者,尚未确认是芯片客户 | 公开披露很薄;销量和价格不可得 | 确认 BYD 是否购买 Kunlunxin 芯片,还是纯财务 / 战略投资者。 |
| IPO 挂钩基石芯片采购承诺 | 据报道,基石投资者需购买价值为其股权投资 3-7x 的芯片 | USD/RMB(比例已披露;绝对金额未披露) | 2026 年 6 月报道称属于香港 IPO 结构的一部分;尚未进入公开招股书 | 结构上与 IPO 本身绑定;引发收入质量上的循环融资疑问 | 招股书公开后获取最终基石投资者芯片采购协议。 |
收入组合比例是媒体估计,不是经审计的分部披露;在所审任何来源中,Kunlunxin 都未发布按收入来源拆分的收入。
[CI001, CI002, CI003, CI005, CI006, CI019]| 厂商 / 合同 | 单价或合同模式 | 标价 vs. 实际价格 | 折扣 / 未知项 | 来源 |
|---|---|---|---|---|
| Kunlunxin P800 芯片 / 集群合同 | 未披露公开标价;所有已知交易都是谈判形成的企业 / 招标合同 | 所审来源无法区分标价与成交价 | 未披露折扣结构 | 关于具名合同赢单(China Mobile、Tencent)的媒体报道;无公司价目表。 |
| Baidu 内部转让定价 | 未知内部销售按市场价、成本加成还是其他基础定价 | 未披露 | 关联方定价条款完全未披露 | 根据 Baidu 自身投资者关系材料中对子公司和内部供应商 Kunlunxin 的描述推断。 |
| 可比同行定价信号(Alibaba T-Head PPU) | Alibaba 声称其自有竞品芯片 BOM 成本比 Nvidia H20 低 ~40% | 不能直接迁移到 Kunlunxin;仅作为粗略行业成本定位代理 | Kunlunxin 专属 BOM 成本未披露 | EE Times 关于 Alibaba T-Head 的报道;这里只作为行业成本定位参考。 |
| J.P. Morgan 芯片收入增长估计 | 隐含的实际收入增长路径(RMB13B 到 RMB83B,2025 到 2026),不是单价 | 不是单价;是收入总额预测 | 六倍增长估计背后的假设未披露 | BigGo Finance 引述的 2025 年 11 月 J.P. Morgan 研究报告。 |
所审来源没有提供 Kunlunxin 的真实单价或合同价值拆分;每一行都是代理指标或推断机制,而非已确认定价。
[CI004, CI007, CI013]4.2 成本结构、利润率驱动与资本密集度代理
已审阅来源没有披露 Kunlunxin 毛利率、R&D 支出或制造成本拆分。最直接代理来自两家阶段相近、刚上市的国产 GPU 同行,它们在 2026 年 3 月提交了经审计 HKEX 年度业绩:Iluvatar CoreX 报告 FY2025 收入 RMB1,033.6 million(+91.6% YoY)、毛利 RMB558.0 million、调整后净亏损 RMB437.7 million;Biren Technology 报告 FY2025 收入 RMB1,034.6 million(+207.2% YoY)、毛利率 53.8%、R&D 费用 RMB1,476.1 million、法定亏损 RMB16,493.0 million(调整后 non-IFRS 口径为 RMB873.8 million,剔除上市和股权激励费用)。Biren 自身招股书另行披露,其毛利率从 2022 年 100%(早期销量接近零)降至 2025 年上半年的 31.9%;公司转向商业规模部署后,毛利率随之下行——本章将这一路径视为 Kunlunxin 最相关的利润率路径代理,因为 Kunlunxin 收入规模大体可比,却没有披露自身利润率。制造上,鉴于 2021 年后出口管制限制 TSMC 访问,市场普遍认为 Kunlunxin P800(Kunlun III)采用 SMIC 7nm DUV 制程生产,但已审阅来源没有正式确认 Kunlunxin 具体代工方;SMIC 7nm 节点估计比 TSMC EUV 等效节点成本高 25-50%,良率也明显更低,这是每个依赖 SMIC 的国产厂商共同承受的结构性成本逆风。[CI007, CI008, CI009, CI010, CI011, CI012]
| 指标 | 数值 / 空值 | 信心 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| 毛利率(Kunlunxin,已披露) | 低 | 外部销售放量后,Kunlunxin 的增长究竟增厚还是稀释利润率,取决于该指标。 | 向 HKEX 招股书索取 FY2024/FY2025 毛利率。 | |
| 毛利率(Biren Technology,可比同业,FY2025) | 53.8% | 中 | 在收入规模和业务阶段相近的公司中,这是最接近的经审计可比项。 | 确认 Kunlunxin 的利润率轨迹是接近 Biren(放量后从 100% 降至 31.9%),还是有所不同。 |
| 毛利率(Iluvatar CoreX,可比同业,FY2025) | ~54% | 中 | 第二个收入和阶段相近的经审计可比项,佐证新商业化国产 GPU 厂商大致可落在 50%+ 毛利率区间。 | Kunlunxin 自身披露后,再交叉核对。 |
| 研发费用(Kunlunxin,已披露) | 低 | 研发强度反映资本密集度和技术投入纪律。 | 向 HKEX 招股书索取 FY2024/FY2025 研发费用。 | |
| 研发费用(Biren Technology,可比同业,FY2025) | RMB1,476.1M | 中 | 展示一家阶段相近的国产 GPU 厂商,在约 RMB1B 收入下仍需维持的研发投入规模。 | 仅作方向性代理,因为 Biren 的产品组合和客户基础不同。 |
| 净亏损(Kunlunxin,已披露) | 低 | 该指标决定公司距离盈亏平衡有多近,也是 IPO 叙事的关键说法(CNBC 报道 2025 年已实现盈亏平衡)。 | 向 HKEX 招股书索取经审计净利润 / 净亏损,以检验盈亏平衡说法。 | |
| 净亏损(Iluvatar CoreX,可比同业,FY2025,调整后非 HKFRS 口径) | RMB437.7M | 中 | 阶段相近的同业即便毛利率约 54%,调整后仍然亏损。 | 交叉核对 Kunlunxin 所称盈亏平衡,采用的是可比调整后口径还是法定口径。 |
| 客户集中度(外部收入占比) | >50%(2025,据报道) | 中 | 直接影响收入质量评估,也决定公司对 Baidu 这个股东兼客户的依赖程度。 | 从招股书索取经审计的客户集中度拆分(前五大客户收入占比)。 |
Biren、Iluvatar CoreX 等同业数字来自各自经审计的 HKEX 年度业绩,仅作方向性代理;这些不是 Kunlunxin 自身披露数字。
[CI008, CI009, CI010, CI021, CI022]Kunlunxin 的收入机制从芯片 / 集群生产开始,经 Baidu 内部和外部客户渠道,最终落到一个仍未对齐的收入数字。
这座桥是定性的;已审阅来源没有提供 Kunlunxin 特定的收入到毛利的量化瀑布。
[CI001, CI002, CI005, CI021, CI022]由于 Kunlunxin 未披露单位经济输入,本图采用定性节点,并标注最接近的同行代理指标(Biren Technology、Iluvatar CoreX)。
本图所有量化值都是同行代理指标(Biren Technology 或 Iluvatar CoreX 经审计 FY2025 结果),不是 Kunlunxin 特定披露;应把这座桥视为行业模式示意,而非 Kunlunxin 特定模型。
[CI008, CI009, CI010, CI012]4.3 资本充足性与融资依赖
Kunlunxin 的资本状况最好通过公司概况章节确立的融资时间线来读,而不是在此重复:2021 年 4 月拆分轮估值约 $1.9 billion,2025 年 7 月 Series D 融资约 $283 million、投后估值约 $3.0 billion。对前瞻资本充足性真正重要的是,Kunlunxin 下一次重大资本事件是自身 IPO——2026 年 1 月保密 HKEX Form A1 申报,同时 2026 年 5 月启动并行 Shanghai STAR Market “辅导”流程——据报道路演指引从 2026 年 5 月约 $14.7 billion 升至 2026 年 6 月下旬约 $50 billion。不同寻常的是,Kunlunxin IPO 结构据报道要求基石投资者承诺购买其股权投资额 3 到 7 倍的芯片,等于在定价前提前塞入多年订单簿——这一结构同时充当融资和收入机制,分析师已提示潜在循环融资疑虑。Kunlunxin 也与 SMIC 和其他国产厂商一道,受益于中国 2024 年启动的约 $47.5 billion “Big Fund III” 国家半导体融资工具,但已审阅来源没有量化其中多少资本具体流向 Kunlunxin。已审阅来源没有披露 Kunlunxin 账面现金、月度烧钱速度、现金跑道,或任何债务 / 项目融资义务;在公开招股书出现前,前瞻资本充足性仍是本章最大尽调缺口之一。[CI014, CI015, CI016, CI017, CI018, CI019]
| 账上现金 | 月度烧钱 | 现金跑道(月) | 计划资金用途 | 下一轮触发条件 | 债务 / 项目融资义务 |
|---|---|---|---|---|---|
| 据报道,IPO 募资将用于产能扩张、研发和营运资金;这与 Biren、Iluvatar CoreX 等同业 HKEX 招股书将募资投向研发和计算解决方案开发的做法一致 | 保密 HKEX Form A1(2026 年 1 月)和同步推进的科创板辅导(2026 年 5 月)已经启动;IPO 本身就是下一次资本事件,而不是再一轮私募融资 |
已审阅来源均未披露 Kunlunxin 的现金头寸、烧钱速度、现金跑道或债务义务;本行每个单元格都是真实公开披露缺口, 而不是假设为零。“计划资金用途”和“下一轮触发条件”单元格只反映目前可获得的前瞻性资本信息——IPO 进程本身—— 不是公司层面的专项披露。
[CI014, CI017, CI018, CI023]在新闻估算尚未对齐的情况下,Kunlunxin 的 2025 年收入和隐含 IPO 市销率倍数区间都非常宽。
2025 年收入区间用 CNBC 约 $500M 作为下限,用 Tencent Tech/BigGo Finance「远超 RMB20B」(约 $2.9B)作为上限;$1.2B 基准情景反映 J.P. Morgan RMB13B(约 $1.8B)2025 年估计,并因汇率 / 估算不确定性打折;P/S 区间用据报道约 $50B 的 IPO 目标除以各收入边界。上述数字均未在公开招股书中确认。
[CI002, CI003, CI021, CI024]Kunlunxin 的资本强度由几件事叠加:推测依赖 SMIC、受产能约束且成本劣势的制造基础,未披露的现金状况,以及即将到来的 IPO 这一近期主要融资事件。
[CI011, CI013, CI015, CI016]4.4 财务结论与披露缺口
Kunlunxin 的财务画像目前还不足以支撑对收入质量、利润率路径或资本密集度的高信心判断。最大问题是 2025 年收入估计无法调和:CNBC 数字为约 RMB3.5 billion(~$500 million,盈亏平衡),而 BigGo Finance 引述 Tencent Tech 的数字为“远超” RMB20 billion(~$2.9 billion)——相差约 5-6x,本报告不尝试向任何一边强行调和。取决于采用哪个数字,Kunlunxin 在 ~$50 billion IPO 目标下的隐含市销率约 17x 到 100x,从激进但仍可辩护(参照 Cambricon 自身 55-85x 区间)到即便按中国 2025-2026 国产芯片 IPO 热潮标准也极端。Kunlunxin 从未披露毛利率、R&D 支出占收入比例、员工数或资本充足性状态(现金、烧钱速度、现金跑道),制造代工方(SMIC 还是任何残余 TSMC 暴露)仍是推断而非确认。未来任何招股书中出现的官方定价,应被解读为标价或名义合同金额,而非已实现利润率,因为当前公开披露基础很薄。本章表格中的每个 null 字段都代表真实的公开披露缺失,而不是假定为零,并且都在下方表格中配有具体尽调路径。[CI021, CI022, CI023, CI024, CI025]
| 缺失的私有指标 | 影响 | 精确尽调路径 |
|---|---|---|
| 经调和的 2025 年收入数字 | 没有该数字,无法有把握地分析市销率或增长率;当前估计大致从 $500M 到 $2.9B,未调和区间达 5-6x | HKEX 招股书公开后,获取经审计 FY2025 收入;在此之前,将所有媒体数字视为估计。 |
| 毛利率和成本结构 | 无法评估增长是否增厚利润率;同业代理指向 50%+ 区间,但 Kunlunxin 自身数字未披露 | 从招股书索取 FY2024/FY2025 毛利率和销货成本拆分。 |
| 员工数 | 限制其与上市同业在规模和成本结构上的比较(Cambricon、Moore Threads、Biren、Iluvatar CoreX 均在申报文件中披露员工数) | 直接向 Kunlunxin/Baidu 投资者关系或 HKEX 招股书索取员工数。 |
| 现金头寸、烧钱速度和现金跑道 | IPO 前无法独立评估资金充足性 | 向 HKEX 招股书索取资产负债表和现金流细节。 |
| 制造代工确认(SMIC vs. 残余 TSMC 敞口) | 影响成本结构和供应链风险评估;行业共识假设为 SMIC 7nm DUV,但公司未确认 | 从招股书索取代工厂和供应协议披露,或取得公司直接确认。 |
| IPO 承销团和最终发行条款 | 无法评估交易可信度、定价纪律和分配结构 | 追踪 HKEX 文件,等待 Kunlunxin 从保密申报转为公开申报后发布公开上市文件。 |
每一行都反映本章审阅来源中识别出的真实公开披露缺口,不是假设为零或负面发现;Kunlunxin 的 HKEX 招股书公开后,这些条目将被替代。
[CI021, CI023, CI024, CI025]05产品与技术
5.1 产品定义与客户工作流契合度
从客户工作流看,Kunlunxin 销售 AI 加速器硬件(Kunlun 芯片家族)加 CUDA 替代软件栈,让模型开发者无需为新硬件目标重写代码,就能训练和服务大模型。模型训练客户的工作流是:在 PaddlePaddle、TensorFlow 或 PyTorch 中编写或移植模型;通过 Kunlunxin 的 XTCL 图编译库(基于 TVM)编译;再借助类似 CUDA 的 kernel/launch 编程模型和 PCIe 上显式 host-device 内存传输,在 XPU-R 硬件上运行。推理客户场景中,Kunlunxin 自身案例材料描述了一个工业质检场景:用 Kunlun-Paddle 方案替换基于 GPU 的 PyTorch 流水线,报告成本下降 65%、性能提升 9%——这是公司来源,但收益量化具体。集群规模上,Kunlunxin 基于 P800 的 “Super-Node” 产品(32/64 卡单柜配置)面向互联网、金融、能源、交通和教育客户,这些客户需要多卡训练或高吞吐推理,但不想自己搭定制集群基础设施。自 2018 年以来,Kunlun 芯片与 PaddlePaddle 的集成从 Paddle Lite 仅推理支持,推进到 PaddlePaddle 2.0 及后续版本的完整训练支持;切换到 XPU 后端据 Kunlunxin 称只需改一行代码——本章将这一说法视为公司来源,而非独立基准验证。[CE001, CE002, CE003, CE004, CE005, CE006]
| 模块 / 资产 / 产品线 | 用户 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| Kunlun I 芯片(K200 / CK10 / CK20) | Baidu 搜索引擎、DuerOS、早期云客户 | 2018-2020 年起量产;上一代产品 | 首款 Kunlun 专用芯片;Samsung 14nm 代工 | 未找到独立拆解,也未找到良率或现场可靠性数据。 |
| Kunlun II(CR20 / R200) | Baidu 云 / 边缘、Apollo 自动驾驶平台、Ernie Bot | 2021 年 8 月起量产;仍在现役集群中 | TSMC 7nm;定位为 Nvidia A100 级性能 | 未找到在标准化工作负载上直接比较 Kunlun II 与 A100 的独立基准测试套件。 |
| P800(Kunlun III) | Baidu 内部(Wenxin 模型)、外部企业 / 国企客户、Tencent(2026) | 当前旗舰;2025 年起投产;30,000 颗芯片集群于 2025 年 4 月启用 | 7nm、96GB HBM3、DeepSeek R1/V3 全版本兼容(2025 年 2 月) | 代工厂未获官方确认(外界认为是 SMIC 7nm DUV);未找到独立 H100/H200 基准测试。 |
| M100(推理优化) | 基于 MoE 的大模型推理客户 | 2025 年 11 月发布;2026 年 1 月起商业爬坡;2026 年 7 月在 CCTV 公开亮相 | 专门面向混合专家推理效率优化 | 截至报告生成日,未发布独立性能数据。 |
| M300(训练优化) | 超大规模多模态模型训练客户 | 2025 年 11 月发布;目标 2027 年发布(尚未出货) | 面向万亿参数多模态模型训练 | 完全前瞻;尚无已出货产品,也没有独立规格确认。 |
| XTCL / XBLAS / XDNN / XCCL 软件栈 | 所有 Kunlunxin 硬件客户;PaddlePaddle、PyTorch、TensorFlow 开发者 | 持续维护;2018 年以来与 PaddlePaddle 多年集成 | 类 CUDA 编程模型;基于 TVM 的图编译器;GitHub 开源组件 | 除 GitHub 仓库页面外,未找到独立开发者采用指标(贡献者数量、下载量)。 |
| Super-Node 集群产品(32/64 卡机柜) | 互联网、金融、能源、交通、教育行业客户 | 2025 年推出;公司称相较传统服务器带宽达 8x | 单柜高密度多卡配置 | 带宽和性能说法来自公司口径;未找到独立集群基准测试。 |
状态 / 成熟度和差异化单元格在有资料时合并了公司披露与独立媒体 / 技术报道;缺少独立佐证的单元格已在尽调缺口列标出。
[CE001, CE007, CE014, CE016, CE019]| 用户任务 | 当前流程 | 公司方案 | 可量化收益 | 限制 |
|---|---|---|---|---|
| 以超大规模训练大语言模型 | 采购 / 调配 Nvidia(或其他供应商)GPU 集群;编写面向 CUDA 的训练代码 | P800 集群(如 30,000 颗芯片部署)加 XTCL/PaddlePaddle 训练流水线 | 据报道,Baidu 集群在全国产芯片栈上训练了 Wenxin 4.5 Turbo,后来又训练 Wenxin 5.1 | 据独立报道,从 CUDA 代码库迁移需要投入可观工程时间。 |
| 为 DeepSeek 级开源模型提供推理服务 | 视可得性和成本,部署在 Nvidia H20/A100/H100 或国产替代方案上 | 单台 8 卡 P800 服务器以 8-bit 运行完整 671B 参数 DeepSeek R1 模型(公司声称) | Morgan Stanley 估计约 1,521 tokens/second,将 P800 置于 Nvidia H20 与 A100 之间 | 仅找到一项第三方性能估计;不是可复现的标准化基准测试。 |
| 为自动驾驶感知和规划(Apollo 平台) | 在车端或云端仿真算力上运行感知 / 规划模型 | Kunlun II 集成进 Baidu 的 Apollo 自动驾驶软件平台 | 公司称 Kunlun II 性能与同期 Nvidia 芯片相当 | 未找到独立车规级认证或安全测试数据。 |
| 自动化工业视觉质检 | 人工目检,或基于 GPU 的 PyTorch 推理流水线 | Kunlun-Paddle 推理方案(公司案例) | 公司称相比原 GPU 流水线,成本降低 65%,性能提升 9% | 单一公司案例;未找到独立复现。 |
| 作为企业客户微调大模型 | 租用 GPU 算力,或自建较小集群 | Baidu P800 集群支持最多 1,000 家企业客户同时微调(公司声称) | 中小企业无需自有专用基础设施,也能微调大模型 | 1,000 家客户的规模说法来自公司;未找到独立使用数据。 |
“可量化收益”单元格混合了独立来源数字(Morgan Stanley 估计)和公司披露说法;限制列标出哪些行缺少独立佐证。
[CE002, CE003, CE004, CE014, CE017]5.2 架构与运行模型
Kunlunxin 的 XPU-R 架构把计算、存储、互连和接口拆成四个硬件模块。计算模块结合 SDNN(用于张量运算的软件定义神经网络引擎)和通用 Cluster 单元;Baidu 自身技术沙龙披露的第二代设计写明,8 个 Cluster 单元和 6 个 SDNN 单元可在 FP16 下提供最高 128 TFLOPS,每个 Cluster 含 64 个核心和分层内存系统(本地内存、共享内存、64MB 片上 L3 SRAM,再到 512GB/s 外部 GDDR6)。当前 P800(Kunlun III)一代转向 7nm 制程,按独立规格汇总配备 96GB HBM3 内存和 2.4-4 TB/s 带宽;TechInsights 独立芯片级 floorplan 分析确认其为 7nm 制程、超过 50 billion 个晶体管,并将矩阵计算与通信单元物理分离——Kunlunxin 称这一架构选择提高了集群扩展效率。软件栈分层包括 runtime(驱动、监控、调试器、profiler)、基于 LLVM 的编译工具链(自定义 Clang 前端、XPU 后端、AOT/JIT 编译)、高性能数学库(XBLAS/XDNN,Kunlunxin 对应 cuBLAS/cuDNN 的库)、集合通信库(XCCL),以及基于 TVM 的 XTCL 图编译器,可从 PaddlePaddle、TensorFlow 和 PyTorch 导入模型。部署通过容器镜像和 Python wheels(经 FastDeploy XPU 后端)以及社区维护的 vLLM-Kunlun 插件实现,后者把 PyTorch 操作映射到 XBLAS/XDNN 用于大模型推理——二者都以 GitHub 开源仓库分发,给 Kunlunxin 带来多数中国纯硬件竞争者没有的公开开发者信号。[CE007, CE008, CE009, CE010, CE011, CE012]
| 层 / 组件 | 作用 | 依赖 | 风险 |
|---|---|---|---|
| XPU-R 计算模块(SDNN + Cluster 单元) | 执行张量运算(SDNN)和通用并行计算(Cluster 单元) | Kunlunxin 自有架构设计团队 | 性能说法(公开资料中该代为 128 TFLOPS FP16)来自公司披露,未经过独立基准测试。 |
| 内存层级(本地 / 共享内存、L3 SRAM、GDDR6/HBM3) | 提供分层、低延迟片上内存,以及更高容量的外部内存 | HBM3/GDDR6 内存供应商(历史上为 Samsung/SK Hynix;国产为 CXMT,Kunlunxin 是否采用未确认) | HBM 供应是全行业瓶颈;Kunlunxin 具体内存来源未确认。 |
| 互连(芯片间链路) | 支持集群内多芯片大规模分布式训练 | 集群级网络设计(自研;细节未充分披露) | 未看到对 5,000+ 颗芯片集群 90%+ 扩展效率说法的独立验证。 |
| PCIe 接口 | 承载训练 / 推理工作负载的主机-设备数据传输 | 行业标准 PCIe 规范(公司技术材料称为 4.0,向后兼容 3.0) | 标准接口;差异化和风险低于计算层或内存层。 |
| 基于 LLVM 的编译器工具链 | 自定义 Clang 前端和 XPU 后端;AOT/JIT 编译;设备 / 主机分离 | LLVM 开源项目作为底层编译器基础设施 | 相较 Nvidia 数十年 CUDA 工具链,工具链成熟度尚无独立基准测试。 |
| XTCL 图编译器(基于 TVM) | 导入并优化 PaddlePaddle、TensorFlow、PyTorch 模型,以适配 XPU 硬件 | Apache TVM 开源编译器框架作为底层图编译引擎 | 相较 Nvidia 成熟图编译工具,框架覆盖度和优化质量未验证。 |
| 运行时和驱动层 | 底层驱动、用户态 API、监控 / 调试 / 性能分析工具;支持多流和 SR-IOV | x86-64 和 Arm64 主机平台 | 虚拟化(SR-IOV)和多租户成熟度来自公司描述,未独立测试。 |
架构细节主要来自 Baidu 托管的技术沙龙文章和独立芯片裸片级分析(TechInsights);风险评估反映每一层可获得独立佐证的程度, 而非已确认缺陷。
[CE007, CE008, CE009, CE010, CE011, CE012]Kunlunxin 的 XPU-R 芯片架构把计算、内存、互联和接口硬件置于多层软件栈之下,开发者获得类似 CUDA 的编程模型。
层级规格结合 Baidu 技术沙龙文章(记录 XPU-R 一代)与独立汇总的 P800 规格(flopper.io)和 TechInsights 独立芯片裸片分析;两代的精确规格并不相同,此处并列只是为了展示架构层次。
[CE007, CE008, CE009, CE010]Kunlunxin 的产品依赖不受其控制的代工厂、外部内存供应商、开源编译器生态,以及 Baidu 自身的云和客户关系来推进商业化。
SMIC 和内存供应商节点标记为「warning」,因为 Kunlunxin 具体代工厂和内存采购安排来自行业报道推断,并非公司官方确认。
[CE016, CE019, CE022]5.3 部署、路线图与差异化
Kunlunxin 最清晰的部署证据,是 Baidu 2025 年 4 月启用自有 30,000 颗 P800 芯片集群;该集群被描述为可训练数千亿参数的 “DeepSeek-like” 模型,同时支持最多 1,000 家企业客户微调工作负载。Baidu 还宣布计划到 2030 年把单个 Kunlun 集群扩展到百万卡,并把 P800 芯片组为 “Tianchi256”(2026 年上半年)和 “Tianchi512”(2026 年下半年)配置,每一代目标都比上一代集群再提升 50%+ 性能。Kunlunxin 2025 年 2 月公告称已完全兼容 DeepSeek R1/V3 训练和推理,单台 8 卡 P800 服务器可用 8-bit 精度运行完整 671-billion-parameter DeepSeek R1 模型,32 台服务器(256 颗 P800)足以做全参数训练——这些说法只有一个第三方估计(Morgan Stanley,DeepSeek R1 推理约 1,521 tokens/second)佐证,而不是可复现基准。差异化方面,Kunlunxin 最耐久的技术护城河是:(1)XTCL/XBLAS/XDNN 软件栈多年融入 PaddlePaddle,生态深度多数国产竞争芯片难以匹配;(2)直接针对 Baidu 自身生产工作负载(Wenxin/Ernie 模型、Apollo 自动驾驶栈)做架构调优。最暴露的差异化缺口是制造:早期 Kunlun I 芯片采用 Samsung 14nm 制程,Kunlun II 采用 TSMC 7nm 制程,但 2021 年后出口管制收紧,使 P800 很可能(虽未获官方确认)由 SMIC 7nm DUV 制程制造,该制程成本估计比 TSMC EUV 等效制程高 25-50%,良率也更低。[CE014, CE015, CE016, CE017, CE018, CE019]
| 控制 / 认证 / 质量指标 | 状态 | 范围 | 缺口 |
|---|---|---|---|
| 美国出口管制 / Entity List 被列名状态 | 截至报告生成日未确认 | 公司整体(Kunlunxin 作为法律实体) | BIS 尚未正式确认任一方向;2025 年全行业出口管制扩围带来潜在风险。 |
| 芯片制造良率 | 未披露 | 专指 P800(Kunlun III)生产 | 没有 Kunlunxin 专属良率数据;Huawei 披露的约 40% SMIC 良率是最接近但不可迁移的代理。 |
| 独立安全 / 可靠性认证 | 未披露 | 车载(Apollo)和数据中心部署场景 | 审阅来源中未识别到认证机构、标准或审计报告。 |
| 独立安全 / 隐私认证(如 ISO 27001 等同认证) | 未披露 | 云 / 集群部署和软件栈 | 未披露认证;部署被推断继承 Baidu Cloud 自身安全态势,但未确认。 |
| XPU 架构的知识产权 / 专利保护 | 审阅来源未量化 | 核心芯片架构和软件栈 | 未找到专利数量或诉讼历史;不同于 Biren 等披露知识产权定位的同业。 |
每一行反映已识别的公开披露缺口,并不代表底层控制已通过或失败;审阅来源中,Kunlunxin 尚未发布专门的信任 / 合规页面。
[CE021, CE022, CE023, CE024, CE025]| 日期 / 阶段 | 功能 / 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2018-07 | Kunlun I(K200)发布 | 已出货;部署于 Baidu 搜索和 DuerOS | 首个 Kunlun 专用芯片验证点 | MIT Technology Review;Wikipedia |
| 2021-08 | Kunlun II 量产 | 已出货;部署于 Apollo 和 Ernie Bot | 确立 Kunlunxin A100 级训练 / 推理供应商定位 | Tom's Hardware;TechInsights |
| 2025(H1) | P800(Kunlun III)发布并兼容 DeepSeek R1/V3 | 已出货;投产 | 把旗舰芯片直接绑定到中国关注度最高的开源模型家族 | Kunlunxin 官方公告;CSDN 开发者报道 |
| 2025-04 | 30,000 颗芯片 P800 集群启用(Baidu 开发者大会) | 已出货 / 运行中 | 证明集群规模部署,而不只是单芯片规格 | Reuters(经 Yahoo Finance);Wikipedia |
| 2025(H2) | Super-Node 集群产品发布(32/64 卡机柜) | 已出货 | 面向需要高密度多卡配置的多行业企业客户 | Kunlunxin 官方公告 |
| 2025-11 | M100/M300 路线图在 Baidu World 发布 | 已发布(发布时尚未出货) | Baidu/Kunlunxin 承诺至少到 2027 年按年迭代芯片 | 来源:SCMP;Yicai Global;China Daily |
| 2026(H1,计划) | Tianchi256 集群发布 | 计划中 | 目标较上一代集群性能提升 50%+ | Tech Wire Asia;TechPowerUp |
| 2026-01 | M100 商业爬坡开始 | 进行中;实物产品于 2026 年 7 月在 CCTV 亮相 | 标志着从路线图发布转向推理硬件出货 | Baidu Baike |
| 2026(H2,计划) | Tianchi512 集群发布 | 计划中 | 进一步瞄准集群性能,卡数较 Tianchi256 翻倍 | Tech Wire Asia |
| 2027(计划) | M300 训练芯片发布 | 计划中;尚未出货 | 面向万亿参数多模态模型训练 | SCMP;Baidu Baike |
| 2030(计划) | 单个 Kunlun 百万卡集群目标 | 长期路线图承诺 | 显示 Baidu 的长期集群规模野心 | Yicai Global |
2026-2030 年里程碑日期来自公司和媒体路线图披露的计划 / 目标,并非已确认出货产品;明确标注已出货 / 运行中的除外。
[CE001, CE014, CE015, CE016, CE018, CE019]Kunlunxin 客户的工作流从标准框架中的模型开发出发,经 Kunlunxin 编译器栈,到集群规模训练或推理部署。
[CE002, CE003, CE006, CE013]Kunlunxin 的产品成熟度在已出货代际(Kunlun I、II、P800)上最强,在前瞻承诺(M300)和全线性能主张的独立验证上最弱。
[CE014, CE015, CE016, CE019, CE020]5.4 信任、质量、合规与证据缺口
