China’s Internet Giants Are Listing AI Units Because the Parent Stock Will Not Price Them

资讯 AI 行业观察 作者:Chelsea_Sun 全文 4369 字 · 约 11 分钟读完 配图 1 张 分类快手百度

Alibaba, Baidu, and Kuaishou are sending chip and model assets to separate markets. The subsidiaries attract lofty private or target valuations; the parents do

所属频道 大厂动态

NextFin News — Alibaba said it will take T-Head, its semiconductor arm, to a Hong Kong main-board listing. The parent will keep control. Fundraising size and final valuation were left open. Within a year, Baidu’s Kunlunxin and Kuaishou’s Kling have followed the same broad path: pull a core AI asset out where the market can assign it a number of its own.

Those numbers have been aggressive. Target or post-money figures discussed for the subsidiaries have, at times, rivaled or exceeded the market value of the parent. The parents’ shares have not moved as if investors were simply adding the pieces together. After Alibaba’s T-Head notice, the U.S.-listed stock barely recovered from an early slide. The pattern is familiar in holding-company math: the market either doubts the subsidiary price or applies a deep discount to the equity the parent still holds.

T-Head is expected to remain a consolidated subsidiary. Tape-out costs and research spending will still hit Alibaba’s combined results. Only a minority slice is likely to be sold. Relative to Alibaba’s quarterly capital expenditure, the cash raised would be secondary. The company itself described the spin as not material against group assets, revenue, and profit.

What changes is the ticker. An asset that used to sit inside cloud cost now gets a daily quote. That is the product being sold to public investors: a price, not a change in who runs the chips or who pays for the next process node.

Alphabet has kept TPU inside the group for more than a decade. Its multiple expanded as investors bought the full stack—from silicon through cloud and models to distribution—and recent external TPU deals still clear through the same set of shareholders. Heavy capital spending is backed by search-advertising cash flow.

Alibaba tells a similar full-stack story in its filings and on stage. On the cash-flow statement the stack shows up first as outflow. Free cash flow has swung negative as infrastructure spending rose, while long-only holders of China internet names have pressed for buybacks and dividends. Waiting for the market to capitalize chip value inside the parent the way it does for Alphabet is a luxury the balance sheet does not clearly afford.

Commercial reasons point the same way. T-Head needs volume beyond Alibaba Cloud to amortize fixed design and manufacturing cost. Domestic buyers in telecom, state finance, and the broader enterprise market care about supplier governance; a clean listed Chinese chip company is easier to put on a vendor list than a division of an overseas-listed conglomerate. Cloud itself wants a neutral shelf—Nvidia and other domestic GPUs beside in-house silicon—so rivals are less wary of sharing road maps. Engineers, watching listed peers, want equity that tracks chip outcomes rather than e-commerce cycles.

Once T-Head has outside shareholders, every internal decision carries two prices. Alibaba Cloud wants cheap silicon to shorten the payback on AI capacity. T-Head’s public investors will want margins closer to specialist chip names. Transfer prices that were once left-pocket, right-pocket become continuing related-party deals with caps and disclosure.

Capacity allocation tightens the same knot. Advanced process, packaging, and memory remain scarce. Each unit sold outside the group is a unit Alibaba Cloud does not get. Exclusive co-design with Tongyi models supports the parent’s efficiency story and weakens the subsidiary’s merchant narrative; open sales do the reverse. Alphabet can resolve that conflict inside one shareholder base. Dual listings cannot.

China’s Internet Giants Are Listing AI Units Because the Parent Stock Will Not Price Them(配图 1)

Baidu spread Kunlunxin’s ownership early and now manages a long shareholder register and procurement-linked allocations in the listing process. Kuaishou sold a minority in Kling at a high mark with repurchase protections if an IPO slips. Alibaba kept T-Head wholly owned for years and is pushing it straight toward a public book, which avoids early-investor politics and leaves no clear private price as a reference—hence valuation ranges that still differ by a factor of two or three.

In each case the parent remains the controller, the consolidator, and the party funding much of the upside. The market’s sharpest AI multiple migrates to the subsidiary’s code. What is left for the parent is e-commerce, cloud, and a discounted claim on the unit that was supposed to redefine the whole.