已审阅来源没有披露 Kunlunxin 的独立安全认证、安全测试标准,或芯片和集群产品的正式质量指标;这与云软件公司通常发布 SOC 2 或 ISO 认证形成对比。现有最相关“信任”信号是间接的:Kunlunxin 芯片部署在通过集中招标赢得的受监管买方基础设施中(China Mobile、China Merchants Bank、China Southern Power Grid),意味着产品通过了某种国企采购尽调,但这些招标采用的具体技术或安全标准并不公开。出口管制合规方面,已审阅来源没有确认 Kunlunxin 当前在美国 Entity List 上;但 2025 年更广泛、针对中国本土半导体生产的出口管制行动,制造了潜在合规风险,而已审阅的 Kunlunxin 信任或合规材料没有回应这一点。制造质量上,已审阅来源没有披露 Kunlunxin 具体芯片良率;最接近的代理是 Huawei 披露其在 SMIC 的良率改善(从约 20% 到接近 40%),而 Kunlunxin 作为规模较小且据称生产优先级低于 Huawei 的 SMIC 客户,未必能复制。凡本章来自 Kunlunxin 官网、技术沙龙或 GitHub 仓库的说法,均标注为公司声称而非独立验证;未解决事项(良率、Entity List 状态、正式认证、独立基准)则作为明确证据缺口带入后续,而非假定有利或不利。[CE021, CE022, CE023, CE024, CE025]
06客户情况
6.1 客户分层与采用轨迹
Kunlunxin 客户基础分为五类,买方 / 用户 / 付款方结构各不相同。Baidu 本身同时是所有者、最大客户和技术共同开发者,为自家 Wenxin/Ernie 模型训练和 Apollo 自动驾驶栈消耗 P800 芯片。电信和金融国企(China Mobile、China Merchants Bank)通过集中采购招标采购,代表案例是 Kunlunxin 2025 年 8 月在 China Mobile 三个通用计算标包中赢得超过 RMB1 billion 的订单。能源和工业国企(China Southern Power Grid)以及汽车 OEM(Geely;BYD 仅确认是股权投资者,而非已确认芯片客户)构成规模更小但战略可见的分层。最新且最重要的分层,是跨竞争对手的超大规模云采购:Tencent 2026 年开始规模采购 P800 芯片,多家媒体将该关系定义为最强可用市场验证信号,原因正是 Tencent 有充分动机挑出竞争对手产品的问题,而不是替它背书。采用轨迹数据很薄:据报道外部(非 Baidu)收入从 2024 年约 40% 升至 2025 年超过 50%;Kunlunxin 2025 年 AI 芯片总出货约 116,000 块(按 IDC 数据与 Cambricon 持平),明显落后于 Huawei 约 812,000 块;但已审阅来源没有按具名客户拆分出货量、活跃部署或利用率。[CU001, CU002, CU003, CU004, CU005, CU006]
| 客户分层 | 买方 / 用户 / 付款方 | 用例 | 规模 | 收入 / 战略价值 | 缺口 |
|---|---|---|---|---|---|
| Baidu(内部) | Baidu 同时是买方、用户和付款方(关联方) | 大语言模型训练 / 推理(Wenxin/Ernie)、Apollo 自动驾驶 | 最大单一客户;确切占比未披露,但历史上占收入多数 | 锚定需求并共同开发技术;也是 Kunlunxin 控股股东 | 关联方定价和收入确认条款未独立披露。 |
| 电信国企 | China Mobile(买方 / 付款方);企业 IT 团队(用户) | 面向通用算力的集中式 AI 计算采购 | RMB1B+(约 $139M)单一标的中标,2025 年 8 月 | 截至目前披露的最大外部商业合同 | 合同期限、续约条款和确认节奏未披露。 |
| 金融行业国企 | China Merchants Bank(买方 / 付款方);风控 / AI 应用团队(用户) | AI 芯片资源项目(BigGo Finance 称 2025 年 3 月中标) | 未披露 | 点名的头部银行客户;增强国企客户可信度 | 部署规模和结果指标未披露。 |
| 能源 / 工业国企 | China Southern Power Grid(买方 / 付款方);电网运营团队(用户) | 智能电网和工业 AI 现代化 | 未披露 | 分散电信 / 金融之外的行业敞口 | 部署规模和结果指标未披露。 |
| 汽车 OEM | Geely Automobile Holdings(买方 / 付款方,已确认客户);BYD(仅确认其为股权投资方) | ADAS / 车载 AI 算力 | 未披露 | 显示云 / 数据中心工作负载之外的跨行业适用性 | BYD 是否采购 Kunlunxin 芯片(而非仅投资)仍未确认。 |
| 竞争对手超大规模云厂商 | Tencent(买方 / 付款方);Tencent Cloud / AI 团队(用户) | 大规模推理 / 训练集群(据报道 10,000+ 卡) | 据报道为 2026 年;合同价值未披露 | 最高等级第三方验证信号;首个确认的非 Baidu 超大规模云厂商客户 | 合同规模、期限以及关系是否持续复购均未披露。 |
收入 / 战略价值单元格合并公司披露和独立媒体报道;规模数据是各细分可取得的最细粒度,不是完整的客户级收入拆分。
[CU001, CU002, CU007, CU008, CU009]| 指标 | 值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 外部(非 Baidu)收入占比 | ~40% | 2024 | BigGo Finance(来源于 Tencent Tech) | 中 | 报道称 2025 年上升之前的外部需求占比基线 | 支撑该占比的总收入数字本身存在争议(见「财务情况」章节)。 |
| 外部(非 Baidu)收入占比 | >50% | 2025 | BigGo Finance;多篇媒体报道 | 中 | 媒体将此视为 IPO 前验证 Kunlunxin 外部商业可行性的里程碑 | 与 2024 年数字一样,底层总收入仍不明确。 |
| 2025 年中国 AI 加速器总出货份额 | ~3%(约 116,000 张卡) | 2025 | IDC 数据,经 Reuters / Yahoo Finance | 高 | Kunlunxin 在国内厂商中并列第三,明显落后于 Huawei 约 20% 的份额 | 该总量没有客户级出货拆分。 |
| 具名外部客户数(已披露) | 6 | 2026-08-19 | 综合 Yicai Global、China Biz Insider、BigGo Finance 报道 | 中 | 相比公司宣称的「数百个真实世界用户」,具名账户基数仍小但在增长 | 没有来源解释 6 个具名账户与 Baidu 所称「数百个」之间的差距。 |
| China Mobile 合同金额 | RMB1B+ (~$139M) | 2025-08 | 来源:Yahoo Finance(Reuters);ZhiDongXi;Zhihu | 高 | 已披露的最大单项外部商业数据点 | 合同期限和续约条款未披露。 |
所有百分比均为媒体估计,而非经审计的分部披露;应把 40%/50% 的外部收入数字视为方向性信号,不是精确值。
[CU003, CU004, CU005, CU007, CU013]6.2 具名客户验证与部署证据
现有最强客户验证证据是 China Mobile 采购中标:中文采购报道(ZhiDongXi、Tencent News、Zhihu)确认,Kunlunxin 在约 RMB5 billion 的 China Mobile AI 计算招标中取得领先位置,三个通用计算标包全数中标,金额超过 RMB1 billion,芯片通过系统集成商 H3C 和 ZTE 供应——这是生产部署,不是试点。Yicai Global 2025 年 11 月报道另行点名 China Merchants Bank(2025 年 3 月中标 AI 芯片资源项目)、China Southern Power Grid、Geely Automobile Holdings 和 Vivo 是外部 P800 部署方,并称其覆盖互联网、金融、能源、交通和教育等领域的“数百个真实用户”——不过只有具名账户可独立核验,“数百个”是公司来源汇总。Kunlunxin 自身案例材料描述了一个工业质检部署,相较此前 GPU 流水线实现 65% 成本下降和 9% 性能提升,这是公司来源但结果量化具体。China Biz Insider 2026 年 7 月报道补充 Tencent 是 2026 年具名客户,在 10,000 卡集群中运行 P800 芯片,并明确将其描述为 Kunlunxin 已通过来自对抗性买方“最严苛真实压力测试”的证据。本章每个具名部署都指向生产使用,而非试点或概念验证;但除 Kunlunxin 自身单一案例外,已审阅报道均未披露同等颗粒度的部署规模、合同期限或结果指标。[CU007, CU008, CU009, CU010, CU011, CU012]
| 客户 | 细分 | 部署 / 用例 | 生产部署 / 试点 | 结果 | 局限 |
|---|---|---|---|---|---|
| Baidu(内部) | 超大规模云厂商(母公司 / 控制方) | Wenxin / Ernie 模型训练与推理;Apollo 自动驾驶栈 | 生产部署(30,000 芯片集群,2025 年 4 月;正扩展到 10,000+ 卡集群,并以 2030 年百万卡为目标) | 据媒体报道,Baidu 用全国产芯片栈训练了 Wenxin 4.5 Turbo(2025 年)以及后来的 Wenxin 5.1 | 关联方关系;不是独立第三方验证。 |
| China Mobile | 电信国企 | 集中式 AI 算力采购(通用计算标包) | 生产部署(2025 年 8 月中标,金额 RMB1B+) | 据 ZhiDongXi / Zhihu 采购报道,在全部三个通用计算标包中均取得领先位置 | 除合同金额外,部署规模、期限和最终用途细节未披露。 |
| Tencent | 跨竞争对手的超大规模云厂商 | 大规模推理 / 训练集群 | 生产部署(据报道 2026 年为 10,000+ 卡集群) | China Biz Insider 将其描述为 Kunlunxin 芯片目前可见的最高等级第三方验证 | 单一来源(The Information,经 China Biz Insider);未找到独立确认。 |
| China Merchants Bank | 金融业国企 | AI 芯片资源项目 | 生产部署(据 BigGo Finance,2025 年 3 月中标) | BigGo Finance 在融资历史报道中将其列为 Kunlunxin 的「顶级客户」之一 | 部署规模和结果指标未披露。 |
| 工业质检客户(未具名) | 制造 / 工业 | 视觉质量检测,替代基于 GPU 的 PyTorch 流水线 | 生产部署(公司案例研究) | 公司称成本降低 65%、性能提升 9% | 客户身份未披露;该案例研究仅来自公司,未被独立复现。 |
本表只列出已审阅来源中有具名、有日期部署引用的客户;Baidu 自称「数百个真实世界用户」(Yicai Global),但除这些具名账户外,没有独立逐项清单。
[CU007, CU008, CU009, CU010, CU011, CU012]Kunlunxin 的客户旅程从 Baidu 内部验证出发,经过国企竞争性采购,最近走到跨阵营超大规模云厂商直接采用。
该旅程表示已审阅来源中识别到的客户采用证据时间顺序,不是公司发布的客户生命周期模型。
[CU001, CU007, CU009, CU011, CU017]具名 Kunlunxin 客户的证据质量差异很大:从 Baidu 深但属关联方的证据,到 Tencent 验证价值高但只有单一来源的披露。
[CU007, CU009, CU011, CU012, CU022]6.3 留存、扩张与集中度风险
已审阅来源没有披露任何 Kunlunxin 客户的留存指标、续约率、流失率、净收入留存或合同期限数据——这是完整披露缺口,本章将其作为 null 字段处理,而不是假定值。最接近的留存代理是跨代重复购买:Baidu 自 2018 年起购买每一代 Kunlun 芯片(Kunlun I、II、P800,以及即将到来的 M100/M300),Geely 与 Kunlun 芯片的关系也可追溯到 Kunlun II/Apollo 时代,说明至少两个账户存在多年延续性,尽管这是从产品代际历史推断,而非披露的续约指标。扩张方面,最清晰信号是外部收入占比在 2024 至 2025 年从约 40% 升至超过 50%,以及 2026 年新增首个超大规模云竞争对手客户(Tencent)。集中度方面,核心风险是结构性的,而非偶发:Baidu 同时是 Kunlunxin 约 59.45% 控股股东和最大客户,意味着 IPO 的收入质量评估高度取决于关联方定价如何处理;据报道 IPO 芯片采购承诺(基石投资者购买其股权投资额 3-7x 的芯片)又加入第二层集中度风险,即从投资者客户提前装入收入,而不是有机获取账户。采购摩擦只在一个方向可见:具名国企中标来自正式、多标包集中招标(China Mobile),Kunlunxin 必须在与其他投标方竞争中胜出,这证明存在真实采购门槛,而不是只靠内部需求的销售模式。[CU013, CU014, CU015, CU016, CU017, CU018]
| 指标 | 值 / 空值 | 细分 | 置信度 | 尽调要点 |
|---|---|---|---|---|
| 净留存率(NRR) | 全部细分 | 低 | 从 HKEX 招股说明书索取队列级收入留存数据;Kunlunxin 是硬件加软件模式,SaaS NRR 的口径未必同样适用。 | |
| 续约 / 复购率 | 全部细分 | 低 | 从公司披露索取按具名账户拆分的重复订单数据。 | |
| 客户满意度 / 类 NPS 指标 | 全部细分 | 低 | 未找到这类硬件 / 芯片供应商的客户满意度调查或评价平台数据。 | |
| 跨代复购(代理指标) | 6 个具名账户中有 2 个(Baidu、Geely)显示跨代延续 | Baidu(内部);Geely(汽车) | 低 | 确认 Geely 和其他具名账户是否跨多个 Kunlun 芯片世代采购,还是迄今只采购单一世代。 |
| 合同期限(具名合同) | 客户:China Mobile;China Merchants Bank | 低 | 索取 China Mobile 和 China Merchants Bank 中标项目的合同期限披露。 |
Kunlunxin 是硬件 / 芯片供应商,不是订阅业务;标准 SaaS 留存指标(NRR/GRR)可能不能直接适用。跨代复购代理指标是本报告根据产品历史作出的推断,不是公司披露的留存统计。
[CU014, CU015, CU016]| 扩张驱动因素 | 集中风险 | 影响 | 尽调路径 |
|---|---|---|---|
| 外部收入占比上升(40% 到 50%+,2024-2025) | 即使外部占比上升,Baidu 内部收入仍是最大单项组成 | 决定 Kunlunxin IPO 增长叙事中有多少真正来自外部需求,而非母公司内部需求 | 索取经审计的客户集中度拆分(前 5 或前 10 大客户收入占比)。 |
| 新增跨竞争对手超大规模云厂商客户(Tencent,2026) | 如果 Tencent 关系不复购或不放量,单一客户依赖风险会上升 | 一次性验证采购会高估可持续外部需求 | 跟踪 Tencent 订单是否在后续披露中变成持续、放量的关系。 |
| 与 IPO 挂钩的基石芯片采购承诺(股权投资额的 3-7x) | 收入来自 IPO 投资者关系,而非自然商业需求 | 引发循环融资疑虑,即已报告外部收入是否持久、独立 | 招股说明书公开后,获取最终版基石投资者芯片采购协议。 |
| 政府 / 国企集中招标中标(China Mobile、China Merchants Bank) | 依赖周期性竞争招标,而非连续商业关系 | 这些账户的收入可能大额集中出现,并受招标周期影响,而非持续复购 | 确认招标续期周期以及 Kunlunxin 在历史招标轮次中的中标率。 |
| Baidu 既是控制方又是最大客户 | 结构性关联方集中,无法靠增加外部客户分散 | 使独立评估 Kunlunxin 真实外部商业牵引力更复杂 | 从 HKEX 招股说明书索取关联方交易披露和定价方法。 |
集中风险按本报告对已审阅证据的定性评估排序,不是公司披露的风险登记表。
[CU003, CU004, CU013, CU017, CU018]| 代理指标 | 指向什么 | 无法确认什么 | 来源 |
|---|---|---|---|
| Baidu 自 2018 年以来持续采购每一代 Kunlun 芯片 | Baidu 内部需求跨芯片世代持续、复购 | 这究竟是真实商业复购选择,还是受控关联方延续 | Wikipedia;Kunlunxin 官方材料 |
| Geely 自 Kunlun II / Apollo 时代起使用 Kunlun 芯片 | 至少一个外部汽车客户具备多年延续 | 当前部署规模,或 Geely 是否已具体采用 P800 | Wikipedia;Yicai Global |
| China Mobile 多标包中标(全部三个通用计算标包) | Kunlunxin 能在一个周期内竞争性赢下多个相关采购标包 | China Mobile 未来招标周期是否会续约或扩大关系 | 来源:ZhiDongXi;Zhihu;Yahoo Finance(Reuters) |
| 未见公开客户投诉、采购争议或部署失败报道 | 未找到已确认的负面客户体验信号 | 这究竟代表履历确实干净,还是只是缺少公开披露 / 报道 | 交叉核对本章审阅的全部来源 |
本表替代按时间桶划分的留存队列图,因为审阅来源没有披露按时间桶划分的百分比留存数据;每行都是推断代理指标,不是披露指标。
[CU014, CU015, CU019]潜在 Kunlunxin 客户从评估走到生产部署,漏斗在竞争性招标和集群集成阶段明显收窄。
各阶段数值只是基于具名客户中标证据给出的定性相对权重,不是来自单一数据集的实测转化率漏斗。
[CU007, CU008, CU013, CU014]07风险
7.1 监管与法律风险
最大未解监管问题,不是 Baidu 或中国 AI 芯片板块整体,而是 Kunlunxin 本身是否会被美国直接列入 Entity List 或纳入出口管制指定。到本报告运行日,已审阅来源没有确认这种指定已经存在,但风险并不轻:阶段相近的国内同业 Biren Technology 自 2023 年 10 月起就在 Entity List 下运营;BIS 2025 年 1 月出台规则,明确收紧晶圆代工尽调要求,并把中国 / 新加坡实体加入 Entity List。这延续了 2022 年 10 月、2023 年 10 月和 2024 年 12 月以来不断加码的管制脉络。GAO 审查确认,Commerce 曾协调另外六个联邦机构落地 2022-2023 年先进半导体出口管制规则;2025 年年终法律回顾(Mondaq/AlixPartners)还记录,一项拟议的「Affiliates Rule」曾被提出,原本会把管制延伸至由已受限主体持股 50% 或以上的实体,但该规则自 2025 年 11 月起暂停一年;2025 年全年另有超过 120 家中国实体被加入 Entity List。另一个口径来自台湾:2025 年 6 月,台湾对 Huawei 和 SMIC 实施新的出口限制,收紧任何 SMIC 客户获取先进封装、EDA 工具和材料的通道;如果 SMIC 代工归因属实,Kunlunxin 也在其中。公司与上市层面,Kunlunxin 同时推进香港 / STAR Market 上市,本身就带来监管流程风险:公司必须同时满足 HKEX 和 CSRC 披露制度;其少见的 IPO 芯片购买承诺结构(据报道,基石投资者需购买相当于其股权投资额 3-7x 的芯片),香港 SFC 或 CSRC 会要求如何写入招股书,目前没有清晰先例。[CR001, CR002, CR003, CR004, CR005, CR006]
| 规则 / 许可 / 案件 | 法域 | 状态 | 可能性 | 严重性 | 缓释措施 | 剩余风险敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| Kunlunxin 可能被美国 Entity List / 出口管制单独点名 | 美国(BIS) | 截至报告生成日未确认;全行业风险升高 | 中 | 严重 | Kunlunxin 未公开披露 | 高 | 专门监测 BIS Entity List 中 Kunlunxin 的更新;在 IPO 尽调中索取法律 / 合规确认。 |
| BIS 2025 年 1 月规则收紧晶圆厂尽调并扩大 Entity List | 美国(BIS) | 已生效 | 高 | 高 | 合规义务落在晶圆厂(SMIC)和出口链条上,不只由 Kunlunxin 承担 | 高 | 确认 SMIC 的尽调合规状态,因为这会影响 Kunlunxin 供应连续性。 |
| 拟议「Affiliates Rule」(覆盖 50% 持股子公司),自 2025 年 11 月起暂停一年 | 美国(BIS) | 已暂停而非废除;可能恢复 | 中 | 高 | 无;暂停是美国监管动作,不是 Kunlunxin 的缓释措施 | 高 | 跟踪 Affiliates Rule 是否在一年暂停期后恢复。 |
| 台湾对 Huawei 和 SMIC 的出口限制(2025 年 6 月) | 台湾 | 已生效 | 高 | 中 | 未披露;影响 SMIC 获取先进封装、EDA 工具和材料 | 中 | 评估台湾限制是否影响 SMIC 服务 Kunlunxin P800 订单的能力。 |
| HKEX / 科创板双重上市监管合规 | 香港 / 中国(HKEX,CSRC) | 推进中(2026 年 1 月起已保密递交 HKEX;2026 年 5 月起接受科创板辅导) | 中 | 中 | 标准双重上市法律 / 顾问流程(CICC 参与科创板路径) | 中 | 跟踪 HKEX / CSRC 文件中针对双重上市结构的招股书披露要求。 |
| IPO 芯片采购承诺的披露处理(无明确监管先例) | 香港(HKEX,SFC) | 未解决;结构已有报道,但尚未体现在公开招股说明书中 | 中 | 高 | 未披露 | 高 | 招股说明书公开后,监测 HKEX / SFC 如何要求披露芯片采购承诺。 |
各行按严重性降序排列。「可能性」和「严重性」反映本报告基于所引证据的定性评估,不是公司披露的风险登记表。
[CR001, CR002, CR003, CR004, CR005, CR006]7.2 运营、制造与质量风险
Kunlunxin 最具体的运营风险,是制造端依赖一家产能受限、成本大概率处于劣势的代工厂。行业共识(公司未确认)认为 P800 使用 SMIC 基于 DUV 的 7nm 工艺生产;相比 TSMC 基于 EUV 的同类工艺,该路线估计成本高出 25-50%,良率也明显更低。据报道,SMIC 在晶圆分配上优先满足 Huawei,Cambricon、Kunlunxin 等其他国内客户排在其后;小厂商很难复制或对冲这个结构性劣势。HBM 内存供应是第二个产能瓶颈:国内 HBM 厂商 CXMT 相比 Samsung/SK Hynix 仍不成熟,Kunlunxin 的具体内存采购安排也未披露。产品质量层面,已审阅来源没有披露 Kunlunxin 自身芯片良率、独立安全 / 可靠性认证或正式质量控制指标;最接近的代理指标是 Huawei 自己披露的良率改善(在 SMIC 从约 20% 升至接近 40%),但考虑到 Huawei 据报道享有生产优先权,这个指标不能外推到 Kunlunxin。已审阅来源也没有找到可独立复现的基准测试(例如 MLPerf 风格测试),在相同工作负载下比较 Kunlunxin 芯片与 Huawei Ascend、Cambricon 或 Nvidia 芯片;因此,竞争力和可靠性主张主要压在一个第三方估计(Morgan Stanley,DeepSeek R1 上 ~1,521 tokens/second)和公司来源材料上。[CR008, CR009, CR010, CR011, CR012, CR013]
| 失效模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余风险敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| SMIC 晶圆分配中 Huawei 优先,Kunlunxin 配额不足 | 高 | 高 | 低;未披露 Kunlunxin 专属缓释措施 | 高 | Kunlunxin 相对 Huawei 和 Cambricon 的实际晶圆分配优先级未确认。 |
| HBM 内存供应短缺(国产 CXMT 尚不成熟,Samsung/SK Hynix 受限) | 中 | 高 | 低;Kunlunxin 具体内存采购安排未披露 | 高 | Kunlunxin 是从国内(CXMT)采购 HBM,还是依赖受限进口,尚未确认。 |
| 相比 TSMC 制造的竞品芯片,制造良率更低 | 高 | 中 | 低;未披露 Kunlunxin 专属良率数据 | 中 | Kunlunxin 自身良率完全未披露;公开信息只有 Huawei 的良率改善。 |
| 缺少独立、可复现的性能基准 | 高 | 中 | 低;只有一个第三方估计(Morgan Stanley) | 中 | 没有标准化基准套件(如 MLPerf 风格)测试 Kunlunxin 芯片与同行对比。 |
| 缺少已披露的安全性 / 信息安全 / 质量认证 | 中 | 中 | 低;审阅来源未披露任何认证、标准或审计 | 中 | 未识别出任何 Kunlunxin 产品的认证机构、标准或审计报告。 |
各行按严重性降序排列。可能性 / 严重性 / 缓释成熟度评级反映本报告的定性评估;未找到公司发布的风险登记表。
[CR008, CR009, CR010, CR011, CR012, CR013]Kunlunxin 的关键依赖穿过单一代工厂、单一控制方 / 客户关系和外部开源软件生态;对任何一个依赖,公司都未披露双源或多元化策略。
SMIC 和 HBM 供应商节点标记为「warning」,因为 Kunlunxin 具体代工厂和内存采购安排来自行业报道推断,并非公司官方确认。
[CR009, CR010, CR016]7.3 合作伙伴与依赖风险
Kunlunxin 最关键的依赖风险是结构性的:Baidu 同时是其约 59.45% 的控股股东、最大客户和 IPO 发起方,治理、商业、资本市场杠杆都集中在同一个交易对手手里;仅靠增加外部客户,无法分散这种集中度。第二层依赖是 SMIC 作为唯一(据信)代工厂,本身产能受限,且据报道把 Kunlunxin 订单排在 Huawei 之后。第三层依赖来自 Kunlunxin 的开源软件生态(PaddlePaddle、基于 TVM 的 XTCL 编译器和 LLVM):这些工具让 Kunlunxin 具备可信的开发者工具链,但也把竞争位置系在其无法完全控制的外部开源项目健康度和迭代速度上。第四层依赖是大客户集中:China Mobile 集采中标(RMB1B+)和 2026 年 Tencent 关系单项规模都足够大,若其中任一关系不续约或不复现,都会实质影响 Kunlunxin 外部收入增长叙事;两项关系的期限和续约条款均未披露。最后,IPO 本身带来融资依赖:据报道,Kunlunxin 的 IPO 结构把投资者资本和芯片购买承诺绑在一起,意味着 IPO 流程中的执行风险(定价、投资者需求或监管延迟)可能同时威胁融资和其试图提前锁入的收入管线。[CR014, CR015, CR016, CR017, CR018, CR019]
| 依赖 | 交易对手 | 角色 | 集中度 | 失效情景 | 严重性 | 缓释措施 | 剩余风险敞口 |
|---|---|---|---|---|---|---|---|
| 控股股东与最大客户重叠 | Baidu, Inc. | 控制方、客户、IPO 发起方 | ~59.45% 股权,加最大收入交易对手 | 关联方定价或治理决策损害 IPO 少数股东 | 严重 | 除 HKEX 上市要求下的标准关联方交易规则外,未披露其他措施 | 高 |
| 唯一(据信)制造晶圆厂 | SMIC | 芯片制造(7nm DUV,未确认) | 如果 SMIC 归因正确,则 P800 生产 100% 由其承担 | 晶圆分配不足或良率问题限制出货量 | 严重 | 未披露;未确认双晶圆厂策略 | 高 |
| 开源软件生态依赖 | PaddlePaddle、Apache TVM、LLVM(外部开源项目) | 支撑 XTCL/XBLAS/XDNN 的编译器和框架基础设施 | 高;Kunlunxin 软件栈直接建立在这些外部项目之上 | 外部项目开发节奏放慢,可能卡住 Kunlunxin 自身软件路线图 | 中 | PaddlePaddle 是 Baidu 关联项目,Kunlunxin 参与贡献,部分缓释该风险 | 中 |
| 最大外部客户集中(China Mobile、Tencent) | China Mobile;Tencent | 已披露的最大外部商业账户 | 两个账户构成已披露外部商业证据的大部分 | 任一关系不续约或不复购,都会实质拖慢外部收入增长 | 高 | 未披露;合同期限和续约条款未知 | 高 |
| IPO 融资与芯片采购承诺结构 | 基石投资者;CICC(已确认的科创板顾问) | 融资加预置收入承诺 | IPO 募资与芯片采购收入在结构上挂钩 | IPO 延期、重新定价或投资者撤回,会同时威胁资本和收入 | 高 | 未披露 | 高 |
各行按严重性降序排列。集中度和严重性评级反映本报告的定性评估;Kunlunxin 未发布依赖风险登记表。
[CR014, CR015, CR016, CR017, CR018, CR019]Kunlunxin 的关键风险通过少数共同卡点传导到收入、融资和估值:SMIC 制造产能、Baidu 同时作为控制方和客户、以及 IPO 流程本身。
[CR009, CR014, CR017, CR018]7.4 团队、执行风险与缓释触发点
Kunlunxin 的领导层风险更像关键人和治理集中,而不是典型风险投资支持的创始团队画像。Ouyang Jian 是唯一具名的 Kunlunxin 专属高管(2021 年分拆以来任 CEO),也是公开记录中唯一拥有 Kunlunxin 运营头衔的人;其他具名领导(Robin Li、Shen Dou、Haijian He)都是 Baidu 公司高管,Kunlunxin 职责从属于其 Baidu 本职,这意味着公开记录中看不到全职、独立的高管层深度。当前项目规模抬高了执行风险:Kunlunxin 必须同时推进香港 / STAR Market 双重上市,把外部商业收入占比在一年内从约 40% 拉到超过 50%,推动 M100 芯片商业爬坡,并为 2027 年发布 M300 芯片做准备;与此同时,其总体出货份额(在中国市场约 3%)仍大幅落后于 Huawei(~20%)。本章的缓释与否决标准表,把上述风险转化为可监控触发点——被列入 Entity List、HKEX 招股书未能调和 5-6x 收入估计差、确认出现 SMIC 分配短缺、Tencent 关系未复现,或 IPO 撤回 / 重新定价——任何一项都会实质改变本报告的投资逻辑。[CR020, CR021, CR022, CR023, CR024, CR025]
| 职责 / 职能 | 依赖项或缺口 | 可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| Kunlunxin CEO(Ouyang Jian) | 唯一被点名的 Kunlunxin 专属高管;未披露继任者或副手 | 中 | 高 | 未披露 | 在 HKEX 招股书中确认,CEO 之外是否存在更完整的 Kunlunxin 专属高管团队。 |
| Baidu 公司高管(Robin Li、Shen Dou、Haijian He)兼任 Kunlunxin / Baidu 职务 | Kunlunxin 缺少独立于 Baidu 管理层的全职高管梯队 | 高 | 中 | Baidu 持续背书可提供过渡期稳定性 | 评估 Kunlunxin 在 IPO 后是否会搭建独立高管团队。 |
| 同步推进双重上市(HKEX 加科创板) | 两条监管与资本市场流程并行推进 | 中 | 中 | CICC 已参与科创板路径;HKEX 流程进度更靠前 | 跟踪两地上市是否按报道时间表推进,或其中一条被推迟 / 放弃。 |
| 外部收入规模化执行(份额目标从 40% 升至 50%+) | 需要持续赢得国企招标,并拓展与超大规模云厂商竞争的客户 | 中 | 中 | 已证明的中标案例(China Mobile、Tencent)提供了部分执行证据 | 跟踪后续披露中外部收入占比是否继续上升。 |
| 按年度芯片迭代承诺执行 M100/M300 路线图 | 五年年度迭代承诺已给出,但截至 2026 年中仅 M100 进入商业爬坡 | 中 | 中 | M100 已在爬坡;M300 仍是 2027 年目标 | 跟踪 2027 年前 M100 出货量和 M300 开发里程碑。 |
各行按严重性降序排列。可能性 / 严重性评级来自本报告基于所引领导层与路线图证据做出的定性评估。
[CR020, CR021, CR022, CR023]| 风险 | 可监控触发因素 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 美国出口管制 / Entity List 点名 | BIS 更新 Entity List 并直接点名 Kunlunxin | BIS 在美国联邦公报发布的任何正式条目点名 Kunlunxin (Beijing) Technology Co., Ltd. | 立即重新评估供应链接入、投资者胃口和投资逻辑可行性。 |
| 未调和的收入估计 | HKEX 招股书披露经审计 FY2024/FY2025 收入 | 确认收入明显落在 ~$500M-$2.9B 估计区间之外 | 用确认数值重新校准估值倍数分析。 |
| SMIC 制造产能 / 良率短缺 | 已披露出货量增速相对路线图承诺(M100 爬坡、Tianchi 集群)的表现 | 出货量明显低于 M100/Tianchi 路线图隐含节奏 | 下调对 Kunlunxin 扩规模以竞争 Huawei / Cambricon 的信心。 |
| Tencent 关系未复现 | 后续披露(媒体或招股书)说明 Tencent 订单规模 / 续单情况 | 首次 2026 年披露后 12 个月内,未确认 Tencent 订单重复或扩大 | 下调 Kunlunxin 跨竞争对手客户验证叙事的强度。 |
| IPO 撤回、延迟或大幅重新定价 | HKEX / 科创板申报状态更新 | IPO 被撤回、无限期延迟,或定价明显低于 ~$50B 路演目标 | 重新评估 Kunlunxin 的资本市场可信度及其私人市场估值的韧性。 |
| 购芯承诺结构遭监管反推 | HKEX / SFC 或 CSRC 就基石投资者购芯结构发布公开指引 | 监管要求实质性重组或撤销与购芯挂钩的基石投资者条款 | 重新评估重组后已报道外部收入承诺是否仍然稳固。 |
触发因素是本报告围绕本章已识别风险自行综合出的可监控事件;没有任何一项是公司发布的终止标准。
[CR001, CR009, CR017, CR024, CR025]Kunlunxin 最高严重度风险集中在出口管制敞口、SMIC 制造依赖,以及 Baidu 治理 / 客户重叠;人员 / 执行风险相对更温和。
可能性、严重度和缓释成熟度是本报告基于本章所引证据给出的定性序数评估;Kunlunxin 未发布可比风险矩阵。
[CR001, CR009, CR014, CR020]08估值
8.1 投资逻辑与反向逻辑
乐观情景靠四根支柱撑着:(1)Kunlunxin 位于中国增长最快的芯片细分赛道,国内 AI 加速器厂商在中国出货份额于 2025 年突破 41%,并在持续美国出口管制和「Buy China」政策支持下继续上升;(2)Kunlunxin 的 P800 芯片已经通过真实且带对抗性的市场检验——Tencent 作为 Baidu 的直接竞争对手,2026 年开始规模采购 P800 芯片,多家媒体将其视为目前最强的第三方验证;(3)外部收入占比从约 40%(2024 年)升至超过 50%(2025 年),说明非 Baidu 自用需求之外确有商业牵引;(4)Kunlunxin 自 2011 年起积累十余年 R&D 历史,且拥有 PaddlePaddle 原生软件栈,相比若干更新的国内芯片创业公司更成熟。反向逻辑至少同样强:Kunlunxin 约 $50 billion 的 IPO 目标大约是 12 个月前 Series D 估值的 16x;其要求基石投资者购买相当于股权投资额三到七倍芯片的少见安排,也让外部需求是否有机产生这一循环融资疑虑具备可信度;其 2025 年收入估计因来源不同从约 ~$500 million 到 ~$2.9 billion 不等,差距达 5-6x 且未调和,削弱了任何基于倍数估值的信心;其制造代工方(据信为 SMIC,未确认)既产能受限,又据报道从属于 Huawei 的优先分配;Baidu 同时作为约 59.45% 控股股东和最大客户,形成关联方结构,限制了 Kunlunxin 增长故事可被独立验证的程度。若要改变这一判断,需要看到:公开 HKEX 招股书调和收入估计,确认 Kunlunxin 的代工厂和良率,并有证据证明 Tencent 关系会复现或扩大,而不是停留在一笔已披露订单。[CV001, CV002, CV003, CV004, CV005, CV006]
| 论点 | 什么会改变判断 |
|---|---|
| 乐观:Kunlunxin 搭上中国增长最快的芯片细分市场(国内出货份额 41%+ 且仍在上升) | 国内份额增长慢于预期,或美国出口管制带来的需求顺风逆转。 |
| 乐观:Tencent 2026 年采购是目前最强的第三方产品验证 | 确认 Tencent 关系会复现并扩大,而不是停留在单笔订单。 |
| 乐观:外部收入占比从 ~40%(2024)升至 >50%(2025),显示真实商业牵引力 | Kunlunxin 的 HKEX 招股书披露经审计、已调和的收入数据,确认这一趋势。 |
| 反向逻辑:不到 12 个月,IPO 目标估值约为 Series D 轮的 16 倍 | 更温和的最终 IPO 定价(例如接近早前约 $13B 的科创板指引)会缓解这一担忧。 |
| 反向逻辑:与购芯挂钩的基石投资者结构引发循环融资担忧 | 监管澄清或招股书披露显示该结构不会抬高已报道需求。 |
| 反向逻辑:各来源的 2025 年收入估计相差 5-6x($500M 至 $2.9B) | HKEX 招股书披露单一、经审计的收入数字。 |
| 反向逻辑:依赖 SMIC 代工,产能受限且据报道排序低于 Huawei | 确认具体且充足的晶圆分配协议,或披露双代工厂策略。 |
| 反向逻辑:Baidu 既是控制方又是最大客户,限制独立验证 | 招股书披露关联交易条款,并由独立审计确认。 |
论点直接来自本章及前文证据;“什么会改变判断”列是本报告自己的尽调框架,不是公司发布的风险因素。
[CV001, CV002, CV003, CV004, CV005, CV006]8.2 融资背景、入场纪律与建议
Kunlunxin 的私募融资历史显示,估值从约 $1.9 billion(2021 年 4 月分拆)到 $3.0 billion(2025 年 7 月 Series D),随后在 IPO 阶段指引中快速抬升:~$14.7 billion(2026 年 5 月)、约 HK$100 billion(~$12.8-13 billion,STAR Market 指引,2026 年 5 月),以及 ~$50 billion(2026 年 6 月路演报道)。已审阅来源没有披露任何私募轮的优先股堆叠、清算优先权或稀释条款;在公开招股书发布前,IPO 最终定价、股份分配和承销团仍未披露(仅通过 GuruFocus 引用的 CSRC 文件确认,CICC 为 STAR Market 路线提供辅导 / 顾问角色)。私募市场入场纪律,首先要确认市场到底在给哪一个收入数字定价——CNBC 估计的 ~$500 million,还是 Tencent Tech/BigGo Finance 估计的 ~$2.9 billion。因为在 $50 billion 估值下,隐含市销率从约 17x 到 100x 不等;这个区间一端是「激进但仍在该赛道泡沫化常态内」(Cambricon 可比区间约 55-85x),另一端则是「按任何可比口径都难以自洽」。收入差距未解、代工厂 / 良率图景未确认、缺少独立基准测试,再加上 IPO 结构本身存在循环融资问号,本报告建议为继续研究(research-more):既不是买入(估值和收入质量证据过于不完整),也不是直接回避(China Mobile、Tencent 以及不断上升的外部收入占比说明底层商业牵引真实且有独立佐证)。风险评级为高;估值立场为偏高到昂贵;本建议置信度为中,反映的是真实不确定性,而不是证据基础薄弱。[CV008, CV009, CV010, CV011, CV012, CV013]
| 建议 | 置信度 | 风险评级 | 估值立场 | 决策含义 |
|---|---|---|---|---|
| 继续研究 | 中 | 高 | 偏高至昂贵 | 在 HKEX 招股书调和收入、确认代工 / 良率并明确购芯承诺结构的处理方式前,不应按报道的 ~$50B 目标承销。 |
这张单行表反映本报告截至 2026-08-19 运行日期的总体建议;Kunlunxin 公开招股书发布后应重新审视。
[CV008, CV009, CV011, CV012, CV013]Kunlunxin 的建议链条从真实但不完整的商业验证出发,穿过尚未解决的财务和制造证据缺口,最后指向继续研究,而不是买入或回避。
[CV008, CV009, CV011, CV013, CV002]8.3 乐观、基准、悲观情景及可比公司
乐观情景下,Kunlunxin 的 2025 年收入向较高的 ~$2.9 billion 估计靠拢,外部收入占比继续升破 50%,Tencent 关系扩大为可复现的多年合同,SMIC 产能 / 良率改善到足以支撑 M100 爬坡和 M300 按期发布,IPO 定价接近或高于 $50 billion 路演目标——这意味着市销率约 17x,对照 Cambricon 的可比区间大体说得通。基准情景下,2025-2026 年收入更接近与 J.P. Morgan 口径一致的 ~$1.2-1.8 billion,外部占比稳定在 50-60% 区间,IPO 定价也更接近此前 HK$100 billion(~$13 billion)的 STAR Market 指引,而不是抬升后的 $50 billion;本报告认为这一估值更有已审阅证据支撑。悲观情景下,收入向 CNBC 较低的 ~$500 million 估计靠拢,Tencent 订单被证明只是一次性验证采购而非持续关系,SMIC 分配短缺限制出货增长,IPO 延期、实质下调定价,或因芯片购买承诺结构受到监管审查;这些因素叠加,会让所报道的 $50 billion 目标在任何可比方法下都难以站住脚。相关可比集合包括四家已上市或新上市的中国 AI 芯片公司(Cambricon、Moore Threads、Biren Technology、Iluvatar CoreX)以及作为全球倍数基准的 Nvidia;没有一家能完美匹配 Kunlunxin 这种 Baidu 关联、仍未上市的结构,但合在一起可以约束一个可辩护的估值区间。[CV014, CV015, CV016, CV017, CV018, CV019]
| 情景 | 假设 | 估值 / 回报逻辑 | 关键风险 | 概率信号 |
|---|---|---|---|---|
| 乐观 | 2025 年收入调和后接近 ~$2.9B;外部占比继续高于 50%;Tencent 关系复现并扩大;SMIC 产能 / 良率支持 M100/M300 路线图按期推进 | $50B 目标对应隐含 P/S 约 17x,在更小收入基数上大体可比 Cambricon 的 55-85x 区间 | 需要收入、制造和客户持续性不确定性同时朝 Kunlunxin 有利方向落地 | 低至中;需要多个当前未解决缺口同时有利化解。 |
| 基准 | 2025-2026 年收入接近 J.P. Morgan 推算的 ~$1.2-1.8B 区间;外部占比稳定在 50-60%;IPO 定价更接近早前 ~$13B 科创板指引,而不是被抬高后的 $50B 数字 | 更保守定价下,可比公司倍数对估值的支撑更强 | 仍取决于收入估计缺口能否解决,以及制造产能能否确认 | 中;与本报告审阅的证据范围最一致。 |
| 悲观 | 2025 年收入调和后接近 CNBC 较低的 ~$500M 估计;Tencent 订单被证明是一次性;SMIC 分配短缺持续;IPO 延迟、下调定价,或因购芯结构受到监管审视 | 在 ~$500M 收入基数上,$50B 对应隐含 P/S 将接近 100x,无论参考哪类可比公司都难以辩护 | 循环融资担忧成为市场 / 监管焦点,广泛损害投资者信心 | 低至中;需要几个当前开放问题都朝负面方向落地。 |
概率信号是本报告基于所审阅证据权衡后的定性评估,不是量化统计模型。
[CV014, CV015, CV016, CV017]| 可比公司 | 指标 | 倍数 / 估值 / 状态 | 相关性 | 局限 |
|---|---|---|---|---|
| Cambricon(寒武纪,688256.SH) | 市值 / 往绩 PE | RMB630B-1T (~$88-140B);200-373x 往绩 PE;FY2025 收入 RMB6.5B(+453%),首次实现年度盈利 | 估值最高的已上市中国纯 AI 芯片公司;最接近的公开市场倍数参照 | 已盈利且已上市;Kunlunxin 仍处 IPO 前阶段,按调和口径尚未盈利。 |
| Moore Threads(摩尔线程,科创板) | 市值 / P/S | IPO 后 RMB282-306B(~$39-43B,2025 年 12 月);三年累计亏损约 RMB6B | 新近上市的国产 GPU 可比公司,首日交易同样带有 IPO 热炒模式 | 全栈图形加 AI 定位不同于 Kunlunxin 的纯 AI 聚焦。 |
| Biren Technology(壁仞科技,HKEX) | IPO 前估值 / IPO 后市值 | IPO 前 ~$2.2-2.9B(2025);IPO 后 ~$12.3B(2026 年 1 月,上市首日上涨 +72.86%) | HKEX 可比上市公司,收入规模(2025 H1 约 $81.8M)与 Kunlunxin 低端估计相近 | 2023 年以来被列入美国 Entity List,这一风险因素尚未在 Kunlunxin 身上确认。 |
| Iluvatar CoreX(天数智芯,HKEX) | IPO 后市值 | IPO 后 ~$5.2B(2026 年 1 月);FY2025 收入 RMB1.03B(+91.6%),毛利率约 54% | HKEX 上市国产 GPU 可比公司,已有经审计财务,可用作利润率结构代理 | ByteDance 客户敞口集中,不同于 Kunlunxin 更分散的客户基础。 |
| Nvidia(NASDAQ:NVDA) | 市值 / 收入倍数 | 全球市值 ~$3T+;收入倍数约 16-25x(2026) | AI 芯片估值倍数的全球基准,也是现状竞争标准 | 不是中国本土可比对象;规模和生态成熟度差了数个数量级。 |
估值数字混合了市值(已上市同行)、IPO 前 / 后估值(新近上市同行)和全球基准倍数(Nvidia);跨行比较只能看方向,不能视为完全可比。
[CV024, CV025, CV026, CV027]在据报道的 $50 billion IPO 目标估值下,Kunlunxin 的隐含市销率(P/S)高度取决于采用哪项 2025 年收入估计。
P/S 值按据报道的 ~$50B IPO 目标估值除以各收入估计计算;基准情景采用 $1.2B 收入估计,口径贴合 J.P. Morgan 人民币预测经汇率调整后的读数。这些数字均未得到公开招股书确认。
[CV009, CV014, CV024]2025-2026 年间,Kunlunxin 据报道的估值指引跨度很大:从最后一轮已确认的私募融资估值,到 2026 年中上调后的 IPO 路演目标。
估值变化区间采用 2025 年 7 月 Series D 轮水平(低位)、2026 年 5 月 STAR Market / HK$100B 指引(基准,~$13B),以及 2026 年 6 月路演目标(高位,~$50B);P/S 区间取自上方的收入敏感性图。
[CV010, CV011, CV014, CV016]8.4 退出准备度与最终尽调问题
现有私募投资者的退出准备度,直接取决于 IPO 本身:Kunlunxin 的保密 HKEX Form A1 申请(2026 年 1 月)和并行 STAR Market 辅导流程(2026 年 5 月),就是 Series A-D 投资者实现流动性的机制;据 BigGo Finance,Jefferies 分析师预计香港上市会在 2026 年第三季度完成。除 IPO 流程本身外,已审阅来源没有披露锁定期条款、老股交易安排或现有投资者退出机制。本章最终尽调问题把前文最关键的开放问题合并成一张优先级清单:调和 2025 年收入估计;确认制造代工厂和良率;取得独立基准测试数据;确认承销团和最终发行条款;澄清芯片购买承诺结构将在公开招股书中如何处理。每一项本身都是足以打破投资逻辑的触发点;合在一起,也解释了为什么本报告把 Kunlunxin 视为一个应在招股书提交过程中密切跟踪的标的,而不是今天就下注的标的。[CV020, CV021, CV022, CV023]
| 触发因素 | 阈值 | 传导到投资逻辑的路径 | 行动含义 |
|---|---|---|---|
| 确认美国 Entity List 直接点名 Kunlunxin | BIS 在美国联邦公报发布的任何正式条目点名 Kunlunxin (Beijing) Technology Co., Ltd. | 立即削弱代工准入、投资者胃口和 IPO 进程本身 | 将建议调整为回避;全面重新评估供应链和资本市场可行性。 |
| 收入调和结果落在或低于 CNBC 的 ~$500M 估计 | HKEX 招股书经审计收入数字位于估计区间低端或更低 | 意味着 $50B 目标对应 P/S 接近 100x,无论参考哪类可比公司都难以辩护 | 将估值立场调整为昂贵 / 缺乏依据;把建议下调至接近回避。 |
| Tencent 关系在首次披露后 12 个月内未复现 | 后续媒体或招股书披露中没有确认重复或扩大订单 | 削弱目前最强的第三方客户验证信号 | 下调对外部需求增长叙事的信心。 |
| IPO 被撤回、无限期延迟,或定价明显低于 ~$50B 目标 | HKEX / 科创板正式申报状态变化,或确认定价低于路演指引 | 表明资本市场并不支持已报道估值 | 重新评估 Kunlunxin 的私人市场估值和此前路演指引的可信度。 |
| SMIC 晶圆分配或良率短缺得到确认 | 已披露出货量增长明显低于 M100/Tianchi 路线图节奏 | 限制 Kunlunxin 扩规模并与 Huawei / Cambricon 竞争的能力 | 下调对 Kunlunxin 缩小与 Huawei 出货份额差距能力的信心。 |
触发因素是本报告围绕全报告识别的风险和投资逻辑要素自行综合出的可监控事件;没有任何一项是公司发布的终止标准。
[CV009, CV020, CV021, CV022]| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 |
|---|---|---|---|
| 已调和 2025/2026 收入 | 经审计收入数字,调和 ~$500M-$2.9B 估计区间 | 锚定本章所有估值倍数计算 | HKEX 招股书公开后核对;过渡期直接向公司确认。 |
| 制造代工厂与良率 | 官方确认 P800 代工厂(SMIC 还是仍有 TSMC 敞口)和良率 | 决定成本结构、供应可靠性和出口管制敞口 | HKEX 招股书;直接向公司确认。 |
| 独立、可复现的基准测试数据 | 用标准化基准测试套件,将 Kunlunxin 芯片与 Huawei、Cambricon 和 Nvidia 对比 | 验证当前仅靠单一分析师估计支撑的竞争定位说法 | 拿到硬件后委托独立基准测试。 |
| IPO 承销团与最终发行条款 | 完整账簿管理人名单、最终股份分配和定价条款 | 反映交易可信度和定价纪律 | 招股书公开后查看 HKEX / CSRC 申报文件。 |
| 购芯承诺结构的信息披露处理 | 监管指引或招股书语言,说明如何披露基石投资者购芯条款 | 决定循环融资担忧在上市后是否得到处理,还是继续存在 | HKEX / SFC 和 CSRC 公开指引;最终招股书。 |
| 员工数以及债务 / 项目融资义务 | 这一可比公司组中所有已上市同行都会披露的基本规模和资产负债表指标 | 支持与 Cambricon、Moore Threads、Biren、Iluvatar CoreX 做标准规模和资本充足性比较 | HKEX 招股书公开后核对。 |
主题排序是本报告按各缺口对可靠估值的约束程度做出的非正式判断,不是公司披露的优先级顺序。
[CV020, CV021, CV022, CV023]Kunlunxin 一边有强劲收入增长信号和真实客户验证,一边仍留下财务、制造和结构层面的证据缺口。
[CV008, CV009, CV011, CV013]免责声明
本报告基于截至 2026-08-19 可获得的公开来源,综合研究与尽调信息而成,不构成投资建议。截至报告生成日,Kunlunxin 仍是未上市公司;文中描述其 2025-2026 IPO 进程的数字来自媒体报道和路演指引,并非公开招股书,官方披露前应视为临时性信息。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Kunlunxin (Beijing) Technology Co., Ltd. is the semiconductor design subsidiary of Baidu, Inc., building AI training and inference chips under the Kunlun brand. | 高 | SO001, SO004 |
| CO002 | Baidu began internal chip R&D in 2011 through its Intelligent Chip and Architecture Department, a decade before Kunlunxin's formal spin-off. | 中 | SO001, SO013 |
| CO003 | In April 2021, Baidu spun off its internal chip team into the independent company Kunlunxin, which closed its first outside funding round the same year. | 高 | SO001, SO013, SO027 |
| CO004 | Kunlunxin's first chip generation (Kunlun I / K200) was unveiled in July 2018, manufactured on Samsung's 14nm process with 16GB of HBM2 memory at 512GB/s bandwidth. | 高 | SO026, SO030 |
| CO005 | Kunlun II, Kunlunxin's second-generation chip, entered mass production in August 2021 on a TSMC 7nm process and was publicly positioned as a rival to Nvidia's A100 for AI workloads. | 高 | SO012, SO001 |
| CO006 | Kunlunxin's third-generation P800 (Kunlun III) chip, launched in 2025 for large-cluster AI training and inference, gained full DeepSeek R1/V3 compatibility in February 2025. | 中 | SO008 |
| CO007 | In April 2025, Baidu activated a 30,000-chip P800 cluster at its developer conference, describing it as capable of training "DeepSeek-like" models with hundreds of billions of parameters. | 高 | SO001, SO028 |
| CO008 | Baidu announced in November 2025 that it would release the inference-optimized Kunlun M100 chip in 2026 and the training-focused Kunlun M300 chip in 2027. | 高 | SO010, SO011 |
| CO009 | Kunlunxin's official website describes the entity using the Chinese "股份有限公司" (joint-stock company) designation, a corporate form Chinese companies typically adopt ahead of a public listing. | 中 | SO007 |
| CO010 | The Kunlun M100 chip entered commercial ramp-up in January 2026, with its physical product first shown publicly in a July 2026 CCTV state-television news report. | 中 | SO009 |
| CO011 | Ouyang Jian, Baidu's chief chip architect, became CEO of Kunlunxin when the unit was spun off in April 2021. | 高 | SO001, SO027 |
| CO012 | Baidu co-founder and CEO Robin Li has publicly framed the Kunlun chip program as a 14-year strategic bet and personally confirmed the M100 chip's 2026 release timeline in a January 2026 interview. | 中 | SO013, SO009 |
| CO013 | Shen Dou, an executive vice-president of Baidu and president of its AI Cloud Group, is Kunlunxin's designated public spokesperson for the chip roadmap, including the M100/M300 and Tianchi256/512 cluster announcements. | 高 | SO010, SO011 |
| CO014 | Baidu CFO Haijian He publicly linked the Kunlunxin spin-off and listing to unlocking shareholder value in Baidu's Q4 2025 earnings release. | 高 | SO005, SO013 |
| CO015 | Baidu, Inc. is Kunlunxin's controlling shareholder, holding approximately a 59.45% equity stake as of Kunlunxin's January 2026 confidential HKEX listing filing. | 高 | SO013, SO010 |
| CO016 | Baidu's ownership stake in Kunlunxin fell from approximately 70.87% in a 2023 report to about 59.45% by early 2026, reflecting dilution from successive external funding rounds while Baidu retained control. | 中 | SO027, SO013 |
| CO017 | As of its January 2026 HKEX filing, Kunlunxin's shareholder register had expanded to 44 distinct entities. | 中 | SO013 |
| CO018 | Kunlunxin's leadership combines Baidu's senior corporate executives (Robin Li, Shen Dou, Haijian He) with dedicated chip-operating leadership (Ouyang Jian), creating a key-person dependence anchored in continued Baidu management sponsorship rather than an independent founder team. | 中 | SO001, SO013 |
| CO019 | Kunlunxin's April 2021 spin-off financing round, led by CPE with IDG Capital, Legend Capital, and Yuanhe Puhua participating, valued the company at approximately RMB13 billion (~$1.9 billion). | 中 | SO013 |
| CO020 | Kunlunxin completed a Series D funding round in July 2025, raising approximately $283 million (~RMB2.1 billion). | 中 | SO013, SO017 |
| CO021 | Kunlunxin's Series D round valued the company at approximately RMB21 billion (~$3.0 billion) post-money. | 高 | SO013, SO015, SO029 |
| CO022 | Series D investors in Kunlunxin included BYD, CICC Guanbo, China Mobile's Zhongyi Hechuang investment vehicle, and a state-backed High-Level Talent Fund under China Reform Holdings. | 中 | SO013 |
| CO023 | Kunlunxin has maintained a financing pace of roughly one funding round per year since its 2021 spin-off. | 中 | SO013 |
| CO024 | On January 1, 2026, Kunlunxin confidentially submitted a Form A1 listing application to the Hong Kong Stock Exchange main board. | 高 | SO002, SO004, SO013 |
| CO025 | Baidu's Hong Kong-listed shares rose as much as 9.35% following the January 2026 disclosure of Kunlunxin's confidential HKEX filing. | 中 | SO013 |
| CO026 | In May 2026, Kunlunxin launched a parallel Shanghai STAR Market IPO "tutoring" (辅导) process, confirming a dual-listing strategy alongside its Hong Kong plans. | 高 | SO016, SO029 |
| CO027 | Press reports in May 2026 placed Kunlunxin's Hong Kong IPO valuation guidance near HK$100 billion (~$13 billion). | 中 | SO016 |
| CO028 | By June 29, 2026, investor roadshow reporting placed Kunlunxin's Hong Kong IPO target valuation near $50 billion, with Baidu shares jumping roughly 7% on the news. | 高 | SO003, SO019 |
| CO029 | An intermediate valuation figure of approximately $14.7 billion for Kunlunxin's IPO was reported in the weeks before the escalation to the $50 billion figure. | 中 | SO029 |
| CO030 | At its ~$50 billion IPO target, Kunlunxin's implied valuation of roughly 1.32x exceeded the approximately $38 billion market capitalization of its own controlling parent, Baidu, on the same trading day. | 中 | SO024 |
| CO031 | CNBC reported that Kunlunxin's 2024 revenue was approximately RMB2 billion with a net loss of about RMB200 million, and that 2025 revenue was expected to exceed RMB3.5 billion (~$500 million) with breakeven achieved. | 中 | SO003 |