本文由本站编辑组整理与排版,内容基于公开信息核实后归纳发布,与文中提及的产品、公司无隶属关系。我们持续跟踪 AI 领域的模型、产品与行业动态,力求信息准确、表述清晰;文中涉及的名称、价格与功能细节可能随官方调整而变化,请以官方最新说明为准。如发现疏漏或表述有误,欢迎通过站内渠道反馈指正。

相关推荐 按相关度与热度排序

全部资讯 →
  1. AI 时代,公司的「新厂房」长什么样?

    十年前的百度世界 2016 大会上,李彦宏曾站在台上提出「AI is the New Electricity」(人工智能是新电能),提前预判了人工智能浪潮。那时,很少有人能想到,走进写字楼,程序员用 Agent 写代码,运营用 AI 跑报表,连行政都在用大模型搭报销流程。不过,索洛悖论也再次产生。你去问企业管理者,十有

    来源 腾讯网
  2. 李彦宏再判AI下一幕:智能组织怎么落地

    作者|孙芮微信|lombredeau2016年百度世界大会上,李彦宏提出:“互联网的下一幕,就是人工智能。”彼时,移动互联网的高速增长开始放缓,互联网行业正在寻找新的增长动力。李彦宏将目光投向人工智能,认为它将成为驱动下一轮产业变革的关键技术。十年过去,人工智能逐渐走进产业生产和日常生活。随着ChatGPT带动生成式A

    来源 腾讯网
  3. Axelera AI 公布数据中心级 AIPU 芯片 Titania 架构,2027Q3 投产

    IT之家 11 月 13 日消息,荷兰芯片初创企业 Axelera AI 当地时间 12 日公布了其第三代 AIPU 芯片 Titania 的架构。其瞄准数据中心推理市场,预计 2027 年第 3 季度量产,已获得超 3 亿美元(IT之家注:现汇率约合 20.06 亿元人民币)的订单。 Titania 芯片设计功耗为

    来源 AI 行业观察
  4. 杨元庆谈AI边界:只出初稿,关键决策还得人来定

    快科技11月12日消息,在今日百度世界大会上,联想集团董事长兼CEO杨元庆与百度创始人李彦宏展开对话,围绕AI在企业经营和供应链管理中的应用进行了交流。 杨元庆表示,AI让供应链决策从“

    来源 快科技
  5. 百度词元所给大模型立标尺:AI比拼转向可靠性

    百度词元所选择的是建立度量衡。把 AI 产业从「比谁便宜」,推向「比谁靠谱」,并让「靠谱」第一次有了被度量、被公示的载体。撰文|焦丽莎今年夏天,硅谷发生了一起反常的闪电并购。一家名叫 OpenRouter 的初创公司,短短 80 多天时间估值从 13 亿美元暴涨到近 80 亿美元。其投资方英伟达原本想要收购,却被支付巨

    来源 腾讯网
  6. AI花了多少钱换回多少活:乌镇闭门会谈RSI与Token账本

    21世纪经济报道记者邓浩 一家企业花了100多万元采购的软件,没能真正用起来。后来,一名熟悉油田业务的工程师借助AI做了个版本,反而被推广开了。 百度秒哒总经理朱广翔在乌镇分享了这个案例。对于这类业务,AI带来的变化很具体:熟悉现场需求的人,可以更直接地把想法做成软件,再拿回工作中使用。 实际上,AI正在接下越来...

    来源 腾讯网
  7. 快手邱慧谈商业Agent:效果时延成本怎么平衡,动态知识库如何落地

    随着大语言模型能力的持续演进,电商场景下的智能对话系统正从传统 FAQ 问答向具备规划、决策与工具调用能力的 Agent 架构迁移。 本文整理自快手大模型应用算法专家邱慧在 AICon 2026 深圳站 的分享《智能互动 Agent 在快手商业场景的落地实践》。他围绕智能互动 Agent 在互动业务场景下的真实落地实践

    来源 AI 行业观察
  8. 张一鸣:字节AI追平OpenAI与Anthropic最早2029年?