| CO032 | A Tencent Tech-sourced figure carried by BigGo Finance put Kunlunxin's 2024 revenue at more than RMB10 billion (~$1.4 billion), with external customers at about 40%, and 2025 revenue "far exceeding" RMB20 billion (~$2.9 billion). | 中 | SO013 |
| CO033 | J.P. Morgan projected in a November 2025 research note that Kunlunxin's chip-related revenue could grow more than sixfold, from about RMB13 billion in 2025 to RMB83 billion in 2026. | 中 | SO013 |
| CO034 | Kunlunxin's most detailed financial metrics remain undisclosed pending a formal HKEX prospectus, leaving 2025 revenue, gross margin, and headcount figures as press estimates rather than audited numbers. | 中 | SO002, SO013 |
| CO035 | Kunlunxin's AI chip shipment volume reached approximately 69,000 units in 2024, according to IDC data cited in Baidu's own encyclopedia entry. | 中 | SO009 |
| CO036 | No public source reviewed for this chapter discloses a specific Kunlunxin employee headcount figure as of August 2026. | 低 | |
| CO037 | In August 2025, Kunlunxin's AI servers won a centralized procurement bid from China Mobile worth more than RMB1 billion (~$139 million) across all three general-computing bid packages. | 中 | SO001, SO013 |
| CO038 | Named external customers of Kunlunxin chips as of late 2025 include China Merchants Bank, China Southern Power Grid, Geely Automobile Holdings, and Vivo, alongside Baidu's own internal deployment. | 高 | SO010, SO013 |
| CO039 | BYD is both a Series D equity investor in Kunlunxin and, per a 2023 DigiTimes report, an earlier small investor tied to autonomous-vehicle chip planning. | 中 | SO027, SO013 |
| CO040 | External (non-Baidu) customers accounted for roughly 40% of Kunlunxin's revenue in 2024, a share press reports say rose to more than 50% in 2025. | 中 | SO013 |
| CO041 | Tencent, a direct rival to Baidu in cloud and AI services, began procuring Kunlunxin P800 chips at scale in 2026, which multiple outlets treat as external market validation beyond Baidu's own ecosystem. | 中 | SO021 |
| CO042 | A single 8-card Kunlunxin P800 server can run the full 671-billion-parameter DeepSeek R1 model at 8-bit precision, according to Kunlunxin's own technical announcement. | 中 | SO008 |
| CO043 | Kunlunxin's Hong Kong IPO structure reportedly asks prospective cornerstone investors to commit to purchasing chips worth three to seven times their equity investment. | 高 | SO021, SO019, SO003 |
| CO044 | Analysts have flagged the chip-purchase-linked IPO structure as raising circular-financing concerns about whether it reflects genuine external demand or artificially inflated near-term revenue commitments. | 中 | SO021, SO022 |
| CO045 | Depending on which 2025 revenue estimate is used (~$500 million to ~$2.9 billion), Kunlunxin's implied price-to-sales multiple at a $50 billion valuation ranges from roughly 17x to 100x, above the 5-20x typical for high-growth semiconductor companies. | 中 | SO022, SO024 |
| CO046 | Kunlunxin is one of several Chinese chip designers -- alongside Cambricon, Moore Threads, Biren Technology, and Iluvatar CoreX -- that completed or advanced a public listing process in the 2025-2026 window amid a surge in China domestic AI chip IPOs. | 高 | SO013, SO015 |
| CO047 | Baidu's Q3 2025 results showed a net loss of RMB11.2 billion, driven partly by a RMB16.2 billion one-time impairment charge, financial pressure that helped motivate Baidu's push to unlock Kunlunxin's standalone value through the spin-off listing. | 中 | SO013 |
| CO048 | Baidu retains SEC reporting obligations as a NASDAQ-listed ADR issuer (ticker BIDU) in addition to its Hong Kong primary listing (9888.HK), giving Kunlunxin's ultimate parent a dual-regulator disclosure regime. | 高 | SO025, SO005 |
| CM001 | This chapter defines the addressable market as China's domestic AI accelerator (training and inference chip) market, excluding non-AI compute spend and AI chip demand outside China. | 中 | SM002, SM019 |
| CM002 | HBM memory supply and advanced-node foundry capacity are adjacent constraint markets rather than buyer-facing accelerator demand. | 中 | SM011, SM012 |
| CM003 | The status-quo substitute for a Chinese AI buyer is importing or renting Nvidia/AMD compute directly or via offshore cloud arrangements, rather than adopting a domestic accelerator. | 中 | SM013, SM014 |
| CM004 | IDC data reviewed by Reuters shows total China AI accelerator card shipments of roughly 4 million units in 2025, split 55% Nvidia, 4% AMD, and 41% domestic Chinese vendors. | 高 | SM002, SM003 |
| CM005 | A second August 2026 report found many Chinese AI developers still accessing Nvidia compute via offshore cloud rental in Southeast Asia specifically to avoid domestic-chip migration costs. | 中 | SM013 |
| CM006 | Bernstein modeling puts China's AI accelerator market at $37 billion in 2025, growing to $55 billion in 2026 and $69 billion in 2027. | 中 | SM001 |
| CM007 | The same Bernstein forecast projects China's domestic AI chip production rising from about $16 billion (2025) to about $91 billion by 2028, against roughly $88 billion of total national demand at that point. | 低 | SM001 |
| CM008 | Huawei shipped approximately 812,000 AI accelerator cards in China in 2025, roughly half of all domestically-branded shipments and about 20% of the total China market. | 高 | SM002, SM004 |
| CM009 | Alibaba's T-Head chip unit shipped approximately 265,000 AI accelerator cards in China in 2025, ranking second among domestic vendors. | 高 | SM002, SM003 |
| CM010 | Kunlunxin and Cambricon each shipped approximately 116,000 AI accelerator cards in China in 2025, tying for third place among domestic vendors and representing roughly 3% each of the total China market. | 高 | SM002, SM003 |
| CM011 | TrendForce expects domestic chips to exceed 50% of Chinese AI server deployments by the end of 2026. | 中 | SM019 |
| CM012 | Hygon, MetaX, and Iluvatar CoreX together accounted for the remaining roughly 12% of domestic-vendor shipments in China in 2025 (5%, 4%, and 3% respectively of total domestic shipments). | 中 | SM002 |
| CM013 | ByteDance reportedly spends about $12.1 billion on AI chips in 2025, rising toward $14.3 billion in 2026, deliberately split across Huawei, Cambricon, and Iluvatar CoreX (with Kunlunxin under evaluation) to hedge supply risk and optimize cost per workload. | 中 | SM009, SM010 |
| CM014 | Hyperscalers (Baidu, Alibaba, Tencent, ByteDance) are the largest and most sophisticated domestic AI chip buyers, typically running deliberate multi-vendor procurement strategies. | 中 | SM009, SM020 |
| CM015 | State-owned and government-linked buyers such as China Mobile procure domestic AI chips through centralized tender processes shaped by Xinchuang (信创) domestic-substitution policy requirements. | 中 | SM009, SM004 |
| CM016 | Automotive OEMs including BYD and Geely form a smaller buyer segment purchasing AI compute for ADAS and in-vehicle autonomous-driving applications. | 中 | SM009 |
| CM017 | National "East Data West Computing" (东数西算) infrastructure hubs represent a subsidy-heavy, sovereign-compute buyer channel distinct from hyperscaler or SOE procurement. | 中 | SM015, SM016 |
| CM018 | Financial-sector institutions beyond state-owned enterprises are evaluating domestic AI chips for data-sovereignty and vendor-diversification reasons. | 低 | SM009 |
| CM019 | US export controls on advanced Nvidia and AMD chips, including the 2025 H20 ban and its partial reversal with a 15% US Treasury fee, have structurally shifted addressable China AI chip demand toward domestic vendors. | 高 | SM007, SM008, SM018 |
| CM020 | China's "Buy China" procurement preference and Xinchuang domestic-substitution mandates create policy-driven demand floors for domestic AI chips among SOE and government buyers. | 中 | SM004, SM009 |
| CM021 | China's National Semiconductor Fund ("Big Fund III"), launched in 2024 with roughly $47.5 billion in capital, subsidizes domestic chip manufacturing to partially offset SMIC's cost and yield disadvantage versus TSMC. | 高 | SM015, SM016 |
| CM022 | SMIC's most advanced available domestic node (7nm, DUV-based) runs at an estimated 25-50% higher cost and meaningfully lower yield than TSMC's EUV equivalent, and capacity is shared across Huawei, Cambricon, Biren, and other national-champion customers. | 中 | SM011, SM012 |
| CM023 | The Council on Foreign Relations assesses that the best US AI chips remained roughly 5x more powerful than China's best in 2025, a gap it expects could widen to as much as 17x by 2027. | 中 | SM017 |
| CM024 | Bernstein's forecast that China becomes a net AI-chip exporter by 2028 is not reconciled with the Council on Foreign Relations' assessment that the US-China compute-performance gap is widening rather than closing over a similar period. | 中 | SM001, SM017 |
| CM025 | The CUDA software ecosystem still requires meaningfully more migration engineering for domestic chips than Nvidia hardware, even where domestic chip specifications are competitive on paper, as documented in reporting as recent as August 2026. | 中 | SM013, SM014 |
| CM026 | No source reviewed for this chapter isolates Kunlunxin's addressable market independent of Baidu's own internal AI infrastructure demand. | 低 | |
| CM027 | Morgan Stanley's cited estimate of roughly 1,521 tokens/second for Kunlunxin's P800 chip on DeepSeek R1 inference, positioning it between Nvidia's H20 and A100, is a single third-party estimate rather than a reproducible, independently verified benchmark suite. | 低 | SM026 |
| CM028 | China reportedly banned foreign AI chips from all state-funded data centers, per Reuters reporting cited by The AI Track, reshaping domestic AI supply chains in favor of vendors such as Huawei. | 中 | SM024 |
| CM029 | The US restricted TSMC's advanced AI chip exports to China specifically targeting Huawei, further tightening control over semiconductor technology access. | 中 | SM024 |
| CM030 | Huawei doubled its AI chip yield rate at SMIC to nearly 40% in 2025, up from about 20% a year earlier, easing a key domestic production bottleneck. | 中 | SM025 |
| CM031 | SMIC reportedly prioritizes Huawei's production allocation over other domestic chip customers such as Cambricon and Kunlunxin, giving Huawei a structural manufacturing advantage that smaller vendors cannot easily replicate. | 中 | SM025 |
| CM032 | Huawei's Ascend 910C chips power approximately 50% of domestic AI data centers in China as of the 2025 shipment data, reflecting deliberate government-backed deployment rather than organic commercial adoption alone. | 中 | SM025 |
| CM033 | Bernstein analysts estimated in January 2026 that Huawei and Nvidia held roughly tied shares of about 40% each in China's AI chip market, an estimate that IDC's later full-year 2025 data showed domestic vendors had already collectively surpassed. | 中 | SM025, SM001 |
| CM034 | Bernstein analyst projections cited by WinBuzzer suggest Nvidia's China AI accelerator market share could fall to as low as 8% under sustained export restrictions, down from 55% in 2025. | 低 | SM025 |
| CM035 | Huawei's Ascend 910C chip delivers approximately 60% of Nvidia H100's inference performance, a gap considered acceptable for most commercial inference workloads even though it matters more for advanced training tasks. | 中 | SM025, SM023 |
| CP001 | Huawei's Ascend chip line (910B/910C, and the newer Atlas 350) is the dominant domestic incumbent, backed by Huawei's own CANN software stack and reported SMIC production-priority status. | 高 | SP020, SP017 |
| CP002 | Cambricon and Moore Threads are the two domestic AI chip peers that had already gone public by mid-2026, each reaching valuations far above Kunlunxin's private mark. | 高 | SP001, SP013, SP004 |
| CP003 | Biren Technology and Iluvatar CoreX both completed Hong Kong listings in the same December 2025-January 2026 window that Kunlunxin is targeting for its own Hong Kong IPO. | 高 | SP007, SP009, SP008 |
| CP004 | Enflame Technology and Alibaba's in-house chip unit T-Head (Pingtouge) remain in pre-IPO or internal-only stages, distinct from the vendors that have already completed a public listing. | 中 | SP002, SP022 |
| CP005 | The status-quo alternative for a Chinese AI buyer remains sourcing Nvidia (and to a lesser extent AMD) chips where export licenses or offshore cloud rental permit it, or building AI chips fully in-house. | 中 | SP024 |
| CP006 | At least six Chinese AI chip companies either listed or advanced a public listing process between December 2025 and June 2026, and the combined market value of the domestic GPU sector approached RMB 2 trillion by mid-2026. | 中 | SP002 |
| CP007 | Huawei's Ascend 910C delivers approximately 60-80% of Nvidia H100's inference performance according to independent technical comparisons. | 中 | SP018, SP019 |
| CP008 | Huawei's Ascend Computing Business president claimed the newer Atlas 350 chip reaches 2.87x the performance of Nvidia's China-specific H20 chip, a company claim not independently verified in sources reviewed. | 低 | SP002 |
| CP009 | Morgan Stanley independently estimated Kunlunxin's P800 chip at roughly 1,521 tokens/second on DeepSeek R1 inference, positioning it between Nvidia's H20 and A100. | 中 | SP027 |
| CP010 | Every major domestic AI chip vendor has built or is building its own CUDA-alternative software stack: Huawei's CANN, Moore Threads' MUSA (with a Musify code-migration tool), and Kunlunxin's XTCL/XBLAS/XDNN. | 高 | SP004, SP027 |
| CP011 | Alibaba's T-Head PPU chip claims a bill-of-materials cost roughly 40% below Nvidia's H20, with Alibaba Cloud reportedly planning a roughly 50% inference price cut using domestic chips. | 中 | SP021 |
| CP012 | Moore Threads' consumer-grade MTT S80 GPU is positioned roughly on par with Nvidia's RTX 3060, giving it the only public consumer-price reference point among domestic AI chip vendors. | 中 | SP004 |
| CP013 | Moore Threads and Biren Technology both operate fabless, full-stack business models (hardware plus software plus systems), while Kunlunxin's commercial motion still leans heavily on Baidu's own cloud and enterprise sales relationships. | 中 | SP004, SP006 |
| CP014 | Kunlunxin's P800 cluster trained Baidu's Wenxin 4.5 Turbo model in April 2025 and later the Wenxin 5.1 model on an all-domestic chip stack. | 中 | SP028 |
| CP015 | Tencent, a direct Baidu rival, began sourcing Kunlunxin P800 chips in 2026, a cross-rival procurement event multiple outlets treat as unusually strong third-party validation, comparable to Apple sourcing OLED panels from rival Samsung. | 中 | SP028 |
| CP016 | On DeepSeek-V4's April 2026 release day, Huawei Ascend, Cambricon, Hygon, Moore Threads, Alibaba's T-Head, and Kunlunxin all completed same-day model adaptation, indicating DeepSeek compatibility has become table stakes across the sector rather than a durable Kunlunxin-specific differentiator. | 高 | SP002, SP020 |
| CP017 | SMIC reportedly prioritizes Huawei's wafer allocation over other domestic chip customers including Cambricon and Kunlunxin, giving Huawei a structural manufacturing advantage smaller vendors cannot easily replicate. | 中 | SP020 |
| CP018 | Huawei has roughly doubled its AI chip yield rate at SMIC to nearly 40% in 2025, up from about 20% a year earlier, a production improvement not reported for any other domestic vendor including Kunlunxin. | 中 | SP020 |
| CP019 | Biren Technology has operated under a US Entity List designation since October 2023, illustrating precedent for a Chinese AI chipmaker facing direct US export restriction as its commercial and technical profile grows. | 高 | SP006, SP008 |
| CP020 | No source reviewed for this chapter confirms that Kunlunxin itself currently faces a direct US Entity List or comparable export-control designation. | 低 | |
| CP021 | Kunlunxin's approximately 116,000-unit 2025 shipment volume (tied with Cambricon) is a distant third behind Huawei's roughly 812,000 units and Alibaba T-Head's roughly 265,000 units, making Kunlunxin's scale-based competitive position the weakest among the top domestic vendors by unit volume. | 中 | SP019, SP020 |
| CP022 | Moore Threads has released four generations of GPU architecture (Sudi, Chunxiao, Quyuan, Pinghu), with the latest Pinghu architecture delivering 32 TFLOPS of FP32 vector compute, above Nvidia's A100 but below the H100. | 中 | SP004 |
| CP023 | Moore Threads won a roughly RMB660 million cluster order in March 2026 for its Kua'e computing cluster, equivalent to about 55% of the company's full-year 2024 revenue, demonstrating it has moved from chip manufacturing into large-scale cluster delivery. | 中 | SP002 |
| CP024 | Nvidia's H20 chip has been subject to intermittent export licensing in China, including a reported 15% US Treasury fee on licensed sales, making its availability politically contingent rather than a stable supply line for Chinese buyers. | 中 | SP024 |
| CP025 | No vendor covered in this chapter, including Kunlunxin, publishes standalone unit list pricing for its AI accelerator chips; all public pricing signals are proxies such as bill-of-materials claims or consumer-GPU comparisons. | 中 | SP021, SP004 |
| CP026 | Cambricon's market capitalization reached roughly RMB1 trillion (~$140 billion) in June 2026, up from an RMB630 billion ($88 billion) valuation on the January 2026 Hurun China AI 50 ranking, making it the highest-valued pure-play Chinese AI chip company. | 中 | SP013, SP015, SP014 |
| CP027 | Cambricon's FY2025 revenue reached RMB6.5 billion (+453% year over year), and the company recorded its first-ever annual net profit, with cloud AI chip revenue accounting for more than 99% of total revenue. | 中 | SP012 |
| CP028 | Moore Threads debuted on Shanghai's STAR Market on December 5, 2025, with shares surging 425-469% on the first trading day and reaching a market capitalization of roughly RMB282-306 billion (~$39-43 billion). | 高 | SP004, SP005 |
| CP029 | Moore Threads' founder and controlling shareholder, Zhang Jianzhong, spent 15 years as an Nvidia global vice president before founding the company. | 中 | SP004 |
| CP030 | Biren Technology's BR100 chip, launched in August 2022, was the first Chinese-designed GPU to adopt Chiplet packaging, PCIe 5.0, and CXL interconnect technology, and its BR106 series (March 2025) supports full-stack DeepSeek-V3 training and inference. | 中 | SP006 |
| CP031 | Iluvatar CoreX's Hong Kong IPO raised approximately $474 million, with shares surging more than 30% on debut, giving the company a post-IPO market capitalization of roughly $5.23 billion. | 中 | SP009 |
| CP032 | ByteDance ordered more than 50,000 Iluvatar CoreX units for chatbot inference workloads, representing a significant customer concentration for Iluvatar CoreX compared with Kunlunxin's more diversified customer list (China Mobile, Tencent, China Merchants Bank, and others). | 中 | SP025, SP028 |
| CP033 | Enflame Technology passed its STAR Market listing review in mid-2026, planning to raise roughly RMB6 billion with proceeds split across an AI software-hardware co-design project and fifth/sixth-generation chip R&D. | 中 | SP002 |
| CP034 | Enflame Technology's pre-IPO valuation reached approximately $2.8 billion in a December 2024 Series E round, a level broadly comparable to Kunlunxin's own private valuation trajectory before its 2026 IPO escalation. | 低 | SP011 |
| CP035 | Alibaba's chip unit T-Head (also referred to as Pingtouge) initiated independent IPO proceedings in 2026, placing it on a similar spin-off timeline to Kunlunxin as Alibaba seeks to capture standalone valuation for its captive chip business. | 中 | SP022, SP028 |
| CP036 | Alibaba's T-Head PPU chip specifications include 96GB of HBM2e memory and 700GB/s of interconnect bandwidth, with hundreds of thousands of cumulative chip shipments across Alibaba Cloud's own infrastructure. | 中 | SP021 |
| CP037 | The Council on Foreign Relations assesses that Huawei (and by extension other domestic Chinese chipmakers) cannot close the absolute compute-performance gap with Nvidia in the near term, arguing US export controls should remain in place. | 中 | SP017 |
| CP038 | Despite rising domestic AI chip revenue and market share, the absolute compute-performance gap between Chinese and Nvidia chips is widening rather than narrowing, according to AInvest's June 2026 analysis. | 中 | SP024 |
| CP039 | No source reviewed for this chapter provides a confirmed, independently reproducible benchmark comparing Kunlunxin's P800 chip directly against Huawei's Ascend 910C or Atlas 350 on the same workload. | 低 | |
| CP040 | ByteDance's roughly $12.1 billion 2025 AI chip budget is deliberately split across Huawei (heavy training), Cambricon (mid-range inference), and Iluvatar CoreX (chatbot inference), with Kunlunxin reported as under evaluation rather than a confirmed primary supplier for ByteDance as of the sources reviewed. | 中 | SP025 |
| CI001 | Kunlunxin's core revenue mechanism is direct chip and cluster sales, sold both to Baidu's own internal AI infrastructure and to external enterprise/government customers via centralized tenders or direct contracts. | 中 | SI020, SI002 |
| CI002 | CNBC reported Kunlunxin's 2025 revenue was expected to exceed RMB3.5 billion (~$500 million), with breakeven achieved. | 中 | SI001 |
| CI003 | A Tencent Tech-sourced figure carried by BigGo Finance put Kunlunxin's 2025 revenue at "far exceeding" RMB20 billion (~$2.9 billion), roughly 5-6x the CNBC figure. | 中 | SI002 |
| CI004 | J.P. Morgan estimated in a November 2025 research note that Kunlunxin's chip-related revenue could grow more than sixfold, from about RMB13 billion in 2025 to RMB83 billion in 2026. | 中 | SI002 |
| CI005 | External (non-Baidu) customers accounted for roughly 40% of Kunlunxin's revenue in 2024, rising to more than 50% in 2025 according to press estimates. | 中 | SI002 |
| CI006 | Kunlunxin's go-to-market still runs substantially through Baidu's own cloud and enterprise sales relationships rather than an independently built channel, consistent with its origin as an internal Baidu chip program before its 2021 spin-off. | 中 | SI020 |
| CI007 | No source reviewed discloses Kunlunxin's gross margin, R&D spend, or manufacturing-cost breakdown. | 低 | |
| CI008 | Iluvatar CoreX reported audited FY2025 revenue of RMB1,033.6 million (+91.6% year over year), gross profit of RMB558.0 million, and an adjusted net loss of RMB437.7 million in its HKEX annual results announcement. | 高 | SI005, SI015 |
| CI009 | Biren Technology reported audited FY2025 revenue of RMB1,034.6 million (+207.2% year over year), a 53.8% gross margin, RMB1,476.1 million of R&D expense, and a statutory net loss of RMB16,493.0 million (RMB873.8 million on an adjusted, non-IFRS basis) in its HKEX annual results announcement. | 高 | SI004, SI013 |
| CI010 | Biren Technology's own prospectus disclosed a gross margin that declined from 100% in 2022 to 31.9% in the first half of 2025 as the company transitioned from early-stage sales to commercial-scale deployment. | 中 | SI013 |
| CI011 | Kunlunxin's P800 (Kunlun III) chip is widely believed to be produced on SMIC's 7nm DUV-based process given post-2021 export-control restrictions on TSMC access, but no source reviewed officially confirms Kunlunxin's specific foundry. | 中 | SI003, SI012 |
| CI012 | SMIC's 7nm node is estimated to run at 25-50% higher cost and meaningfully lower yield than TSMC's EUV equivalent, a structural cost headwind shared by every SMIC-dependent domestic chip vendor. | 中 | SI003, SI012 |
| CI013 | No public price list exists for any Kunlunxin chip; the closest sector cost-positioning reference is Alibaba's own claim that its competing T-Head PPU chip carries a bill-of-materials cost roughly 40% below Nvidia's H20. | 低 | SI002 |
| CI014 | Kunlunxin's April 2021 spin-off round valued the company at roughly $1.9 billion, and its July 2025 Series D round raised about $283 million at a roughly $3.0 billion post-money valuation. | 中 | SI002, SI024 |
| CI015 | Kunlunxin confidentially filed a Form A1 HKEX listing application on January 1, 2026, and launched a parallel Shanghai STAR Market "tutoring" process in May 2026. | 高 | SI020, SI015 |
| CI016 | Reported Hong Kong IPO valuation guidance for Kunlunxin escalated from about $14.7 billion in May 2026 to roughly $50 billion by late June 2026. | 高 | SI006, SI001 |
| CI017 | Kunlunxin's Hong Kong IPO reportedly requires cornerstone investors to commit to purchasing chips worth three to seven times their equity investment, pre-loading a multi-year order book ahead of pricing. | 高 | SI010, SI001 |
| CI018 | Analysts have flagged Kunlunxin's chip-purchase-linked IPO structure as a potential circular-financing concern about whether reported external demand reflects the structure itself rather than organic pull. | 中 | SI010, SI011, SI032 |
| CI019 | Kunlunxin's chip unit is working with China International Capital Corp. (CICC) on its STAR Market IPO process, according to a CICC filing to the China Securities Regulatory Commission cited by GuruFocus. | 中 | SI006 |
| CI020 | China's roughly $47.5 billion "Big Fund III" state semiconductor financing vehicle, launched in 2024, is intended to support domestic chip and equipment makers including SMIC, though no source reviewed quantifies how much of that capital reaches Kunlunxin specifically. | 中 | SI007, SI008 |
| CI021 | Depending on which 2025 revenue estimate is used, Kunlunxin's implied price-to-sales ratio at a ~$50 billion IPO target ranges from roughly 17x to 100x. | 中 | SI011, SI018 |
| CI022 | Cambricon, the highest-valued listed domestic AI chipmaker, traded at roughly 55-85x the range analysts have used as a comparison point when assessing whether Kunlunxin's IPO multiple is justified. | 中 | SI016, SI027 |
| CI023 | No source reviewed for this chapter discloses Kunlunxin's cash position, monthly burn rate, runway, or any debt/project-finance obligations. | 低 | |
| CI024 | No source reviewed for this chapter confirms Kunlunxin's employee headcount, in contrast to listed peers (Cambricon, Moore Threads, Biren Technology, Iluvatar CoreX) that disclose headcount in regulatory filings. | 低 | |
| CI025 | Baidu's Q3 2025 results showed a net loss of RMB11.2 billion, driven partly by a RMB16.2 billion one-time impairment charge -- financial pressure that helps explain Baidu's motivation to unlock Kunlunxin's standalone value through the spin-off and IPO. | 中 | SI002, SI009 |
| CI026 | Baidu's own Q4/FY2025 earnings release explicitly named the Kunlunxin spin-off and separate listing as progressing and expected to unlock shareholder value, directly linking Kunlunxin's financial narrative to Baidu's own consolidated results commentary. | 高 | SI009, SI020 |
| CI027 | Chinese AI chip stocks broadly traded at 250-373x trailing price-to-earnings multiples as of mid-2026, versus Nvidia's 16-25x revenue multiple, illustrating the extreme valuation premium the sector commands relative to global semiconductor peers. | 中 | SI017 |
| CI028 | Sina Finance reported in December 2025 that Kunlunxin's 2026 revenue was expected to reach RMB8 billion (~$1.1 billion), a figure between the CNBC and Tencent Tech/BigGo Finance 2025 estimates when annualized growth is considered. | 中 | SI014 |
| CI029 | Hong Kong Economic Times reported that Kunlunxin's 2024 net loss was approximately RMB200 million on revenue of approximately RMB2 billion, broadly consistent with the lower-bound CNBC revenue estimate rather than the higher Tencent Tech/BigGo Finance figure. | 低 | SI019 |
| CI030 | Baidu retains SEC reporting obligations as a NASDAQ-listed ADR issuer (ticker BIDU) in addition to its Hong Kong primary listing, meaning any future Kunlunxin-related disclosures could also surface indirectly through Baidu's own SEC filings. | 高 | SI021, SI020 |
| CI031 | Kunlunxin's own official DeepSeek-compatibility announcement describes technical capability (running the 671-billion-parameter DeepSeek R1 model on a single 8-card P800 server) but discloses no pricing, cost, or margin information alongside the technical claim. | 中 | SI022 |
| CI032 | Kunlunxin's earlier IPO spin-off coverage (BigGo Finance, January 2026) frames the listing as part of a broader wave of China AI chip IPOs, alongside Moore Threads and Biren Technology, rather than an idiosyncratic single-company event. | 中 | SI023 |
| CI033 | China Mobile's centralized procurement tender that Kunlunxin won in August 2025 was valued at more than RMB1 billion (~$139 million), the single largest disclosed external commercial contract for Kunlunxin to date. | 中 | SI025 |
| CI034 | Press coverage in Taiwan's TechNews described Kunlunxin's P800 cluster as powering Baidu's Wenxin 4.5 Turbo model, tying the chip's technical narrative directly to Baidu's own AI product roadmap rather than an independent commercial storyline. | 中 | SI026 |
| CI035 | No source reviewed for this chapter provides an audited customer-concentration breakdown (e.g., top-5 customer revenue share) for Kunlunxin, despite named customer wins across telecom, finance, energy, and automotive sectors. | 低 | |