    【TechWeb】11月13日消息,据澎湃新闻消息,近期字节跳动创始人张一鸣再次在内部分享中谈到AI业务。他表示,字节有能力与OpenAI和Anthropic展开竞争,但真正追平前沿技术,“最早也要到2029年或2030年”。 张一鸣坦言,与海外顶尖AI公司相比,字节Seed系列模型的预训练数据集规模仍然少了,这是字节

    来源 AI 行业观察
  9. 百度智能云伐谋3.0:AI自进化如何进入产业一线

    编辑|杜伟AI 自进化(RSI),自年初以来按下了加速键。从编写代码到设计实验、运行测试,AI 逐步进入 AI 研发的核心流程。当 AI 利用研发成果反过来改进下一代 AI,一条自我加速的技术演进路径便开始显现。这正是 RSI 概念受到关注且热度持续飙升的原因。11 月 10 日,AI Futures Project

    来源 腾讯网
  10. 亚马逊云科技Continuum:AI攻防提速下漏洞修复如何跑通

    Amazon Continuum:把“发现—排序—验证—修复”跑通今年10月,亚马逊云科技宣布针对代码漏洞的Amazon Continuum正式可用,该服务能够以机器速度发现、优先级排序、验证并修复安全风险。在发现阶段,Continuum分析代码库并形成候选漏洞发现,记录支持每个候选的源代码证据。江学森介绍,Conti

    来源 TechWeb
  11. Ring Drone Cam取消手动操控,AI Pilot全权接管飞行

    科技媒体 Android Headline 昨日(11 月 12 日)发布博文,报道称由于在演示过程中连续出现 2 次坠机故障,亚马逊旗下智能家居品牌 Ring 确认其首款安防无人机摄像头 Drone Cam 不再开放手动操控功能。

    来源 IT之家
  12. 中科院合肥团队研发肺结节AI诊断新模型M³Net,准确率更高成本更低

    IT之家 11 月 13 日消息,中国科学院合肥物质院健康所 / 合肥肿瘤医院科研团队近期联合国内外多家单位,在肺癌诊断研究方面取得进展,发展了一种基于胸部 CT 影像的肺结节 AI 辅助诊断新方法,研究结果发表在 Information Fusion 上。 肺癌是高发恶性肿瘤,早期筛查是提高治愈率和降低死亡率的关键环

    来源 AI 行业观察
  13. 文心5.5 Flash Preview实测:编程提速,网页游戏双场景交付

    在11月12日开幕的百度世界大会2026上,这家曾开国内AI先河又略显波折的巨头发布了一系列前沿新品。其中尤为引人注目的,就是最新基础模型文心5.5 Flash Preview(ERNIE 5.5 Flash Preview)。百度基础模型研发部负责人孙天祥也首次公开亮相,介绍了这款新模型的参数规模和核心优化方向。按照

    来源 腾讯网
  14. 六款国产AI模型现Blackwell硬件痕迹:约12000块受管制GPU

    2026年11月12日发布的法证报告显示,DeepSeek-V4、Kimi K3等六款中国模型权重初始化痕迹指向NVIDIA Blackwell GPU,涉及约12000块算力资源。此前同实验室旧版模型均匹配Hopper特征。报道从权重可验证硬件痕迹出发,拆解取证机制,分析对评测透明度与管控交叉领域的影响,并给出战略层

    来源 AI 行业观察
  15. 亚马逊云科技中国团队谈AI落地:黄金三角闭环怎么走

    亚马逊云科技中国团队的经验是,以业务战略为中心,围绕场景、数据和人才构建“黄金三角”。亚马逊云科技产品专家团队总经理崔玮表示,智能体输出是否可信、执行是否有边界,除了安全思路,更重要的准确性来自企业内部数据。安全能力方面,亚马逊云科技持续深化“安全赋能AI”和“AI赋能安全”。

    来源 TechWeb
  16. 亚马逊云科技:AI落地为何个人提效企业不增效,黄金三角怎么补

    亚马逊云科技中国团队的经验是,以业务战略为中心,围绕场景、数据和人才构建“黄金三角”。亚马逊云科技产品专家团队总经理崔玮表示,智能体输出是否可信、执行是否有边界,除了安全思路,更重要的准确性来自企业内部数据。安全能力方面,亚马逊云科技持续深化“安全赋能AI”和“AI赋能安全”。

    来源 TechWeb
  17. 亚马逊云科技江学森:AI漏洞年产千级,Continuum如何闭环修复?

    【TechWeb】11月13日消息,人工智能正在改变网络安全行业。随着新一代AI模型具备自主挖掘漏洞的能力,漏洞发现正从“几年一遇”变为“一年几遇”,甚至走向工业化批量发现。研究机构Epoch AI的数据显示,仅2026年6月,全球高危及关键漏洞数量约为1500个,是此前月度历史纪录的3.5倍。10月刚刚爆出的PHP2

    来源 腾讯网
向下滚动自动加载更多