| CI036 | Comparable domestic GPU peers that have already gone public direct IPO proceeds primarily toward R&D and intelligent-computing capacity expansion rather than debt repayment or shareholder distributions, per Biren Technology's own disclosed use-of-proceeds plan, the closest available proxy for how Kunlunxin's own IPO proceeds are likely to be allocated absent its own prospectus. | 中 | SI013, SI004 |
| CI037 | Baidu's roadmap announcement for the M100 (2026, inference-optimized) and M300 (2027, training-optimized) chips implies continued capital expenditure on chip R&D beyond what Kunlunxin's IPO proceeds alone would cover, since the roadmap was announced by Baidu in November 2025 before the IPO structure was finalized. | 中 | SI028, SI029 |
| CI038 | US export-control compliance obligations impose additional administrative and legal cost burdens on foundries and chip designers operating in China's advanced semiconductor supply chain, per a US Government Accountability Office review of BIS rule implementation. | 中 | SI030 |
| CI039 | China invested over $6.1 billion in national computing-hub infrastructure under its East Data West Computing program, a state-financed demand channel distinct from Big Fund III's supply-side semiconductor manufacturing subsidies. | 中 | SI031 |
| CI040 | China Mobile's centralized AI-computing procurement program was itself part of a broader roughly RMB5 billion tender, of which Kunlunxin's RMB1 billion-plus award was one portion, alongside other winners including Huawei-affiliated entities. | 低 | SI033 |
| CE001 | Kunlunxin sells AI accelerator hardware (the Kunlun chip family) plus a CUDA-alternative software stack that lets developers train and serve large models without fully rewriting code for a new hardware target. | 中 | SE001, SE005 |
| CE002 | A model-training customer's workflow with Kunlunxin runs from writing or porting a model in PaddlePaddle, TensorFlow, or PyTorch, through XTCL graph compilation, to execution on XPU-R hardware via a CUDA-like kernel/launch programming model. | 中 | SE001 |
| CE003 | Kunlunxin's own case-study material describes replacing a GPU-based PyTorch industrial quality-inspection pipeline with a Kunlun-Paddle solution, reporting a 65% cost reduction and a 9% performance improvement. | 低 | SE001 |
| CE004 | Kunlunxin's P800-based "Super-Node" products use 32/64-card single-cabinet configurations targeting internet, finance, energy, transportation, and education sector customers. | 中 | SE009 |
| CE005 | Kunlun chip integration with PaddlePaddle progressed from Paddle Lite inference-only support (starting 2018) to full training support in PaddlePaddle 2.0 and later versions. | 中 | SE001, SE005 |
| CE006 | Switching a PaddlePaddle workload to the Kunlunxin XPU backend requires what Kunlunxin describes as a single code-line change, a company-sourced claim not independently benchmarked. | 低 | SE001 |
| CE007 | Kunlunxin's XPU-R architecture separates compute (SDNN plus Cluster units), storage (L3 SRAM plus GDDR6), interconnect, and interface (PCIe) into four hardware blocks. | 中 | SE001 |
| CE008 | The documented XPU-R generation specifies 8 Cluster units and 6 SDNN units delivering up to 128 TFLOPS at FP16, with 64MB of on-chip L3 SRAM and external GDDR6 memory at 512GB/s bandwidth. | 中 | SE001 |
| CE009 | Independent spec aggregation places the P800 (Kunlun III) chip at 7nm, 345 TFLOPS FP16, 96GB of HBM3 memory, and 2.4-4 TB/s of bandwidth. | 中 | SE003 |
| CE010 | TechInsights' independent die-level floorplan analysis confirms the P800 chip uses a 7nm process with over 50 billion transistors and a physical separation of matrix computation from communication units. | 高 | SE004, SE003 |
| CE011 | Kunlunxin's software stack layers a runtime (drivers, monitors, debuggers, profilers), an LLVM-based compiler toolchain, math libraries (XBLAS/XDNN), a communication library (XCCL), and the TVM-based XTCL graph compiler. | 中 | SE001 |
| CE012 | Kunlunxin distributes a FastDeploy XPU backend and a community-maintained vLLM-Kunlun plugin as open-source GitHub repositories, giving it a public developer-signal footprint most hardware-only Chinese competitors lack. | 中 | SE006, SE024 |
| CE013 | The Paddle-Lite XTCL integration documented on GitHub supports deployment on both x86 and ARM servers via a subgraph-based online model analysis and conversion pipeline. | 中 | SE005 |
| CE014 | Baidu activated a 30,000-P800-chip cluster in April 2025, pitched as capable of training "DeepSeek-like" models with hundreds of billions of parameters while supporting fine-tuning for up to 1,000 enterprise customers. | 高 | SE018, SE015 |
| CE015 | Baidu plans to scale a single Kunlun cluster to a million cards by 2030 and to launch "Tianchi256" (first half of 2026) and "Tianchi512" (second half of 2026) cluster configurations, each targeting a further 50%+ performance gain. | 中 | SE020, SE011 |
| CE016 | Kunlunxin's February 2025 announcement claims full DeepSeek R1/V3 training and inference compatibility, with a single 8-card P800 server able to run the full 671-billion-parameter DeepSeek R1 model at 8-bit precision. | 中 | SE009 |
| CE017 | Morgan Stanley's independent estimate of roughly 1,521 tokens/second for P800 on DeepSeek R1 inference, positioning it between Nvidia's H20 and A100, is the only third-party performance estimate located for Kunlunxin's chips. | 低 | SE010 |
| CE018 | On DeepSeek-V4's April 2026 release day, Kunlunxin was among the domestic chips (alongside Huawei Ascend, Cambricon, Hygon, Moore Threads, and Alibaba's T-Head) that completed same-day model adaptation. | 高 | SE028, SE009 |
| CE019 | Kunlunxin's earlier chips used Samsung's 14nm process (Kunlun I) and TSMC's 7nm process (Kunlun II), but no source reviewed officially confirms which foundry manufactures the current P800 chip; industry consensus assumes SMIC's 7nm DUV-based process given post-2021 export-control restrictions on TSMC access. | 中 | SE008, SE007, SE025 |
| CE020 | SMIC's 7nm DUV-based process is estimated to run at 25-50% higher cost and meaningfully lower yield than TSMC's EUV-based equivalent, a cost/yield disadvantage that would apply to P800 if the SMIC attribution is correct. | 中 | SE025, SE026 |
| CE021 | No source reviewed for this chapter discloses independent security certifications, safety-testing standards, or formal quality metrics for any Kunlunxin chip or cluster product. | 低 | |
| CE022 | No source reviewed confirms Kunlunxin is currently on the US Entity List, though broader 2025 export-control actions targeting indigenous Chinese semiconductor production create latent compliance risk not addressed in any Kunlunxin-published material reviewed. | 低 | SE013 |
| CE023 | No source reviewed discloses a Kunlunxin-specific chip yield rate; the closest available proxy is Huawei's disclosed yield improvement from about 20% to nearly 40% at SMIC, which is not directly transferable to Kunlunxin given Huawei's reported production-priority status. | 中 | SE014 |
| CE024 | No source reviewed discloses a patent count, IP litigation history, or formal IP-protection disclosure for Kunlunxin's XPU architecture, in contrast to peers such as Biren Technology that reference IP positioning in their own disclosures. | 低 | |
| CE025 | No source reviewed discloses independent safety-testing or reliability data for Kunlun II's deployment in Baidu's Apollo autonomous-driving platform. | 低 | |
| CE026 | Kunlun I was deployed in over 20,000 units within Baidu's own search engine and DuerOS voice-assistant products, according to Kunlunxin's own technical salon disclosure. | 中 | SE001, SE016 |
| CE027 | Kunlun II was integrated into Baidu's Apollo autonomous-driving software platform and the Ernie Bot deep learning model, extending the chip's use beyond cloud/search workloads into automotive AI. | 中 | SE015, SE017 |
| CE028 | Tencent began sourcing Kunlunxin P800 chips at scale in 2026, running them in 10,000-card clusters, per China Biz Insider's July 2026 reporting. | 中 | SE027 |
| CE029 | Kunlunxin's P800 cluster reportedly trained Baidu's Wenxin 4.5 Turbo model in April 2025 and later the Wenxin 5.1 model on an all-domestic chip stack. | 中 | SE027 |
| CE030 | Kunlunxin's M100 chip's physical product made its first public appearance in a July 2026 CCTV state television news report, following a January 2026 start of commercial ramp-up. | 中 | SE021 |
| CE031 | Baidu's M100/M300 roadmap announcement (November 2025) commits the company to releasing new Kunlun chip generations annually for at least five years. | 中 | SE019, SE022 |
| CE032 | Baidu's Kunlun chip architecture was originally researched and emulated on FPGAs before the company built dedicated silicon in 2018, a multi-year R&D path predating most rival Chinese AI chip startups founded after 2019. | 中 | SE015, SE016 |
| CE033 | No independently reproducible benchmark suite (e.g., MLPerf-style) comparing Kunlunxin's chips against Huawei Ascend or Cambricon chips on identical workloads was located in any source reviewed. | 低 | |
| CE034 | CSDN developer-community coverage of Kunlunxin's DeepSeek support included downloadable documentation and confirmed a vLLM-Kunlun interface for one-click deployment, corroborating Kunlunxin's own official DeepSeek-compatibility announcement. | 低 | SE023 |
| CE035 | Kunlunxin's XPU-R programming model follows a CUDA-like paradigm with kernels, host-side launch parameters, and explicit host-device memory transfers over PCIe, lowering the conceptual migration barrier for CUDA-experienced developers even though real-world migration still requires meaningful engineering effort. | 中 | SE001 |
| CU001 | Baidu is simultaneously Kunlunxin's controlling shareholder, largest customer, and technology co-developer, consuming P800 chips for its own Wenxin/Ernie model training and Apollo autonomous-driving stack. | 中 | SU014, SU023 |
| CU002 | Telecom and financial SOEs (China Mobile, China Merchants Bank) buy Kunlunxin chips through centralized procurement tenders rather than direct negotiated sales. | 高 | SU001, SU006 |
| CU003 | External (non-Baidu) revenue accounted for roughly 40% of Kunlunxin's total revenue in 2024, according to press estimates. | 中 | SU006 |
| CU004 | External (non-Baidu) revenue share reportedly rose to more than 50% of Kunlunxin's total revenue in 2025. | 中 | SU006, SU015 |
| CU005 | Kunlunxin's 2025 China AI accelerator shipment volume was approximately 116,000 units (tied with Cambricon), roughly 3% of the total China market and well behind Huawei's approximately 812,000 units. | 高 | SU016, SU025 |
| CU006 | No source reviewed provides a customer-level breakdown of unit shipments, active deployments, or utilization for any individual named Kunlunxin customer. | 低 | |
| CU007 | In August 2025, Kunlunxin's AI servers won a leading position across all three general-computing bid packages in a China Mobile centralized procurement tender worth more than RMB1 billion (~$139 million), with chips supplied through system integrators H3C and ZTE. | 高 | SU001, SU003 |
| CU008 | China Mobile's centralized AI-computing tender that Kunlunxin won was part of a broader roughly RMB5 billion procurement program, with Kunlunxin one of multiple winners across the program. | 低 | SU003 |
| CU009 | Yicai Global's November 2025 reporting names China Merchants Bank, China Southern Power Grid, Geely Automobile Holdings, and Vivo as external Kunlunxin chip deployers, alongside a company claim of "hundreds of real-world users" across internet, finance, energy, transportation, and education sectors. | 高 | SU005, SU027 |
| CU010 | Kunlunxin's own case-study material describes an industrial quality-inspection deployment achieving a 65% cost reduction and a 9% performance improvement versus a prior GPU-based PyTorch pipeline. | 低 | SU010 |
| CU011 | Tencent began sourcing Kunlunxin P800 chips at scale in 2026, running them in 10,000-card clusters, according to China Biz Insider's reporting citing The Information. | 中 | SU004 |
| CU012 | Kunlunxin's P800 cluster reportedly trained Baidu's Wenxin 4.5 Turbo model in April 2025 and later the Wenxin 5.1 model on an all-domestic chip stack. | 中 | SU004, SU014 |
| CU013 | Kunlunxin's Hong Kong IPO reportedly requires cornerstone investors to commit to purchasing chips worth three to seven times their equity investment, pre-loading revenue from investor-customers rather than organically sourced accounts. | 高 | SU007, SU018 |
| CU014 | No source reviewed discloses a retention metric, renewal rate, churn figure, net revenue retention, or contract-length data for any named Kunlunxin customer. | 低 | |
| CU015 | Baidu has purchased every Kunlun chip generation since 2018 (Kunlun I, II, P800, and the announced M100/M300 roadmap), the clearest available repeat-purchase proxy among Kunlunxin's named accounts. | 中 | SU013, SU010 |
| CU016 | Geely's relationship with Kunlun chips traces back to the Kunlun II / Apollo autonomous-driving era, suggesting some multi-year continuity, though current deployment scale and whether Geely has adopted P800 specifically are not disclosed. | 低 | SU013, SU005 |
| CU017 | Baidu's dual role as Kunlunxin's ~59.45% controlling shareholder and largest customer creates a structural related-party concentration risk that cannot be diversified away simply by adding more external customers. | 中 | SU023, SU006 |
| CU018 | Named SOE customer wins (China Mobile, China Merchants Bank) came through formal, competitively bid centralized tenders rather than a captive-only sales motion, evidence of a real procurement gate for at least some of Kunlunxin's external revenue. | 中 | SU001, SU006 |
| CU019 | No public customer complaint, procurement dispute, or reported deployment failure involving Kunlunxin was located in any source reviewed for this chapter as of the run date. | 低 | |
| CU020 | BYD is confirmed only as a Kunlunxin equity investor (an early ~0.3% stake per DigiTimes, and a participant in the 2025 Series D round), with no source reviewed confirming BYD as a Kunlunxin chip customer. | 中 | SU011, SU012 |
| CU021 | ByteDance reportedly evaluated Kunlunxin as a potential supplier for chatbot inference workloads alongside Iluvatar CoreX, without confirming Kunlunxin as a finalized primary ByteDance supplier as of the sources reviewed. | 中 | SU020 |
| CU022 | Iluvatar CoreX depends heavily on a single customer relationship (ByteDance ordering 50,000+ units for chatbot inference), a customer-concentration profile that contrasts with Kunlunxin's more diversified named customer list across telecom, finance, energy, automotive, and hyperscaler segments. | 中 | SU022, SU020 |
| CU023 | On DeepSeek-V4's April 2026 release day, Kunlunxin was among the domestic chips that completed same-day model adaptation, alongside Huawei Ascend, Cambricon, Hygon, Moore Threads, and Alibaba's T-Head, a customer-relevant proof point since it demonstrates continued compatibility with the workloads Kunlunxin's customers are most likely to run. | 中 | SU021 |
| CU024 | China's East Data West Computing (东数西算) national infrastructure program invested over $6.1 billion in computing hubs as of August 2024, a state-financed demand channel that could plausibly include Kunlunxin chips, though no source reviewed confirms Kunlunxin's specific participation in these hubs. | 低 | SU009 |
| CU025 | Kunlunxin's overall China AI accelerator market position (approximately 3% shipment share) means its customer base, while diversifying, remains far smaller in scale than Huawei's, which supplies roughly 20% of the total market to a similarly broad set of hyperscaler, SOE, and government customers. | 中 | SU016, SU025 |
| CU026 | The gap between Kunlunxin's six independently named external customers and Baidu's own "hundreds of real-world users" claim is not reconciled by any source reviewed, leaving the true breadth of Kunlunxin's external customer base unconfirmed beyond the named accounts. | 低 | SU005 |
| CU027 | No source reviewed discloses the contract length or renewal cycle for the China Mobile or China Merchants Bank tender wins. | 低 | |
| CU028 | Sahm Capital's December 2025 coverage frames the August 2024 China Mobile chip order as a "watershed moment" that first confirmed Kunlunxin's external commercial viability beyond Baidu. | 中 | SU017 |
| CU029 | TrendForce's May 2025 industry tracking notes that Baidu's Kunlun chips, alongside Tencent's Zixiao and Alibaba's T-Head PPU, were entering production for cloud inference at scale as part of a broader domestic AI chip independence trend. | 中 | SU019 |
| CU030 | Analysts have flagged the chip-purchase-linked IPO structure as raising circular-financing concerns about whether Kunlunxin's reported customer growth reflects genuine external demand or artificially inflated near-term revenue commitments tied to the IPO itself. | 中 | SU007, SU008 |
| CU031 | Kunlunxin's 30,000-chip P800 cluster, activated in April 2025, was designed to simultaneously support fine-tuning workloads for up to 1,000 enterprise customers, a scale claim not independently itemized by named account. | 中 | SU014 |
| CU032 | No source reviewed discloses a customer satisfaction survey, Net Promoter Score, or review-platform rating for any Kunlunxin chip or cluster product. | 低 | |
| CU033 | Kunlunxin's shipment volume trails Alibaba's T-Head by more than 2x (approximately 116,000 versus 265,000 units in 2025), placing Kunlunxin fourth rather than second or third among all domestic AI chip vendors by unit volume when T-Head is included. | 中 | SU016, SU025 |
| CU034 | BYD's Series D equity investment in Kunlunxin (2025) followed an earlier, much smaller ~0.3% stake acquired around 2023, showing a strategic relationship that deepened over time even without confirmed chip purchases. | 中 | SU011, SU012 |
| CU035 | Kunlunxin's named customer roster spans six sectors (Baidu-internal, telecom, finance, energy, automotive, and hyperscaler-rival), a broader sector spread than most single-generation domestic GPU peers disclose in press coverage reviewed. | 中 | SU005, SU004 |
| CU036 | Huawei's roughly $12 billion 2026 AI chip revenue projection dwarfs Kunlunxin's disclosed and estimated commercial scale, underscoring the gap between Kunlunxin's named-customer diversification and Huawei's much larger customer base across the same buyer segments. | 中 | SU026 |
| CU037 | Tech in Asia's July 2026 reporting corroborates that Kunlunxin's external customer revenue rose to roughly 50% of the total, from about 40% previously, and separately highlights the China Mobile RMB1 billion server procurement win as proof of commercial expansion beyond Baidu. | 中 | SU027 |
| CU038 | Biren Technology's customer base, per Tech in Asia's coverage of its pre-IPO funding round, includes China Mobile and China Telecom among named telecom customers, a segment overlap with Kunlunxin's own China Mobile relationship that illustrates shared SOE buyers across competing domestic chip vendors. | 低 | SU028 |
| CU039 | ByteDance and Alibaba's competition for constrained Huawei Ascend chip allocation illustrates that demand for domestic AI chips broadly exceeds supply, a market backdrop that could favor Kunlunxin capturing overflow demand from buyers unable to secure sufficient Huawei allocation. | 低 | SU029 |
| CU040 | Independent reporting corroborates that Baidu's 30,000-chip P800 cluster, activated in April 2025, was specifically used to train the Wenxin 4.5 Turbo model, consistent with Kunlunxin's own claims about internal deployment scale. | 中 | SU030, SU014 |
| CR001 | No source reviewed confirms Kunlunxin is currently on the US Entity List, but the risk is elevated given its growing commercial and technical profile and the broader pattern of 2025 export-control actions. | 低 | SR001 |
| CR002 | BIS issued rules in January 2025 tightening foundry due-diligence requirements and adding PRC/Singapore entities to the Entity List, building on prior controls from October 2022, October 2023, and December 2024. | 高 | SR003, SR004 |
| CR003 | A proposed "Affiliates Rule" that would have extended US export controls to entities 50%-or-more owned by already-restricted parties was introduced then suspended for one year starting November 2025. | 高 | SR002, SR005 |
| CR004 | More than 120 new Chinese entities were added to the US Entity List during 2025, per a year-end legal review. | 中 | SR002 |
| CR005 | Taiwan imposed new export restrictions on Huawei and SMIC in June 2025, tightening access to advanced packaging, EDA tools, and materials for any SMIC customer. | 中 | SR006 |
| CR006 | Kunlunxin's dual Hong Kong/STAR Market listing process requires simultaneous compliance with HKEX and CSRC disclosure regimes, adding regulatory-process complexity beyond a single-market listing. | 中 | SR020, SR012 |
| CR007 | No source reviewed identifies a clear regulatory precedent for how Hong Kong's SFC or the CSRC would require Kunlunxin's IPO chip-purchase-commitment structure to be disclosed in a public prospectus. | 低 | |
| CR008 | Industry consensus, not company-confirmed, places Kunlunxin's P800 production on SMIC's 7nm DUV-based process, which runs an estimated 25-50% higher cost and meaningfully lower yield than TSMC's EUV-based equivalent. | 中 | SR008, SR030 |
| CR009 | SMIC reportedly prioritizes Huawei's wafer allocation over other domestic customers including Cambricon and Kunlunxin, a structural manufacturing disadvantage smaller vendors cannot easily replicate. | 高 | SR007, SR025 |
| CR010 | Domestic HBM producer CXMT remains immature relative to Samsung/SK Hynix, and Kunlunxin's specific memory-sourcing arrangement is not disclosed in any source reviewed. | 中 | SR008 |
| CR011 | No source reviewed discloses a Kunlunxin-specific chip yield rate; the closest available proxy is Huawei's disclosed yield improvement from about 20% to nearly 40% at SMIC, which is not directly transferable to Kunlunxin given Huawei's reported production-priority status. | 中 | SR007 |
| CR012 | No independently reproducible benchmark suite comparing Kunlunxin's chips against Huawei Ascend, Cambricon, or Nvidia chips on identical workloads was located in any source reviewed. | 低 | |
| CR013 | No source reviewed discloses any independent safety, security, or quality-control certification for any Kunlunxin chip or cluster product. | 低 | |
| CR014 | Baidu is simultaneously Kunlunxin's ~59.45% controlling shareholder, its largest customer, and the sponsor of its IPO, concentrating governance, commercial, and capital-markets leverage in a single counterparty. | 高 | SR015, SR012 |
| CR015 | Kunlunxin's dependence on SMIC as its believed sole foundry is compounded by SMIC's own capacity constraints and reported prioritization of Huawei's orders. | 中 | SR007, SR008 |
| CR016 | Kunlunxin's software stack (XTCL/XBLAS/XDNN) depends on external open-source infrastructure -- PaddlePaddle, Apache TVM, and LLVM -- that Kunlunxin does not fully control, though its Baidu affiliation with PaddlePaddle partially mitigates this dependency. | 中 | SR031 |
| CR017 | Kunlunxin's two largest disclosed external commercial relationships (China Mobile and Tencent) are each individually significant enough that non-renewal or non-recurrence of either would materially affect Kunlunxin's external-revenue growth narrative. | 中 | SR025, SR023 |
| CR018 | Kunlunxin's reported IPO chip-purchase-commitment structure ties investor capital to revenue commitments, meaning execution risk in the IPO process could simultaneously threaten both the capital raise and the revenue pipeline it is designed to pre-load. | 高 | SR010, SR021 |
| CR019 | Analysts have flagged the chip-purchase-linked IPO structure as raising circular-financing concerns about whether reported external demand reflects the structure itself rather than organic commercial pull. | 中 | SR010, SR011 |
| CR020 | Ouyang Jian is the sole named Kunlunxin-specific executive in the public record, serving as CEO since the 2021 spin-off with no disclosed successor or deputy. | 中 | SR014, SR022 |
| CR021 | Every other named Kunlunxin-related leader (Robin Li, Shen Dou, Haijian He) is a Baidu corporate officer whose Kunlunxin responsibilities are secondary to their Baidu roles, meaning Kunlunxin lacks disclosed full-time, independent C-suite depth beyond its CEO. | 中 | SR015, SR022 |
| CR022 | Kunlunxin must simultaneously execute a dual Hong Kong/STAR Market listing process, a chip-purchase-linked IPO structure, and an accelerating external-revenue growth target, compounding execution risk across several fronts at once. | 中 | SR020, SR021 |
| CR023 | Kunlunxin's external revenue share must continue rising from over 50% (2025) to sustain its IPO growth narrative, an execution target dependent on continued SOE tender wins and hyperscaler-rival customer growth like the 2026 Tencent relationship. | 中 | SR027, SR023 |
| CR024 | A confirmed BIS Entity List designation naming Kunlunxin (Beijing) Technology Co., Ltd. directly would be an unambiguous thesis-break trigger given its potential effect on foundry access and investor appetite. | 中 | SR003 |
| CR025 | Publication of Kunlunxin's audited FY2024/FY2025 revenue in a public HKEX prospectus would resolve the current 5-6x estimate gap between CNBC's (~$500M) and Tencent Tech/BigGo Finance's (~$2.9B) 2025 revenue figures, materially changing any valuation-multiple analysis. | 中 | SR027, SR028 |
| CR026 | Biren Technology's continued operation under a US Entity List designation since October 2023, alongside its successful January 2026 Hong Kong IPO, shows that an Entity List designation does not necessarily prevent a Chinese chipmaker from completing a public listing, though it does add investor-facing risk disclosure. | 高 | SR009, SR018, SR019 |
| CR027 | The Council on Foreign Relations assesses that the compute-performance gap between the best US and Chinese AI chips is structural and likely to widen rather than narrow, an adverse signal for the long-term competitiveness of all domestic Chinese chipmakers including Kunlunxin. | 中 | SR013, SR026 |
| CR028 | Kunlunxin's risk profile differs from already-listed peers Cambricon and Moore Threads in one important respect: both peers have already cleared public-market scrutiny and disclosed audited financials, while Kunlunxin's financial and governance risks remain estimated rather than independently verified pending its own prospectus. | 中 | SR032 |
| CR029 | As of the report run date, Kunlunxin's confidential HKEX Form A1 filing (submitted January 1, 2026) remains in process, with a parallel Shanghai STAR Market "tutoring" process underway since May 2026 and Jefferies analysts (per BigGo Finance) expecting the Hong Kong listing in the third quarter of 2026. | 中 | SR022, SR020 |
| CR030 | Manufacturing-capacity risk at SMIC transmits directly into Kunlunxin's shipment-volume growth, which in turn affects reported revenue, the credibility of the IPO valuation narrative, and ultimately investor pricing of the reported ~$50 billion target. | 中 | SR007, SR027 |
| CR031 | Across regulatory, operational, partner, and people/execution risk categories, the highest-severity, lowest-mitigation-maturity risks for Kunlunxin cluster around export-control exposure, SMIC manufacturing dependency, and the Baidu governance/customer overlap. | 中 | SR007, SR001, SR015 |
| CR032 | No litigation, safety incident, or confirmed regulatory enforcement action involving Kunlunxin specifically was located in any source reviewed for this chapter as of the report run date. | 低 | |
| CR033 | The regulatory and legal risk exposures identified for Kunlunxin span at least three jurisdictions -- the United States (BIS export controls and potential Entity List action), Taiwan (export restrictions on SMIC), and Hong Kong/China (dual HKEX/CSRC listing compliance). | 中 | SR003, SR006, SR012 |
| CR034 | No source reviewed discloses any dual-sourcing or foundry-diversification mitigation strategy for Kunlunxin's dependency on SMIC. | 低 | |
| CR035 | The IFP's policy analysis argues that the H20 export ban was counterproductive and primarily accelerated China's domestic chip ecosystem without materially slowing Chinese AI development, a viewpoint that complicates simple assumptions about the effectiveness of further export-control escalation against vendors like Kunlunxin. | 中 | SR017 |
| CR036 | CSIS's analysis concludes that while EUV lithography access remains blocked to China, domestic chip ecosystems including Kunlunxin's continue to advance within DUV-based constraints, suggesting export controls slow but do not halt Kunlunxin's manufacturing progress. | 中 | SR016 |
| CR037 | Reporting as recent as August 2026 found many Chinese AI developers still accessing Nvidia compute via offshore cloud rental specifically to avoid domestic-chip migration costs, an adverse signal for the durability of demand for domestic chips including Kunlunxin's. | 中 | SR029 |
| CR038 | Kunlunxin's M100 chip is in commercial ramp-up as of January 2026, while the M300 chip remains a 2027 target with no shipped product yet, meaning roughly half of the disclosed forward roadmap remains execution risk rather than delivered product. | 中 | SR024, SR015 |
| CR039 | Kunlunxin's overall 2025 China AI accelerator shipment share (approximately 3%) trails Huawei's (approximately 20%) by a wide margin, meaning Kunlunxin's scale-based execution risk is more acute than that of the market leader. | 高 | SR025, SR007 |
| CR040 | Kunlunxin's IPO process is being advised in part by China International Capital Corp. (CICC) for its STAR Market track, per a CICC filing to the China Securities Regulatory Commission cited by GuruFocus. | 中 | SR020 |
| CR041 | Kunlunxin's HKEX Form A1 filing remained confidential as of the report run date, meaning the specific regulatory risk factors and disclosure requirements that will govern its public prospectus are not yet known. | 高 | SR012, SR034 |
| CR042 | The gap between Kunlunxin's roughly $3.0 billion Series D valuation (July 2025) and its reported ~$50 billion IPO target (June 2026), a roughly 16x increase in under twelve months, is itself flagged by multiple analysts as a valuation-credibility risk independent of the chip-purchase-commitment structure. | 中 | SR027, SR011 |
| CR043 | No source reviewed discloses whether Kunlunxin carries any debt or project-finance obligations that could compound the risks associated with its capital-intensive manufacturing dependency. | 低 | |
| CR044 | Kunlunxin's Series D investor base includes state-linked capital (China Reform Holdings' High-Level Talent Fund) alongside strategic industrial investors (BYD, China Mobile's Zhongyi Hechuang), meaning any governance or disclosure risk also touches state-affiliated capital providers. | 中 | SR022 |
| CR045 | No public source reviewed discloses Kunlunxin's employee headcount, limiting assessment of whether its organizational scale is sufficient to support both its current manufacturing/commercial execution and a dual public-listing process simultaneously. | 低 | |
| CR046 | The reported IPO chip-purchase-commitment structure has drawn coverage from multiple international outlets (The Next Web, Startup Fortune, CryptoBriefing) as an unusual condition rarely seen in global IPO practice. | 中 | SR033, SR010 |
| CR047 | Kunlunxin formally initiated STAR Market IPO "tutoring" (辅导) -- a mandatory pre-IPO regulatory coaching phase under Chinese securities rules -- in May 2026, adding a distinct domestic regulatory process alongside its Hong Kong filing. | 中 | SR034 |
| CR048 | Chinese regulators investigated Nvidia's H20 chip over reported backdoor/remote-disable capabilities in 2025, leading some state-owned enterprises to suspend H20 purchases even after the US export ban was partially lifted -- a regulatory dynamic illustrating that Chinese authorities can independently restrict foreign chips in ways that indirectly benefit domestic vendors like Kunlunxin. | 中 | SR035 |
| CR049 | Manufacturing costs for China's DUV-based domestic semiconductor ecosystem run an estimated 25-35% higher than TSMC EUV equivalents, with lower yields, though state subsidies partially compensate -- a bifurcation dynamic that applies to Kunlunxin if the SMIC foundry attribution is correct. | 中 | SR036 |
| CV001 | Kunlunxin sits inside China's fastest-growing chip segment, with domestic AI accelerator vendors' collective China shipment share crossing 41% in 2025 and continuing to rise under sustained US export controls and "Buy China" policy support. | 中 | SV018 |
| CV002 | Tencent, a direct Baidu rival, began sourcing Kunlunxin P800 chips at scale in 2026, which multiple outlets treat as the strongest available third-party validation of Kunlunxin's product. | 高 | SV001, SV013 |
| CV003 | Kunlunxin's external (non-Baidu) revenue share reportedly rose from roughly 40% in 2024 to over 50% in 2025, suggesting genuine commercial traction beyond captive Baidu demand. | 中 | SV017 |
| CV004 | Kunlunxin's reported Hong Kong IPO target of roughly $50 billion is approximately 16x its July 2025 Series D valuation of about $3.0 billion, an escalation achieved in under twelve months. | 高 | SV001, SV002 |
| CV005 | Kunlunxin's requirement that cornerstone investors buy chips worth three to seven times their equity stake has drawn analyst concern that reported external demand may partly reflect the IPO structure itself rather than organic commercial pull. | 中 | SV009, SV010 |
| CV006 | Kunlunxin's 2025 revenue is estimated anywhere from approximately $500 million to $2.9 billion depending on the source, a 5-6x unreconciled gap that undermines confident multiple-based valuation. | 中 | SV002, SV008 |
| CV007 | Baidu's dual role as Kunlunxin's ~59.45% controlling shareholder and largest customer creates a related-party structure that limits how independently Kunlunxin's growth story can be verified. | 中 | SV016 |
| CV008 | Given unresolved revenue, manufacturing, and IPO-structure evidence gaps, this report's recommendation for Kunlunxin is research-more rather than buy or avoid. | 中 | SV001, SV008 |
| CV009 | This report assigns Kunlunxin a high risk rating given export-control, manufacturing-dependency, and governance-concentration risks documented in the Risks chapter. | 中 | SV009, SV011 |
| CV010 | Kunlunxin's valuation progressed from approximately $1.9 billion (April 2021 spin-off) to approximately $3.0 billion (July 2025 Series D), then escalated through IPO-era guidance of ~$14.7 billion (May 2026), ~HK$100 billion / ~$13 billion (STAR Market guidance, May 2026), and ~$50 billion (June 2026 roadshow). | 高 | SV017, SV003, SV012, SV001 |
| CV011 | This report assigns Kunlunxin a stretched-to-expensive valuation stance given the wide implied price-to-sales range (17x to 100x) depending on which unreconciled 2025 revenue estimate is used. | 中 | SV008, SV010 |
| CV012 | No source reviewed discloses preference-stack, liquidation-preference, or dilution terms for any of Kunlunxin's private funding rounds. | 低 | |
| CV013 | Only China International Capital Corp.'s (CICC) role advising Kunlunxin's STAR Market IPO track is confirmed, via a CICC filing to the China Securities Regulatory Commission cited by GuruFocus; the Hong Kong underwriter syndicate remains undisclosed. | 中 | SV012 |
| CV014 | In a bull scenario, Kunlunxin's 2025 revenue reconciles near the higher ~$2.9 billion estimate, implying a price-to-sales multiple around 17x at the $50 billion target, broadly comparable to Cambricon's 55-85x range on a smaller revenue base. | 中 | SV008, SV004 |
| CV015 | In a base scenario, Kunlunxin's revenue settles nearer a J.P. Morgan-consistent ~$1.2-1.8 billion range for 2025-2026, with the IPO pricing closer to the earlier ~$13 billion STAR Market guidance rather than the escalated $50 billion figure. | 中 | SV003, SV025 |
| CV016 | In a bear scenario, Kunlunxin's revenue reconciles toward the lower ~$500 million CNBC estimate, the Tencent order proves to be one-off, and the IPO is delayed, repriced downward, or draws regulatory scrutiny over the chip-purchase-commitment structure. | 中 | SV002, SV009 |
| CV017 | This report assigns the base scenario the highest probability signal among the three scenarios, reflecting that it is most consistent with the balance of evidence reviewed across this report. | 低 | SV003, SV025 |
| CV018 | Kunlunxin's relevant comparable set spans four already-public or newly-public Chinese AI chip companies (Cambricon, Moore Threads, Biren Technology, Iluvatar CoreX) plus Nvidia as the global multiple benchmark. | 中 | SV004, SV006, SV007 |
| CV019 | None of Kunlunxin's comparables is a perfect match: Cambricon and Moore Threads are already profitable or further along their public-market track, while Biren Technology and Iluvatar CoreX are closer in scale but differ in Entity List status or customer concentration. | 中 | SV006, SV007 |
| CV020 | Kunlunxin's confidential HKEX Form A1 filing (January 2026) and parallel STAR Market tutoring process (May 2026) are the mechanism by which existing private investors (Series A-D) would realize IPO-driven liquidity. | 高 | SV016, SV025 |
| CV021 | Jefferies analysts, per BigGo Finance/GuruFocus reporting, expect Kunlunxin's Hong Kong listing to occur in the third quarter of 2026. | 中 | SV012 |
| CV022 | No source reviewed discloses lock-up terms or secondary-sale provisions for Kunlunxin's existing private investors ahead of the IPO. | 低 | |
| CV023 | The most critical unresolved diligence asks for Kunlunxin are, in priority order: reconciled revenue, confirmed manufacturing foundry/yield, independent benchmark data, confirmed underwriter syndicate, and clarity on the chip-purchase-commitment structure's prospectus treatment. | 中 | SV001, SV012 |
| CV024 | Cambricon's market capitalization reached roughly RMB1 trillion (~$140 billion) in mid-2026 and trades at a 200-373x trailing PE, the highest valuation multiple among domestic Chinese AI chip comparables. | 高 | SV004, SV005 |
| CV025 | Biren Technology's post-IPO market capitalization reached approximately $12.3 billion after a 72.86% debut pop in January 2026, on FY2025 revenue of RMB1,034.6 million and a 53.8% gross margin per its audited HKEX annual results. | 中 | SV006 |
| CV026 | Iluvatar CoreX's post-IPO market capitalization reached approximately $5.23 billion after its January 2026 HKEX debut, on FY2025 revenue of RMB1,033.6 million (+91.6% YoY) and an adjusted net loss of RMB437.7 million per its audited annual results. | 中 | SV007 |
| CV027 | Nvidia's global market capitalization exceeds $3 trillion and it trades at a 16-25x revenue multiple, a benchmark against which Chinese AI chip comparables' 55-373x multiples represent a substantial premium. | 中 | SV011 |
| CV028 | BigGo Finance's analysis frames Kunlunxin's $50 billion IPO target as exceeding its own controlling parent Baidu's approximately $38 billion market capitalization by roughly 1.32x, an unusual "subsidiary out-valuing parent" dynamic. | 中 | SV008 |
| CV029 | China's roughly $47.5 billion Big Fund III state semiconductor financing initiative, alongside broader sovereign-compute policy support, underpins investor appetite for domestic AI chip IPOs generally, though no source quantifies its specific effect on Kunlunxin's valuation. | 低 | SV023 |
| CV030 | Kunlunxin's decade-long chip R&D history (internal Baidu development since 2011, independent since 2021) gives it a maturity edge over several newer domestic chip startups founded after 2019, a qualitative valuation-supportive factor not directly captured in any single financial metric. | 中 | SV020, SV019 |
| CV031 | Kunlunxin's P800 chip specifications (7nm, 345 TFLOPS FP16, 96GB HBM3) support its positioning between Nvidia's A100 and H100 generations, a technical bull-case data point independent of the disputed revenue figures. | 中 | SV022 |
| CV032 | Kunlunxin's M300 chip, targeting trillion-parameter multimodal model training in 2027, represents unpriced forward optionality in the current valuation that is not yet reflected in any comparable-company multiple. | 低 | SV021 |
| CV033 | Sina Finance's December 2025 reporting projected Kunlunxin's 2026 revenue at approximately RMB8 billion (~$1.1 billion), a figure that sits within this chapter's base-case revenue range. | 中 | SV024 |
| CV034 | No target return, hold period, or exit-window figure is supportable from current public evidence given the unresolved revenue, valuation, and IPO-timing uncertainty. | 低 | |
| CV035 | Kunlunxin's SEC-reporting parent (Baidu, NASDAQ:BIDU) provides an indirect disclosure channel through its own 20-F filings, though these consolidate Kunlunxin's results rather than disclosing them on a standalone audited basis. | 高 | SV015, SV016 |
| CV036 | Kunlunxin's shareholder register, per Tracxn's funding database, spans at least seven institutional investors across its funding history, though a complete cap table is not publicly disclosed. | 低 | SV014 |
| CV037 | The absence of any independently reproducible benchmark comparing Kunlunxin's chips to Huawei, Cambricon, or Nvidia chips is itself a valuation risk factor, since competitive-positioning claims underlying the bull case rest on a single analyst estimate rather than verifiable data. | 低 | |
| CV038 | Domestic chip shipment growth (41%+ share, 2025) and the DeepSeek-V4 same-day adaptation event (April 2026) together support the market-scale component of Kunlunxin's bull case, independent of its disputed firm-specific revenue figures. | 中 | SV018, SV019 |
| CV039 | Kunlunxin's IPO timeline risk is compounded by the need to simultaneously satisfy HKEX and CSRC disclosure regimes across its dual-listing structure, a process-execution factor relevant to both the bear scenario and the exit-timeline diligence ask. | 中 | SV003, SV025 |
| CV040 | No public source discloses Kunlunxin's employee headcount or debt/project-finance obligations, both standard disclosure items already available for its listed comparables (Cambricon, Moore Threads, Biren, Iluvatar CoreX). | 低 | |
| CV041 | A Congressional Research Service review of US export controls on advanced Chinese semiconductors frames the broader regulatory environment Kunlunxin operates within, noting continued scrutiny that could affect investor risk-pricing of its IPO even absent a direct designation against Kunlunxin itself. | 中 | SV026 |
| CV042 | Baidu's official announcement of the Kunlunxin spin-off states the listing is intended to "unlock significant value for shareholders" and broaden Kunlunxin's access to financing channels, framing the IPO explicitly as a value-unlocking event from the parent's own perspective. | 中 | SV027, SV016 |
| CV043 | TrendForce's December 2025 reporting on Baidu's exploration of the Kunlunxin spin-off framed it as part of a broader surge in China chip IPOs, providing early confirmation of the listing process later formalized in the January 2026 HKEX filing. | 中 | SV028 |
| CV044 | Cambricon's market capitalization crossing RMB1 trillion (~$140 billion) in June 2026 set a ceiling reference point for what STAR Market and HKEX investors are currently willing to pay for a Chinese AI chip company, directly relevant to assessing whether Kunlunxin's ~$50 billion target is achievable. | 中 | SV029, SV004 |
| CV045 | SMIC's capacity-constrained 7nm DUV manufacturing strategy underpins the cost and yield disadvantage Kunlunxin would face if the SMIC foundry attribution is correct, a structural factor relevant to assessing the durability of Kunlunxin's margin path implied by its valuation. | 中 | SV030 |